This episode focuses on Machine Learning and Artificial Intelligence. These topics are massively discussed in investor and entrepreneurial circles, as well as the media in general. There is a trickle of that starting to move into areas of health tech and health data. There is a lot of potential and discussion around what that could mean.
It was about time that we tackled this subject to see what potential it has to help make better use of all the data that we are collecting on health. I have been spending more time on the conference circuit around this topic looking for technology that is adding value in this area. This means helping us make better decisions with less error and less effort.
This is a health data podcast, and as you will have understood through listening to previous episodes, there are a lot of challenges to getting actionable information and value out of today’s health data. So machine learning looks promising to potentially help us bridge that gap.
This will be the first of many episodes where we look into the subject, and today is a bit of an intro into the subject.
– Christopher Kelly
This episode’s guests are the Nourish Balance Thrive team, Christopher Kelly and Tommy Wood. Chris and Tommy are friends of mine whom I bump into often on the functional medicine conference circuit. Chris and Tommy run the Nourish Balance Thrive podcast and are constantly digging into functional medicine and related areas to see what they can extract to help athletes perform better.
They’ve used the data they’ve collected over the last three years that they’ve been working with athletes – as an input to a machine learning tool, to cheaply predict what an athlete should prioritize working on to improve his or her performance. This is, to my knowledge, the first time that machine learning has been applied to the area of functional medicine.
“We have an output from the algorithm; for each individual prediction, we have a sensitivity and specificity. Our H. pylori prediction has 100% sensitivity and a 98% specificity. That’s basically a gold standard test. That’s as good as doing the real test.”
– Tommy Wood
You can run the test yourself to understand what we’re talking about better by going to TheQuantifiedBody.net/machinetest. That will take you through a series of questions before predicting the issues that blood, urine, and stool tests would uncover for you, without actually investing in those tests.
The episode highlights, biomarkers, and links to the apps, devices and labs and everything else mentioned are below. Enjoy the show and let me know what you think in the comments!
What You’ll Learn
- Mutual respect between guests and host for their work (4:23).
- Using machine learning to determine patterns in very large data sets (4:33).
- Artificial intelligence is on the rise in the health market – will be the topic in future podcasts (7:29).
- Machine learning is useful in functional medicine because of the ability to produce simplifying algorithms for detecting complex physiological processes (8:06).
- The process of developing the Nourish Balance Thrive (NBT) questionnaire for assessing five major performance issues (10:47).
- The basic and advanced biomarkers tested on individuals during the development of the NBT algorithm (13:36).
- In some cases, machine learning algorithms determine health issues with more precisions compared to the judgment of individual medical practitioners (15:47).
- Power output in athletic performance depends on oxygen deliverability – in large part, determined by levels of oxygen-binding protein (hemoglobin) in red blood cells (19:04).
- The importance of choosing the right study – population of people based on whose results machine learning algorithms are developed (20:28).
- How algorithms are re-validated and the reasons Chris is confident in the predictive power of the developed model (22:47).
- The NBT questionnaire retains high sensitivity and specificity in predicting results which individual athletes would obtain when actually testing the algorithm’s predictions (25:31).
- The logic behind how algorithms make predictions in answering input questions (28:05).
- Combining small decision trees into an overall big algorithm with real-life predictive power (29:36).
- The background leading Chris to journey into artificial intelligence and machine learning fields of study (30:12).
- A practical walk-through into how the NBT machine test works and how they interpret results (33:17).
- The NBT machine test provides test clients with rankings of where each person stands in terms of 5 main performance issues and determines which issue to focus on the most (34:59).
- Based on algorithm results predictions, clients are usually asked to come to the US for getting specific follow-up tests done (37:49).
- Tommy hopes to accumulate success stories of tackling athletes’ performance issues, thus prove the actionability of the developed algorithm (39:49).
- After detecting weak points in athlete performance, the used interventions have a base in low-risk diet and lifestyle modifications (41:12).
- The potential of machine learning to revolutionize important aspects of life, including human health (42:29).
- Compared to traditional medicine, functional medicine considers the multi-complexity of factors influencing health (45:48).
- Developing useful applications in health doesn’t always require really big data – ex. NBT uses data from a relatively small study population of 1000 athletes (48:15).
- The amount of data necessary for machine learning application in health depends on the artificial intelligence tool used for computing patterns (49:09).
- Machine learning applied to detecting specific root causes of chronic illness (50:52).
- Modern medicine solves acute health conditions but there is a strong need to utilize proactive approaches in chronic illness prevention (52:49).
- Resources for learning more about the complexities and applicability of machine learning (56:50).
- Picking up machine learning is accessible and available, even for beginners with no programming skills (58:00).
- How best to connect with Chris and Tommy and learn more about their work (58:50).
- Influential people in the field of functional medicine (59:52).
- The biomarkers Chris regularly tracks to uncover and solve underlying causes of health issues (1:01:45).
- The importance of optimizing both mind and body towards better health – including diet, exercise and meaningful relationships with others (1:03:56).
- Monitoring blood glucose is an effective self-experiment which has a big payoff for health, performance, and longevity (1:05:35).
- Using the Wim Hof method towards improved management of blood glucose metabolism (1:09:11).
- Ketogenic dieting and why aiming for overall stability in blood glucose regulation is among the most important health strategies (1:11:01).
Thank Chris and Tommy on Twitter for this interview.
Click here to let them know you enjoyed the show!
Christopher Kelly & Tommy Wood, Nourish Balance Thrive
- Main Website: Introduction to the Nourish Balance Thrive project.
- Nourish Balance Thrive Podcast: Features over 100 episodes covering functional medicine and related topics.
Machine Learning Applied in Functional Medicine
- Nourish Balance Thrive Test: This recently developed 7 minutes questionnaire-based test is able to detect performance issues in athletes. It serves as a filter for which aspect of optimizing performance an individual should focus on improving the most. Give it a try!
- Short Explanation Video: How the Machine Learning test uncovers underlying root causes of physiological weak-points which are holding athletes back from their peak performance.
Monitoring Blood Glucose
This experiment involves tracking measurements of glucose (blood sugar molecules) concentration in your system. It reflects the body’s ability to properly metabolize food and feed cells with essential energy in the form of glucose molecules. Fasting glucose means testing first thing in the morning before eating anything. As such people are enabled to follow overall functioning of the body’s energy metabolism – whether glucose levels are used up in a stable way.
By making use of continuous glucose monitoring (see below), more specific information about glucose metabolism can be derived. For example, Chris has detected that elevated levels of blood glucose after a meal (post-meal glucose spikes) are sufficiently reduced when he takes a walk after eating. He has also discovered that intense exercise drives his glucose levels up to 180 milligrams/deciliter meaning that eating food is not the only reason for elevated glucose concentrations.
- Glucose Tolerance Tests
- Fasting Glucose: One of the most researched biomarkers in human health. Optimal fasting glucose levels are between 83 to 88 milligrams/deciliter.
- Fasting Insulin: The cells in the pancreas release insulin into the bloodstream in response to increases in blood glucose concentrations. Insulin functions to enable the intake of glucose from the bloodstream into the cells of your body. Optimal fasting insulin is above 5 microunits per milliliter.
- Hemoglobin A1C: One of the most useful markers in testing for glucose intolerance. Its interpretative power comes from the connection between glucose and hemoglobin – the protein in red blood cells (RBCs) which carries oxygen. Because RBCs live approximately 3 months, Hemoglobin A1C reflects the average blood glucose levels over this period. Higher levels of hemoglobin A1C indicate poorer control of blood glucose levels with optimum HbA1c levels being below 5%.
- Lipid Profile Panel
- High – Density Lipoprotein (HDL): The traditional measure of ‘good cholesterol’ used by doctors and healthcare. For example, levels above 60 mg/dL are protective of cardiovascular disease.
- Low-Density Lipoprotein (LDL): The traditional measure of ‘bad cholesterol’. Less than 100 mg/dL is an optimal level, while levels between 160-189 mg/dL increase the risk for cardiovascular disease.
- Lipoprotein(a): Lipoprotein molecules carry cholesterol and similar substances through the blood. Tests can measure a specific type of lipoprotein called lipoprotein-a. Higher levels of this marker imply risk of artery damage. Dr. Kahn states that in most labs normal reference ranges for lipoprotein(a) should be under 30 mg/dL.
- Thyroid Functional Test Panel: This panel of tests typically includes testing for circulating levels of thyroid hormones such as Thyroid Stimulating Hormone as well as the thyroid hormones triiodothyronine (T3) and thyroxine (T4). Proper functioning of the thyroid gland is key to athletic performance. The thyroid serves as a regulator for speeding up or slowing down human metabolic processes (conserving vs. using up energy, based on energy availability).
- Liver Function Tests: When liver functioning is physiologically stressed, the blood levels of liver enzymes Alanine Transaminase (ALT), Aspartate Transaminase (AST), and Alkaline Phosphatase (ALP) get to a higher level.
- Hemoglobin: The protein in red blood cells that carries oxygen throughout your body and is usually known for its role in diagnosing anemia – a sometimes serious health condition characterized by low oxygen delivery throughout the body. Because the Nourish Balance Thrive team strives to enable athletes to perform optimally, the team sets hemoglobin ranges higher than the standard upper ranges for eliminating anemia. As such, ranges are above 13 grams/deciliter in females and above 14.5 in males aiming for peak athletic level oxygen deliverability throughout the body.
- The 25-hydroxy Vitamin D Blood Test: The most accurate way to measure how much vitamin D is bioavailable to your body is the 25-hydroxy vitamin D blood test. Optimum vitamin D levels range between 50-70 ng/ml.
Lab Tests, Devices and Apps
- Differential Blood Cell Counts: includes measuring concentrations and important ratios between different types of cells found in blood including white and red blood cells, platelets, or immune system specific cells such as neutrophils or basophils. This is a very common test and differential diagnosis uses it – seeing whether particular aspects of a person’s physiology are functioning more strongly than others or if there is a need for follow-up medical tests.
- Blood Chemistry Panel: This test includes measuring of blood chemistry parameters including sodium, potassium, glucose, urea nitrogen, creatinine, total protein, albumin, globulin, cholesterol, triglycerides, total iron and other markers. This is also a very common test which serves to examine the overall physiological functioning of organ systems which are most important in human health.
- Dried Urine Test for Comprehensive Hormones (DUTCH): A lab test which uses mass spectrometry analytical methods. These methods are significantly more precise in measuring hormone levels compared to blood or saliva tests – most of which use antibody-based immunoassay analytical methods. Mass-spectrometry also allows for comprehensive analysis of metabolites of hormones and thus provides a more – comprehensive physiological picture. For athletes, the Nourish Balance Thrive team suggests that optimal scores are above 4 points, considering the test’s internal reference ranges for providing scores on hormonal balance in a person’s body.
- Organic Acid Test: This test provides an accurate evaluation of gut yeast and bacteria functioning, thus offering a snapshot of an individual’s nutritional and metabolic profile.
- GI-Map Test: Stool testing which uses DNA sequence analytical detection techniques of gut microorganisms including opportunistic organisms, normal gut bacteria flora, parasites, and fungi. The strong confidence in results stems from the ability to quantify the amounts of specific individual microorganisms instead of merely detecting their presence in the gut.
- Comprehensive Stool Analysis: A test which measures key markers of digestion, nutrient absorption (intake into the bloodstream after digestion), and inflammation in the gut.
Tools & Tactics
Diet & Nutrition
- Ketogenic Diet:1 A high fat, moderate protein and low carbohydrate diet. This diet is particular in that it changes the metabolism so that it burns ketones instead of glucose for fuel. 2 A ketogenic diet usually leads to elevated fasting glucose levels but it would be a mistake to apply standard fasting glucose reference ranges for long-term ketogenic dieters. This is because fasting glucose epidemiological studies do not consider special ranges for subpopulations of people who make use of a ketogenic diet.
- Matula Tea: A type of herbal tea which is potentially effective in removing H. Pylory – related gut dysbalances causing health and performance issues. This is a relatively low-risk intervention compared to taking antibiotics as first treatment.
- Sulforaphane: A chemical found in abundance in broccoli sprouts that people can either grow at home or grind up the seeds. Sulforaphane can potentially eradicate H. pylori infections.
- Time-Restricted Eating:3 Eating only daylight hours. This strategy potentially improves glucose tolerance.
- Whole Body Resistance Training:4 Moving your body more creates a deficit for glucose molecules in cells of your body. In turn, blood glucose levels can move more easily into cells and an overall effect of improved glucose tolerance is potentially achieved.
- Wim Hof Method: This method consists of breathing exercises, training of mindset or concentration and gradual exposure to cold. One important effect of this combined method is gaining control of voluntary activation of the sympathetic nervous system and improving the innate immune system response5.
Tech & Devices
- Measuring Blood Glucose6
- Pin-Prick Glucose Tracking Devices: The most popular and easily accessible devices for checking blood glucose. The most popular devices, and ones we’ve discussed before, are the Precision Xtra Blood Glucose and Ketone Monitoring System in the U.S. and the Freestyle Optium Neo Glucose/ Ketone meter in the UK.
- Continous Glucose Monitoring (CGM): A device containing a small sensor just underneath the skin that measures glucose continuously (ex. every 5 min). A transmitter then sends wireless data to a receiver which displays glucose trends. One of the most popular CGMs on the market is the Dexcom G4.
- FitBit: This company offers wearable devices which include cardiovascular fitness tracking. The Fitbit Surge is a fitness watch that offers GPS tracking, heart rate monitor, all-day tracking, and sleep tracking. The Fitbit Charge HR monitors physical activity and sleep quality.
Other People, Books & Resources
- Quest Diagnostics: A company in the United States offering easy access to most of the basic lab tests, ex. blood cell counts or lipid profile panels.
- Great Plains: A company which offers an Organic Acids Test (OAT) featuring testing of more than 70 markers from a urine sample.
- Genova Organix: While the NBT team mainly utilize data from the company Great Plains, they sometimes also use data from Genova Organix because this company also offers organic acid testing and some clients have already done it.
- Diagnostics Solutions: This company offers the (Gastro Intestinal) GI-Map Test and are among leaders in the field of precision genetic profiling of gut microbiome.
- Doctor’s Data: A company which offers stool sample test featuring testing of microorganisms functioning in the gut flora.
- 23andMe genetic testing: The largest personalized genetics company offering direct to customers testing. Analogous to the Nourish Balance Thrive test serving as a strategic filter, 23andMe genetic testing also does not lead to diagnostic results but focuses on guiding individuals to focus on specific aspects of their health and performance.
- Regenerus Labs: This company is in the United Kingdom and focuses their services in the area of functional medicine. The company is discussed in the context of there currently being difficulties in obtaining functional medicine – relevant tests in countries other than the US.
- Doctor’s Data: A laboratory testing company which offers heavy metal burden, nutritional deficiencies, gastrointestinal function, cardiovascular risk, liver and metabolic abnormalities testing profiles.
- Dale Bredesen: An expert in the mechanisms of neurodegenerative diseases including Alzheimer’s disease. He offers courses in his approach to treating this health condition.
- Ben Greenfield:7 A professional competitor and endurance-training athlete. Previously Chris has discussed the story of his health decline and recovery on Ben Greenfield’s podcast and this story has strongly resonated with listeners who are athletes.
- Robb Wolf:8 A former research biochemist who is quite influential in bringing Paleo to the mainstream.
- Jeremy Howard: Offers courses for people who have basic coding skills but are beginners in the machine learning field. Compared to Chris, Jeremy uses a different sub-branch of machine learning known as Deep Learning which is currently very popular. Tommy discusses how Deep Learning can help to detect lung cancer from Computer Assisted Tomography (CAT) scans of people’s lungs.
- Bryan Walsh: A naturopathic doctor who produces a youtube video series on interpreting blood chemistry results, as part of the Wellness FX company. Brian has previously participated in Robb Wolf’s podcast discussing adrenal fatigue and the effects of low cortisol.
- Chris Kresser: Works in ancestral health, Paleo nutrition, and functional and integrative medicine.
- Mark Hyman: A doctor in the field of functional medicine who works to tackle the root causes of chronic disease.
- The Master Algorithm: A book written by Pedro Domingos in which he discusses the applicability and future potential of machine learning. Previously he has been on the Nourish Balance Thrive podcast to discuss how machines can learn.
- NIH PROMIS: The National Institute of Health has sponsored the development of the Patient-Reported Outcomes Measurement Information System. This is a set of person-centered measures that evaluates and monitors physical, mental, and social health in adults and children. Based on personal experience with athletic performance and informed intuition about health, Chris selected the questions which are part of the Nourish Balance Thrive machine test from this free NIH PROMIS database of questions.
- XGBoost: The algorithm used by Nourish Balnce Thrive to develop their test in the field of functional medicine. This algorithm is most popular among the Kaggle community. Kaggle is a place where the world’s leading practitioners complete (usually, but always, for a prize) to solve machine learning problems. The particular model of this algorithm used by Chris does not require much computing resources in order to train the models.
- fast.ai: Jeremy Howard’s website where he teaching online courses in Deep Learning. While Deep Learning turned out not to be the algorithm applied in the Nourish Balance Thrive project, fast.ai DeepLearning courses are useful for machine learning practitioners of any type. Most Deep Learning applications are computationally expensive and require more than only your laptop to perform. For example, you might have to sign up to Amazon S3 data storage services and purchase computer hardware (likely from the company Nvidia) in order to be able to train models using Deep Learning algorithms.
- IBM Watson Health: Overview of healthcare applicability of the IBM Watson’ artificial intelligence platform. Functional medicine differs from traditional medicine in that it focused more on personalized, integrative and preventative health. It might be argued that IBM Watson’s program is, in fact, using traditional medicine approaches while adding machine learning as another layer of understanding patterns in health.
- Artificial Intelligence in Medicine Conference: Chris attended this AI conference, focused on using Deep Learning to uncover the root causes of chronic disease. One main argument discussed was that root causes may be less important compared to being able to diagnose health issues and apply treatment solutions. However, while machine learning is powerful in diagnosing health issues, it remains a tool which requires an understanding of the process.
- Google Deep Mind: By analyzing million hand-labeled images of diabetic retinopathies (damage in the eye caused by high blood glucose levels), the team created a learning algorithm that predicts diabetic retinopathies better than a human could9 The same potential science can also be applied in preventative medicine measures – by integrating diet, exercise and lifestyle factors to study, for example, prevention of diabetic retinopathies.
- Python: The programming language Chris uses which is a language that is readable and non-obscured in any way – thus offering user-friendly access to programming.
- Amazon & Yahoo: Chris has worked in these companies which use machine learning to optimize their business in analytical ways.
Full Interview Transcript
(0:04:23) [Damien Blenkinsopp]: Hey Chris and Tommy, welcome to the show.
[Christopher Kelly]: Thank you for having us. I am delighted to be here. It’s a privilege and an honor. I’m a long time listener, so it’s very exciting to be here.
[Tommy Wood]: Yeah, likewise.
(0:04:33) [Damien Blenkinsopp]: It’s awesome. Yeah, having seen you guys at various conferences over time and obviously having many discussions, it’s about time.
So today we’re going to dive a little bit into machine learning, because you guys have been playing around with that. Chris and Tommy, what is machine learning? Is it the same as artificial intelligence?
I think, first of all, we’d better just give a bit of background. I think what people are looking at in the news and everything you could think it’s anything, and maybe nothing, and maybe it’s the end of the world, Terminator style, pretty soon. So what is it really, and what is it today?
[Christopher Kelly]: That’s a really good question.
I get the sense that people are starting to use the term ‘Machine Learning’ like people used the term ‘Internet’ in 1999. There were internet companies popping up all over the place. And there are machine learning companies popping up all over the place now and I think that maybe is a bit of a warning sign that maybe there’s some hype going on here.
Like anything else, machine learning is just a tool and really what you care about is the application. So I think that’s maybe an important point to note.
I should make it clear that I’m a practitioner of machine learning, and not necessarily an expert of the academic sort. And this may be important for people listening because I want to encourage people to take part in this activity, especially if you’re already a code or computer programmer of some sort.
I would say that you need to know how the controls work.
Imagine you’re driving a car. It’s important that you understand what happens when you turn the wheel, and it’s important to know what happens when you press the pedals but you don’t necessarily need to know how internal combustion works in order to drive the car. And I think the same is true of machine learning; it really doesn’t need to be very, very complex unless you’re going to be researching and developing new algorithms.
So to answer your question specifically, machine learning, in my mind, is a sub-branch of artificial intelligence.
And I spent most of my life writing computer programs, very carefully, by hand coding algorithms. IF-THEN-ELSE, that type of construct that some people will be familiar with. In machine learning, I’m doing something different.
Over the last three years, we’ve collected lots and lots of data–about 100,000 total features–from about 1,000 athletes. And then I’ve used that data to train an algorithm. So I’ve shown an algorithm many, many examples of the pattern that I would like to identify in the future, and then the machine–it’s kind of a funny thing to say–learned how to predict the patterns that I was interested in.
At no point did I ever hand code an algorithm, and I think that’s what makes machine learning different from regular programming.
(0:07:29) [Damien Blenkinsopp]: Thanks for that overview, Chris.
So what is artificial intelligence?
I’ve seen a lot of the hype. I can tell you, I go to conferences now and a lot of start-ups are talking about adding machine learning and AI to their apps just to be a bit cooler and to attract investment and so on. So there’s definitely a bit of hype around it, which I think is why it’s worth talking about.
In contrast to machine learning, which is what you’ve been doing, what is artificial intelligence?
[Christopher Kelly]: Yeah, that’s a good question I would rather not answer.
[Damien Blenkinsopp]: Okay, yeah it’s fine. We can explore it in a later podcast.
This is a topic I’ve been fascinated with and digging into. And I know it’s pretty complex. So, let’s just skip that one, shall we?
[Christopher Kelly]: Yeah.
(0:08:06) [Damien Blenkinsopp]: Right. So one of the unique things about what you’ve done is you’ve applied it in the area of functional medicine, which I don’t think I’ve seen done before. We’ve started to see a few applications for health.
But what do you think, if you’re looking at the area of health, where do you think it could be applied usefully just from how you’ve got to it? So you’ve done it for prediction of results. Is that the main area you see it as useful? Or are there other areas which you see that it could be applicable?
[Christopher Kelly]: Oh no, so yeah. Our application is just one tiny thing.
So to give people a bit more background, we’ve worked with about 1,000 athletes over the past three years, and the way that we’ve helped those people is we’ve uncovered the underlying root causes of the things which are holding them back from their peak performance.
We’ve used blood chemistry, and urinary organic acids, and urinary hormone testing, and then also stool microbiology, and then also PCR DNA analysis. And obviously it’s quite difficult to get some of these tests done. Blood chemistry is ubiquitous, obviously, but the other tests I talked about are quite difficult to get done. And they’re also quite expensive.
So the thing that I would like to achieve is first to make it much easier to do our program. See, you know, can I predict the results of these tests without you doing them. And then of course potentially it could bring the costs down in the long-run.
So when someone does one of our tests, somebody in a lab somewhere is putting a sample into a machine. And they’re doing some mass spectrometry, obviously that’s an expensive machine that is taking somebody’s time, and that costs money.
So one of the things I think that machine learning will be able to do in medicine is reduce the cost in the long term. And then it will provide greater access to people who perhaps might not otherwise be able to get a hold of these fancy tests.
[Damien Blenkinsopp]: Right, so it’s like a filter, so that people don’t necessarily have to do all of the tests.
Because when we look at the functional medicine process today, basically a functional medicine practitioner takes your history, right. He talks to you, and then he decides on an array of tests. As you said, this can be pretty expensive depending on how many you’re going to run.
And I think that’s one of the biggest issues for functional medicine right now, for it’s greater acceptance. Some people basically can’t afford the tests that they’re being told to do.
So what you’re saying is you’ve used a questionnaire that you give people, and you’re using the data from that to predict what the results would be in the tests.
(0:10:47) [Christopher Kelly]: Yes, exactly.
So we have 53 questions which form part of our standardized health assessment questionnaire. Those questions I chose personally from a large data bank of questions that are available online for free. It’s the NIH Promis data bank of questions.
And it’s a whole other story that I won’t get into now, but I was not feeling good in 2014. I chose these questions based on the way that I felt. Some of them were very relevant to me, and then others that I saw in the data bank I thought, well now that’s not really right.
They’ll ask you things like, ‘I was so tired I couldn’t get out of the bath.’ Or, ‘I couldn’t even leave the house, I didn’t have enough [energy].’ So this sort of chronic fatigue type questions. And I was definitely feeling bad, but not that bad. So I selectively cherry-picked these questions from this data bank.
Then we had 1,000 athletes go through our program. It almost became a standing joke that we would see the same person and the same problems over, and over, and over again.
So that’s what got us thinking. Is there some way that we could predict the results of these fancy tests using just these 53 close-ended questions that you could answer in seven minutes by clicking on radio buttons.
[Damien Blenkinsopp]: Right, so you’ve been giving that selection, your cherry-picked selection of questions, to everyone that you’ve worked with over the last three years.
[Christopher Kelly]: Exactly right.
So, everybody that’s been through the program, they’ve done [it]. We’ve worked a bit differently maybe from some other functional medicine providers that you’ve met in the past in that we always do the same set of tests.
Obviously, each person is unique. They have their unique history, situation, and goals. But the tools that we use to identify the underlying root causes don’t vary much from person to person. We use the same set of tests on everyone. And then at the same time they do the tests, we have them do the health assessment questionnaire.
I always have that data for every single person that goes through our program. So that’s how I was able to train the machine. I had the 53 close-ended questions and then alongside that, I have all the blood chemistry, the urinary organic acids, the DUTCH test, the stool culture, the stool PCR test.
So if you can imagine a great big spreadsheet with all of these things in columns. Then the final thing I’m trying to predict is do you have circadian dysregulation, or do you have gut dysbiosis, or do you have a glucose tolerance problem, or do you have an oxygen deliverability problem?
So that’s a higher order function that I’ve calculated using some of the other biomakers which form the columns of the spreadsheet.
(0:13:36) [Damien Blenkinsopp]: Okay, excellent.
Could you just go through the lists of tests you used? Because we talk about tests all the time on this show, so people will have run into them in past ones, and so on. So what’s the blood chemistry you’re running specifically?
[Christopher Kelly]: Sure. Do you want to talk about the blood chemistry, Tommy, because it was largely you that designed that panel.
[Tommy Wood]: Yeah. So we’re obviously based in the United States, and most of the blood panels are run through Ulta. So you do the tests at Quest Laboratory. And it’s stuff that people will be very familiar with.
I’m a big fan of doing the basics, because we know the basics work, and that include doing the history. So everybody comes in and does the history; that’s really important and sort of gave us the basis for how we could predict some of the results, like Chris was talking about.
And then the blood tests are a basic blood count, an extensive thyroid panel, a liver function test, a kidney function test, there’s calcium and [unclear 0:14:32], vitamin D, insulin, HbA 1 c, fasting glucose, and basic lipids test.
All the things that people will be familiar with that they can get from their doctor. And it’s also something that even if somebody is not in the United States it’s usually something they can get locally as well.
So those are the real… We cover those basics just because we know what they mean and how they apply to the physiology. Then it gives us some grounding to then expand into the newer tests.
[Damien Blenkinsopp]: Excellent. And so the newer tests you’re talking about, is that The Great Plains?
[Tommy Wood]: Generally. We do have some data from the Genova Organix, because some people have done that too, but it’s mainly The Great Plains organic acids test.
Then, like Chris mentioned, we do the urinary hormones, the DUTCH. The stool tests that we’re currently using are the Doctor’s Data–a comprehensive stool analysis with parasitology–and the Diagnostics Solutions GI-Map with the PCR. That’s the whole panel.
[Damien Blenkinsopp]: Right, great. So you’ve taken this data for everyone, and what you’re saying is you’ve seen correlations which will lead to five different outcomes that you’re looking for. Five problems to target.
(0:15:47) [Tommy Wood]: Yeah, so what the machine learning particularly–and Chris knows more about this than I do, definitely–what it’s really good at doing is predicting patterns. There’s the well-known example of the algorithm that was trained to identify lung cancer on X-rays, and it was able to do that better than the best radiologist in the world.
So if you give it enough X-rays which say this X-ray shows lung cancer then it learns what that looks like. And then you give it future X-rays and then it says okay this is lung cancer, this isn’t lung cancer. And it can do that more or better, more accurately, than a human radiologist can.
So, this makes me think of a time Chris and I went to Dale Bredesen’s training course last year to learn about how he treats Alzheimer’s disease. And Chris stands up and tells Dale Bredesen’s personal radiologist that at some point machine learning is going to make radiologists completely irrelevant because the machines are going to be able to do all the radiology for us.
That’s what machine learning is really good at. So if we give it specific patterns we want to look for, and the ones that Chris mentioned were low oxygen deliverability – that’s basically just another word to describe lower than optimal hemoglobin, which people probably will have heard of.
Then we talked about glucose intolerance. That is three different predictions in one group; so it’s high fasting blood glucose, high HbA 1 c, and high fasting insulin.
We obviously have tighter levels than most people would probably think of. We’re talking above 88 milligrams/deciliter blood glucose or a fasting insulin above 5. So that’s kind of our level of where we’d like to see things.
Then, if we talk about the dysbiosis we’re predicting things like H pylori, or clostridium, or a general bacteria overgrowth, yeast overgrowth on the OAT, something like that. That’s based on the lab values that you get from, say, The Great Plains Organic Acid Test.
And then hormonal balances, so that’s low estrogen in females, low testosterone in males, again based on the DUTCH references ranges. And then circadian dysregulation, which is basically having a cortisol marker outside of the normal range at a given time point during the day. Again on the DUTCH you need at least a 4 point. Now it can be a 5 point if you take a sample in the middle of the night.
So based on all of those things, we can kind of drill down, and the machine will tell you what the ranking of those different problems is for you. So maybe glucose intolerance is the most likely issue that you have, and then it will rank the other ones too.
Then sort of the back-end we can look at the percentages, and we know how accurate the machine is at predicting each individual thing. Have like a sensitivity and specificity for each individual one, so we know how accurate it is, and what the likely issue is. And that means we can get people started very quickly.
Because we know if someone’s going to come to us with blood glucose dysregulation. We take a little bit more of a history and we know exactly what we need to do. We don’t need to do any blood tests first because we know what issues there are going to be. We can start people very quickly without them having to do all the tests first.
(0:19:04) [Damien Blenkinsopp]: Yeah, excellent.
The only one I don’t think we’ve come across before is low oxygen deliverability. Could you give us a little bit more background on that? Where does it come from, and what type of people have it?
[Tommy Wood]: Yeah, basically it’s based on hemoglobin.
People will have heard of hemoglobin when it comes to anemia. So if you have low hemoglobin–that’s the protein in your red blood cells that carries oxygen–then you’re considered to be anemic. It’s one of the markers of anemia. So people may have heard of that.
We have ours slightly higher. We know what levels athletes need to be at in order to perform optimally; so it’s above 14.5 grams/deciliter in males, above 13 in females. So those are higher cutoffs. Those don’t define anemia, but it defines what we’d like to get an athlete to if we can so they can perform optimally.
So we don’t call it anemia because we’re not detecting anemia. We’re detecting low oxygen deliverable, which is basically your blood doesn’t have as much hemoglobin as it could hopefully have. This means that you’re not delivering as much oxygen. You don’t have the capacity to deliver as much oxygen as you’d quite like to.
So the phrasing is important because we’re not detecting frank anemia. We’re detecting something else that we know is important for athletic performance because power output tracks very nicely with hemoglobin. So if you can increase that, you’ll definitely increase somebody’s athletic performance.
(0:20:28) [Damien Blenkinsopp]: Right, excellent.
And that brings us to a very important point of where your data comes from, and what the focus of it is.
I understand that data set of people that you have is quite important. You have to be quite careful of the selection and use of it. Why is it that important?
[Christopher Kelly]: So I’ll take that. I think it’s really, really important.
I went on to the Ben Greenfield podcast in 2014 with another one of my doctors – Jamie, my founding medical doctor, she’s a pro mountain biker. And I told this story of my health decline and then recovery and the use of some of the testing that we’ve talked about so far in that recovery.
And that story resonated with a particular type of athlete that listens to then Ben Greenfield podcast. And they were the people that came forward to work with us. I also talked on the Robb Wolf podcast and from my perspective, it was difficult to identify the two different types of people.
They seem to be very similar in their personality and their problems and their goals. All completely wonderful people and I’ve had a fantastic time over the past three years. But I already mentioned that I cherry-picked those questions out of this huge data bank of questions that were supposed to be for all people and all things.
So I think that the algorithms would not be particular good at predicting the results of these tests of people who didn’t fall into that same category. I don’t know, I haven’t tested this. But that is my suspicion because remember we said that machine learning was teaching a machine how to learn based on labeled examples.
So when Tommy talked about the X-rays there, a real radiologist had labeled this X-ray as this one is a malignant tumor, this one is not. And in my data set, we’ve said this one has low hemoglobin, this one does not. And we’ve taught the machine how to learn to identify this pattern using examples.
If I then went into a completely different population of people, let’s say people who only had chronic fatigue syndrome, well they might answer something completely different to my health assessment questionnaire. And so I don’t know whether I would be able to predict with the same accuracy.
I think this is something that we should experiment with.
(0:22:47) [Damien Blenkinsopp]:And when you’re going into this, how can you re-validate it for the other populations? Is that something you’re going to be doing on a long range basis, or how does that work? I imagine for some people you’re going to be collecting test data as well, rather than simply relying on the questionnaire for everyone.
So how do you see this going forward? How do you think it might work on insuring that it’s continuing to be valid? Is it going to be continuing to machine learn, or have you basically done a cut-off based on the training it’s already had?
[Christopher Kelly]: Sure. The analysis is already live on my website, and so I’m collecting some data already through people who are just visiting my website and seeing the analysis there, and taking it.
And then I’ve also spoken on a few different podcasts about the analysis. For each podcast that I’ve spoken on – and I should do the same for this one – I will provide a custom link that you can find in the show notes and that custom link allows me to identify the source of the traffic.
By definition, you are a particular sort of person if you listen to The Quantified Body podcast, and I think that might be important in the predictions. So the custom link, I think, is going to be really important. And it’s only once I’ve collected a certain amount of data will I be able to say this is a very strong prediction and maybe this is not so good.
And of course some of the people that do the analysis will go on to do the real tests. You can get started more quickly when you do the analysis, but for now we’re still doing all of the tests. So once I get back the real data – the real blood chemistry, the real urinary organic acids, all of that – I’ll be able to compare what the machine predicted versus what we actually tested.
Now, I’m not really expecting any surprises for some groups of people because when I was training these models, I deliberately held out 20% of our data. So I said we had data from 1,000 athletes. I held out 20% of that, set it to one side and then I only used that data once I had finished training the models. So I used that to test the models, and so that’s how we know how accurate they are.
I wouldn’t be here talking about it if I didn’t think the models were any good. And the reason I know they’re so good is because of this held out data set.
[Damien Blenkinsopp]: Great. And so, does it give back a correlation or something like that? Did you get a number like this is 90% accurate with the last 20% you used, or something like that, to give you that confidence?
[Christopher Kelly]: Yeah. So, Tommy, do you want to talk about the sensitivity and the specificity of the tests?
(0:25:31) [Tommy Wood]: So people maybe have heard of sensitivity and specificity, which is basically something we often use or calculate in medicine if we’re comparing a new test to a gold standard test. This is exactly what we want to do.
And basically the sensitivity tells you the likelihood that a positive result is truly positive. And the specificity tells you the likelihood that a negative result is truly negative. So you want both sides of that coin.
You could say that if you have 100% sensitivity, you’ll pick up everybody who is going to be truly positive about one thing. But if you don’t have any specificity then you’ll have loads of false negatives. There are lots of ways to balance that out. So you want both to be, essentially, as high as possible.
We have an output from the algorithm; for each individual prediction we have a sensitivity and specificity. So I’m looking at one right now. Our H pylori prediction has 100% sensitivity and a 98% specificity. That’s basically gold standard test. That’s as good as doing the real test.
Some of the other things are not going to be as accurate. Bacterial overgrowth has a 94% specificity. So they’re up there; I think the lowest one is maybe in the 80% in terms of specificity. If somebody has a negative prediction there’s a small chance they might still have a yeast overgrowth on the actual test results.
So it’s really close. It’s at the level where you could say that we’re close to being able to predict something as well as the test would be able to.
[Damien Blenkinsopp]: Wow. That’s pretty impressive. Just through 53 questions.
[Tommy Wood]: I just have to say that I actually couldn’t believe how good this was. And Chris has run it multiple times.
So originally we were going to do tests or predict urine results and stool results from blood test results. Then eventually we sort of worked our way back, and we got to the point where we were just using the questions. And it’s almost too good to be true, but I promise you it is actually true.
[Christopher Kelly]: That was my original idea. I thought blood chemistry is ubiquitous; anyone in the world – or that’s not true, but most people have access to blood chemistry. If you give me your CBC, for example, can I then predict the arabinose, which is a marker of candida overgrowth on the urinary organic acid test? Because that would still be quite cool.
And it turns out that does work, but what works even better is just me asking you these 53 close-ended questions.
(0:28:05) [Christopher Kelly]:But one thing I’d like to point out is there are five different answers, 53 different questions. So I believe that is 5 to the power 53… It’s 1 times 10 to the power 37 different permutations. So that’s a lot of different ways to answer this health assessment questionnaire It’s really a lot.
[Damien Blenkinsopp]: Right. It’s like this huge tree of permutations that’s going on there.
[Christopher Kelly]: Exactly.
[Damien Blenkinsopp]: So you’re getting people to take a lot of different paths, and eventually they’re coming to one outcome. So that’s where that specificity is coming from, from all of those permutations you’re driving them through.
[Christopher Kelly]: Right, exactly.
And that’s exactly how this particular algorithm works. We’ve used this algorithm called XGBoost, which is very popular from the machine learning website I would encourage people to visit called Kaggle.
Kaggle is a place where you could launch a competition and have the world’s leading practitioners compete – usually for prize money but not always – to solve your machine learning problem. And XGBoost, the algorithm that we used, has been a constant winner in the Kaggle space. And that is exactly how it works; it’s a boosted decision tree.
So, think about what happens when you call up the electricity board; you get presented with all these different options. You have to press one for customer service, two for sales, and all of that. So you can see that pans out into a decision tree, and that’s exactly how our XGBoost algorithm works. It’s a large number of these small decision trees.
(0:29:36) [Christopher Kelly]: And another really interesting thing that’s so simple it’s almost worth not talking about, and you can’t believe how well it works.
Each one of these small decision trees, they’re slightly better than chance. So if I’m trying to predict the results of a coin flip, then it gets it slightly better than chance. And it turns out that when you have thousands, or even millions, of these small decision trees that are slightly better than chance. And you combine them all together amd get a really strong learner that’s very good at predicting things.
So that’s how this algorithm XGBoost works.
(0:30:12) [Damien Blenkinsopp]: Great.
Chris, I know you’ve been going through artificial intelligence and machine learning for a few years now. I was just wondering if you could talk a little bit about your experience in this. You ended up choosing this particular approach to it.
Was it easy when you jumped into it and you wanted to learn it? I myself have been looking at it, also, because it’s this whole new world with all this potential. How have you found it? How has your journey been through it?
[Christopher Kelly]: I’m glad you asked that question, that’s a great question.
I have an undergraduate degree in computer science, and I’ve worked my whole life for big tech companies. I’m 41 years old now, and I’ve worked for Yahoo – they were the company that brought me from London to Sunnyvale to their headquarters. I’ve worked for Amazon. And I’ve worked for a search company within Amazon, and I’ve worked for two hedge funds.
All of those companies make heavy use of machine learning, but somehow the technology evaded me for the longest time. And the reason was, every time I tried to get into it I just found the subject matter so incredibly dry.
So if you go and read some academic papers on some machine learning algorithm, typically what you encounter is an abstract, or a small amount of text at the beginning of the paper that makes a lot of sense. Then you turn to page two and there’s this wall of equations, and you’re like, okay. And then you just put page one back on top of page two, move it to one side, and carry on hand coding your algorithm.
And that’s just been the way with the academic computer science community. It seems to be dominated by people who are very strong in mathematics and mathematics is the language that they use to communicate. But it’s not necessarily the best language for all computer scientists.
And so I’ve found some other resources very, very helpful. In particular, Jeremy Howard has been running some classes in San Francisco designed exactly for people like me. Luckily, those classes are now available online. And those were wonderfully helpful.
And it turns out that Jeremy Howard is using a different sub-branch of machine learning. He’s using something called Deep Learning, which is very, very popular at the moment. I had tons and tons of fun.
So Tommy mentioned the trend to identify the malignant tumor on an X-Ray; it’s a Deep Learning algorithm that’s doing that. So it’s different from what we’re using. We’re using XGBoost.
So Jeremy is arguably more state of the art, but he’s solving different types of problems. Deep Learning is better at solving these computer vision problems, and other things too.
So those courses, I think, were absolutely fantastic. That was what allowed me to get past this wall of mathematics and become a machine learning practitioner.
[Damien Blenkinsopp]: Excellent, thank you for going into that, because I think it would be amazing if more people started to apply this to health and functional medicine. And there’s a lot of listeners on this show – entrepreneurs, venture capitalists, and all sorts of types – who might find it a little bit easier and approachable knowing that there are ways around that.
(0:33:17) [Damien Blenkinsopp]: Another thing I wanted to do on this podcast is take the people listening through a practical walk-though of how it’s being used. So people are going to click on this link and go to the page where it is, and then what happens?
[Christopher Kelly]: Sure.
I’m a very visual person; I like to learn with audio and visual stuff. So I’ve paid someone to make some whiteboard explainer videos, because obviously this stuff is complex. So there’s a video, if you come to the show notes you can see the video. It’s the whiteboard explainer video that hopefully summarizes things that we’ve been talking about and explains how these things work.
And then as you go through the analysis, it’s really quite simple to do. All you do is you click on radio buttons and answer the questions. I’m going to ask you things like: ‘In the past seven days I’ve felt tired.’ And then the answers will be something like always, sometimes, never. I can’t remember the five different permutations, but you just answer the questions honestly.
They’re grouped into sub-categories that you’ll recognize, and anyone who has spent any amount of time not feeling good will recognize these questions intimately, I’m sure. And that’s it really.
Just walk through for seven minutes, and then at the end you’re presented with the results, which, as Tommy alluded to slightly earlier, we don’t give you the output of the model because it’s kind of confusing. You need to know quite a lot about how the model is made in order to interpret the output. It’s actually talking in probabilities, which are quite difficult to understand.
So the model is going to say whether it thinks you probably do match the criteria or you probably don’t. It’s mostly a binary classification.
(0:34:59) [Damien Blenkinsopp]: So to just highlight that point.
Basically as Tommy was saying earlier, it’s going to highlight whether you’re in a specific range in one of the tests. Is that the output for you guys? It’s going to say there’s this risk of H pylori, for example?
[Christopher Kelly]: Yeah, that’s exactly right.
The gut dysbiosis model, for example, is a composite of the H pylori prediction, the bacterial overgrowth, and yeast overgrowth. So we just lump all of those things together and call it gut dysbiosis. And so if the model thinks that any one of those things is true, then it’s going to predict a binary classification for the most part.
So we kind of argued about it–not argued but debated it–for a while about what we should show the user. In the end what we went for was just a rank. What you see on the results page are the things that the model thinks are most important for you. Because it’s kind of hard to interpret the output of the model, this probability as a percentage.
[Damien Blenkinsopp]: Right. So, I went through it myself, and you have the display of the five areas. And then it looks like a percentage, basically, right?
[Christopher Kelly]: Yes, that’s the number that in the end we decided to hide from the user, because it was confusing.
And so we can still see on the back-end, but for the user now what they’re seeing is just the rank of things. So these are like the order of importance of the five different categories.
[Tommy Wood]: Damien you used a slightly earlier version where you could still see the percentages.
It eventually turned out that that was becoming kind of confusing. So we thought that people could just focus on what the most important thing is, and that’s how people would then follow up through the system. But we have obviously all of the data to help.
[Damien Blenkinsopp]: I see. So it’s just going to highlight one of the items that is the most important to look into, for example if low oxygen deliverability is the thing you should focus on. Is that the point?
[Tommy Wood]: Yeah, well, you’ll get your ranking for all five. The order of importance for all five. So you’ll get some follow up, that Chris can tell you about, and that will be based on whatever was ranked number one.
[Damien Blenkinsopp]: Okay, got it.
And that’s the use of the tool, really, helping people to focus on the area. I mean, I used to talk about DNA tests like 23andMe as something useful to help you focus on things.
It’s not entirely accurate, it doesn’t give you a diagnosis or anything, but using it as a strategic filter to say: ‘There’s a lot of things popping up in lung cancer risk in my genetics, I should probably have a deeper look into that.’
So it sounds like you’re kind of proposing this to be used in the same way. Basically it’s a strategic device to look at where should I focus my efforts and have a look more into it.
[Tommy Wood]: Yeah.
(0:37:49) [Damien Blenkinsopp]: Great.
Have you got any case studies of people who have used it already? Anything that’s come out of it since you’ve been playing around with it?
[Christopher Kelly]: No. It’s totally brand new. In fact, just this morning I just signed up our first client who is a British guy living in Spain. So he could do it; it’s still possible to get the tests done, but it’s not easy. So…
[Damien Blenkinsopp]: It’s impossible in Spain. It’s really, really hard.
When I lived in Spain, I ended up moving to the United States because I got so frustrated. I was getting an MRI done and they gave me just the completely wrong results, and I thought, I’m done, I’m out of here. And I left, it was the end.
[Christopher Kelly]: Yeah, so that’s what we normally do. We do have clients from all over the world, and that’s what we normally say, ‘Can you come to the United States?’
And for most of the athletes that we work with, they can. So if you’re an IronMan Triathlete, for example, there’s a good chance you’re going to want to come to the United States for a race. And when you do the race, you can just stay in a hotel or an AirBnB, whatever it is. And then you can do all of the tests either at home or you can take a trip to Quest and get the blood drawn there.
So for this guy in Spain, I didn’t show him the exact output of the models, as we discussed previously. But when I looked on the back-ends, the models were really, really confident about several things which I know how to fix right away.
[Damien Blenkinsopp]: So it’ll be interesting to see how that goes.
[Christopher Kelly]: Exactly.
[Damien Blenkinsopp]: It’s incredibly good use case because so many people struggle with tests outside of the United States.
[Christopher Kelly]: Right.
[Damien Blenkinsopp]: It’s getting a little bit better in the United Kingdom now. There are some guys called Regenerus Labs who are doing a fair number of the functional medicine and other tests now by post and they handle that. But overall, it’s still really, really complicated, and I’m constantly getting questions about it.
This sounds like a really useful use case for it, for people who are also in Europe or even in places like this where they can’t get their hands on the tests in the first place.
(0:39:49) [Tommy Wood]: This is a really important part of the process. We think we can predict things with a very high degree of accuracy, but how well can we treat those things when we don’t have the full set of data. And we’re very confident that we can, but the only way you can find out is to actually do it.
Particularly with people who fit very nicely into the group that we used to train the data. So just more of the same kind of client that we’re used to working with and we get very good results with, that’s the ideal test bed. And then we can show that we can really do what we think we can.
[Damien Blenkinsopp]: Yeah. It would be really interesting to have you guys a few months or whatever down the line, once you’ve run it for a while and got some test results and some experience and so on.
And maybe it sounds like basically trial and error. You’ll just put someone through a program, say they’re living in Spain, and if it fixes him you’ll be like, okay, that worked. That’s a good data point for the model.
[Tommy Wood]: And maybe we’re just doing what Voltaire said, which is that we’re just entertaining the patient enough while nature cures the disease.
[Damien Blenkinsopp]: It would be great.
[Tommy Wood]: But in reality, I think we know how we would approach each of those different things. So if we’ve got a model that predicts something with a very high degree of certainty, then the likelihood that this person will see benefits based on what we suggest based on the algorithm is really, really good.
(0:41:12) [Christopher Kelly]: And we should talk about some of the interventions as well, because I think that’s important. It’s not like we’re predicting things and then asking people to take drugs that may have unwanted effects. We’re talking about lifestyle medicine here.
So let’s say the model predicted that you had a glucose intolerance problem. Well I can coach you, my wife can coach you, and any one of my coaches can coach you with how you can improve your glucose tolerance.
So you could do things like time-restricted eating where you only eat during daylight hours. That could improve glucose tolerance. Or you could move your body more. Maybe you could do some whole body resistance training that’s going to create an intracellular glucose deficit and make the glucose that’s in your blood go into cells more easily. And maybe that would improve your glucose tolerance.
Do you see what I’m saying? It’s mostly diet and lifestyle interventions.
[Tommy Wood]: Really low risk.
[Christopher Kelly]: Very low risk.
[Damien Blenkinsopp]: I’m guessing the ones where we get closest to actually some kind of medicine are gut dysbiosis, where you guys are using herbals and probiotics and things like that, primarily, aren’t you?
[Christopher Kelly]: Exactly. For example, this tea, there’s a Matula Tea. There’s a company on the internet that guarantees you that it gets rid of H pylori. And it’s very expensive, but they give you your money back if you send them a test and you’ve still got the bug.
And there’s other things like broccoli sprouts, sulforaphane, that people can either grow at home or just grind up the seeds. That may help with eradicating an H pylori infection. So fairly low risk compared to taking antibiotics, I would say.
(0:42:49) [Damien Blenkinsopp]: I thought we would take a little bit of the big picture look at this machine learning.
Having gone through this experience for yourselves, how transformative do you think machine learning could be? Or will not be, for that matter, for health over the next 10 years, given the examples you’ve seen? I know Chris, you’ve been to conferences and stuff and seen some examples as well.
What do you think the power of this is? Or isn’t?
[Christopher Kelly]: Oh yeah, I mean it’s going to completely revolutionize everything, I think. Almost everything. And it’s interesting that some of the jobs that I think are going to go are the white collar jobs.
So I know this from talking to Pedro Domingos, he’s been on my podcast. And I would highly recommend his book, The Master Algorithm, where he talks obviously in detail about this.
But it’s the white collar jobs, so anything where you’re doing something over and over again that doesn’t really require any manual movement. So some people I think mistakenly believe it’s the workers that are going to go, that they’re all going to be replaced by robots. But that’s not true.
When you look at, say, the skills of somebody building a house, those motor skills they’re using and their dexterity took millions of years to evolve. Computers haven’t got there yet. Whereas, identifying a malignant tumor on an X-ray, that’s just a pattern recognition thing that computers have already learned how to do.
So, if you’re laying bricks and mortar for a living, I think your job is safe. If you’re a lawyer or a radiologist, or somebody who issues patents, then I’m not so sure your job is safe. It’s very interesting.
[Damien Blenkinsopp]: I was reading a case study on J.P. Morgan and they were talking about deals, like mergers and acquisitions, and it was taking hundreds of thousands of hours of lawyer work before. Now it’s being done by a computer in a day.
[Tommy Wood]: There’s one thing we discuss a lot, sort of on the back-end. We’re basically it’s like we’re discussing things pretty much continuously. And one thing that comes up a lot, particularly as it relates to health, is the machine is only going to be as good as the person who is training it and the way that they train it.
So if you think about, I was reading something recently about how IBM’s Watson in health hasn’t produced as much or as fast as they thought it would. We wonder if part of the problem is the fact that you’re taking traditional medicine approaches and then just trying to add machine learning on top.
And as we know, the current approaches we have to chronic diseases or cancer aren’t necessarily the right ones. And these aren’t getting us anywhere as fast as we originally hoped. Because we’re still working around an acute care system for chronic diseases.
So there’s definitely the possibility that until we keep trying and failing this in various different arenas we’re just going to get the same wrong answers, but we’re just going to get them faster.
(0:45:48) [Damien Blenkinsopp]: Right. And I think Chris you brought this up in an email that sometimes the system we have is focusing on one marker, or is focusing on one diagnosis driven by one single input to that. There’s one reason why you get sick.
Whereas in the world of functional medicine we’re looking at a multifactorial, multicomplex, everyone is kind of different with different inputs, sort of problem situations. And from what I’ve seen with machine learning is it could be the answer to this because it will just look at all of the data and it will say if you look at these five things and how they vary, you get these different situations.
Whereas I guess the limitation of our human brain is we tend to focus on one thing and we’re just trying to say this leads to that, and it’s a linear fashion.
[Christopher Kelly]: Yeah, absolutely. So that’s a really good example, actually.
If you think about my analysis, most people could hold it in their working memory that maybe gas, bloating, and diarrhea might be related to gut dysbiosis. And a practitioner can hold that in their working memory.
But what about these 50 other questions? Maybe you can’t go for very long without eating, and that is a sign of gut dysbiosis. How many of these things can you hold in working memory at once? It’s really, really complex.
When I went to a conference last autumn, the Artificial Intelligence in Medicine Conference down at Dana Point, which overall was very good, I enjoyed my experience there. They where taking questions at one point and I asked the question, ‘Could we use Deep Learning to uncover the root causes of chronic disease?’
And the commentator, he turned to the panel and he said: ‘What do you think? Do you think we really need to understand the root causes, or is it just enough to be able to diagnosis the problem? Because once we have the diagnosis, then there’s the treatment, right? So do we really need to understand the root causes?’ And I just like put my head in my hands.
So it’s frightening because machine learning, obviously it’s so very powerful, but like Tommy said, it’s just a tool. You still have to understand how to use the tool in the most effective way in order to get the result that you’re looking for.
[Damien Blenkinsopp]: Absolutely.
A lot is going on in the world of health, right? Conventional medicine is starting to use big data and train algorithms and so on. But there’s not a lot going on in functional medicine, which is the area and the conferences which we explore more because it’s related to the origins of problems and so on.
(0:48:15) [Damien Blenkinsopp]: Have you seen any other examples of people trying to apply some kind of machine learning? It would be something I’d really love to see more of. I’ve been thinking about it for a little while, that’s why when you guys told me about this I was like, yes!
[Christopher Kelly]: Yeah, so, maybe part of the problem is that everybody thinks that they need big data.
When I was listening to some of these talks that were presented at this conference last year, there were hospitals there who were doing 16,000,000 blood tests per month. That’s probably more than Nourish Balance Thirve will ever do in our lifetime, I think. That truly is big data.
But I think we’ve been able to do something really good without actually having big data. We’ve only had data from 1,000 athletes. So maybe this idea of big data needs to go away. Perhaps we don’t need big data; each individual practitioner already has enough data in order to do something useful.
(0:49:09) [Damien Blenkinsopp]: Well yeah, and especially the ones that have been practicing for 10 years. There’s many of those, and they’ve got a ton of data. I think that’s one of their biggest attributes there, this asset of data they’re sitting on from past patients.
How much data is necessary then? Did you just try this and it worked out?
Or do you think it’s because you were focusing on a niche and there was a tight correlation between the people and that’s why it worked out? Whereas if we do these population studies, I think the view is it can be a bit all over the place, so it can be harder to see those patterns potentially.
[Christopher Kelly]: Yeah. We only had 100,000 total features, which is really quite a small data set. But, there’s no reason why we can’t keep these data sets separate for specific populations.
So let’s say Mark Hyman wants to train models based on his data set. And then Chris Kresser is over here and he sees a lot of thyroid patients so maybe he wants to train on his specific data set. You could still use the same code base, and you could still use the same algorithms.
With the particular model that I’ve used, XGBoost, it doesn’t take that much compute resources in order to train the models. So this is in stark contrast to Deep Learning, for example, where really it’s not possible to do much on your laptop.
You really have to spin up an S3 instance, a Cloud computer with lots of fancy hardware probably made by Nvidia that will allow you to do the training of these algorithms. So computationally it’s very expensive.
That’s not true of the algorithms I’ve used. There’s no reason why people couldn’t just run separate instances of the algorithms on their own personal data sets.
(0:50:52) [Damien Blenkinsopp]: Great. That’s great to hear. I hope this episode inspires a few more people to look into this.
Are there any specific areas you think it should be applied to beyond, or at, or where you think it’s going to be more exciting?
[Christopher Kelly]: That’s a good question. It’s a question I don’t know if I’ve got any good answers [for].
Yeah, so we talked about how there’s so much complexity in the root causes that are causing chronic illness. Tommy has a really good talk that he did on the underlying root causes of insulin resistance.
It’s tempting to believe that the only thing that causes that are refined carbohydrates. And that’s technically true, they maybe do cause insulin resistance. But then there’s endotoxins in the gut, and there’s circadian dysregulation, and there’s loneliness, and there’s other types of stress. There are all these different things, and I feel like it’s going to be almost infinitely complex.
What we really need is some kind of algorithm that could really uncover all of these root causes. Keep everything in working memory at once, and figure this out in a way that no human ever could. I think that Tommy has done a better job on insulin resistance than any other human I’ve met so far. And that includes all the people I’ve interviewed on my podcast. But I have a feeling that a computer might do even better, should someone choose to sit down and apply one in that area.
[Damien Blenkinsopp]: Yeah, that’s great, thanks for that feedback.
I think functional medicine is actually an area where they’re dealing with some of the most complex problems. If you look at things like Lyme disease, where there’s of course a ton of controversy because it’s so complex; people say it doesn’t exist or it exists.
I’d love to see this kind of thing applied to those areas to finally bring some clarity to it and say this is what the machine is coming up with, based on just data. To get past all the opinion and everything which seems to kind of cloud these types of areas.
And chronic fatigue syndrome, you brought that up earlier, that’s another one of these dubious areas where…
[Christopher Kelly]: Yes, that’s death by a thousand cuts, I’m sure.
(0:52:52) [Damien Blenkinsopp]: Yes, excellent.
So is there anything we’ve missed that’s important about your thinking on this subject, or your application and what it’s doing currently?
[Christopher Kelly]: The only thing I wanted to say, another example I saw which I thought was a bad use of this technology, was there’s a paper that came out of the Deep Mind group, which is now part of Google. They did something astonishingly clever, it’s absolutely amazing.
They took a million – a million – hand labeled images of diabetic retinopathies, this is damage done to the eye through high blood glucose. And they created a learning algorithm that would predict diabetic retinopathies better than a human could. And so it’s kind of all amazing, that’s absolutely brilliant.
But then you realize that had the person who’s retina was being scanned done an oral glucose tolerance test with insulin 20 years previous. They maybe could have altered their eating patterns, the food they’re eating and when they’re eating it, and their movement patterns, exercise. Then potentially we could have saved their eyesight, which I think is a much greater win than being able to diagnose them with diabetic retinopathy 20 years later.
So I really wanted people to know about some of the uses and abuses of this type of technology.
[Damien Blenkinsopp]: Yeah so it’s a question of thinking about where they greatest impact is going to be had. And also there’s this question of trying to diagnose the end conditions rather than trying to proactively trying to tackle a problem for the future.
It’s just that mindset, which I don’t know if it’s a lobbying philosophy or how eventually that mindset switch is going to take hold. It seems to just be engrained in the education system, I guess. The systems and everything and the process people are taught on how to approach problems.
[Christopher Kelly]: Yeah, medicine is a really funny beast.
Initially when I started Nourish Balance Thrive, I thought medicine was broken. And then more recently I’ve come to understand that medicine is not broken at all. It’s doing exactly what we designed it to do, which is treat acute and episodic illness.
So if you get hit by a bus or you get an infection, then medicine is really, really good at treating that, for the most part. Where it really falls down is with these diseases of modernity, like diabetes or obesity. Medicine is just not designed to solve those types of problem.
And so we need something completely different, and that is what the Nourish Balance Thrive program is. Now we’ve got a machine learning algorithm that will identify the problem sooner and more easily, but the solution remains the same; you need to move your body, you need to eat appropriately, you need to handle stress appropriately. All of those things.
So, I’m hoping that by doing the cheap and easy diagnosis sooner, it’s going to bring people’s attention to the real problems more easily and sooner so they can rectify them before it really becomes a chronic disease.
[Damien Blenkinsopp]: Yeah, it’s interesting. As you were talking about the system there, I was thinking basically the same problem we have with machine learning we have on a society level, right?
If medicine’s focus, or any organization’s focus, is on something else, it doesn’t matter what you put into it it’s not going to get the ideal outcome. Just like with the machine learning programs; if you set it on the wrong task or the wrong focus, it’ll get the wrong result. And the more money you put into it, the worse it will get.
So it’s interesting like that. Maybe it’s just mimicking humans.
[Christopher Kelly]: Well that’s exactly right. That’s a really good point that you bring up.
Especially with the Deep Learning algorithms, it’s a deep convolutional neural network. It is a model of what happens inside the human skull, literally. That’s how it works. So if you set it on the wrong task, it’s going to get the same wrong answer that humans did.
(0:56:49) [Damien Blenkinsopp]: Yeah, exactly.
OK, so where should someone look first to learn more about this? Are there any good books or presentations on the subject? You’ve mentioned a couple of resources already. Are there any others?
[Christopher Kelly]: No, I don’t think so.
Definitely my two favorite things are, The Master Algorithm by Pedro Domingos–and it was Pedro, when he came on my podcast he said: ‘Oh you should use XGBoost for that.’ And I said, ‘Okay.’
Until then I had been trying to use Deep Learning to solve my problem, not really getting very far. And then Pedro said that one word and he was right. Absolutely amazing book; I absolutely love that book, The Master Algorithm.
And then check out fast.ai, which is Jeremy Howard’s website where he’s teaching these online courses in Deep Learning. And even though I just said that Deep Learning turned out to not be the algorithm that was best for me, Jeremy is an amazing practitioner that will teach you all of the skills that you need in order to become a machine learning practitioner of any type.
So even if you end up using XGBoost or some other algorithm, you’re still going to need all of these other tools that sit around the periphery that will be very valuable no matter what algorithm you use. So that’s fast.ai.
(0:58:00) [Damien Blenkinsopp]: Excellent.
Just to clarify there, if you have no programming background is this still something you can look into and learn more about it and it would be useful, do you think?
[Christopher Kelly]: I would like to say yes, that learning how to program shouldn’t be much more difficult than learning how to speak. It’s really getting that easy.
Python is the programming language that I use and I really don’t think it’s that hard. You can read it just like you can read English. It’s not obfuscated in any way.
Having said that, Jeremy Howard’s course is designed to teach machine learning to people who can already code. I know that some of the people on his classes were coming from a background that was completely different, like mathematics, for example. So maybe they didn’t have any ability to code.
But if you’re smart, you’re going to be able to solve this. Learning is the only skill that really matters.
(0:58:50) [Damien Blenkinsopp]: Cool.
What are the best ways for people to connect with you and learn more about you and your work, you and Tommy’s work at Nourish Balance Thrive? Are you on Twitter, have you got a podcast, or Facebook? Where are you most active?
[Christopher Kelly]: Sure. So what I would really like people to do is come to Damien’s show notes and use the custom link and do the seven minute analysis. Then once you’ve done the analysis, I’m going to follow up on email and send you links to my best podcast episodes on some of the problems that we found.
Tommy has done some fantastic interviews all over the internet–even I have trouble keeping track of them all. So we sat down and thought okay, which are the best things on glucose intolerance? So you’re going to get an email with links to our very best stuff.
I do have my own podcast, a Nourish Balance Thrive podcast, but yeah I would encourage people–you do fantastic show notes, Damien. You do the best show notes I’ve ever seen on the Internet. They’re amazing. So if people are listening and they’ve never seen Damien’s show notes, they should definitely come and check those out.
(0:59:52) [Damien Blenkinsopp]: Thanks, I appreciate it. Of course everything you’ve mentioned in the whole show will be in the show notes, as usual. So, thanks.
Who besides yourself would you recommend to learn about machine learning, or just functional medicine? On your journey, because I know we’ve been to the same space, who do you recommend to check out their work?
[Christopher Kelly]: That’s a really good question. I think my favorite person, the guy that’s been most influential to me, is Bryan Walsh.
Bryan Walsh is a naturopathic doctor from Maryland. If you search online, this is like a hidden gem on the internet. Nobody knows about this. His videos on Youtube, some of them still only have a 1000 views, and I swear most of them are me.
So if you search for Bryan Walsh WellnessFX, which is the blood testing company you’ve probably heard of, you’ll find these videos on Youtube and they are amazing. No one teaches blood chemistry interpretation like Bryan Walsh.
Bryan also has a biochemistry training course for health and fitness professionals called Metabolic Fitness Pro. By the way, I have no financial affiliation with any of this stuff. This is just someone that’s been really, really helpful to me in learning over the years.
Bryan is now on the road teaching weekend seminars on how to do blood chemistry interpretation. I’ve done a whole bunch of training courses. I’ve done FDN, I’ve done Kalish, I’ve done other things, and really Bryan’s stuff is by far the best for me.
[Damien Blenkinsopp]: Excellent. I’ve seen some great stuff of him looking at cortisol dysregulation and adrenal fatigue, if it exists or not.
[Christopher Kelly]: Yeah, exactly, The Artist Formerly Known as Adrenal Fatigue. Bryan has been talking about how that’s nonsense for at least five years, probably longer.
He did a really good interview with Rob Wolfe about this. Really, really good. I love Bryan. He’s been really helpful to me.
(1:01:45) [Damien Blenkinsopp]: Great. Thanks for that.
So I’d love to get to know you a bit more, as well, just in terms of what you actually do to improve your body, and how you use tracking today.
Do you track any metrics or biomakers for your own body on a routine basis? And if you do, why?
[Christopher Kelly]: Well, I still do all of the testing that we talked about. I was at the lab yesterday getting the blood panel that Tommy designed done. I still do urinary organic acid, I still do stool testing.
Already, I said that I ran into some health problems a few year ago. When I did the stool testing, I found a pinworm infection, I found a raging yeast overgrowth. I had almost certainly what most people would call SIBO, although I never did the breath test. I had a belly like a basketball, where I’m still quite lean but for some reason I look like I’m six months pregnant.
So those tests, they were really helpful in uncovering the root causes of my health problems. I took a whole bunch of botanical herbs to solve those issues and that worked really well. So I still do all of that type of testing.
In terms of tracking things on a daily basis, in the end I found it more helpful to track the behavior that leads to the desired outcome. And I can explain that with an example.
So I know that I don’t walk enough–that was one of my problems. I’m a mountain biker, so I pedal lots and lots, and I sit lots and lots when I’m working, but I don’t really walk around too much. The reason I didn’t use to walk around too much was because I found it really, really boring.
So I first thought I would get a FitBit and track the number of steps I’m taking each day. And that was horrifying. I was doing 400 steps a day or something on some days, working from home. Really, really low. And in the end, the solution was not FitBit, it was to get a dog.
A few months ago we got a dog – and I apologize, he barked earlier and I had to kick it out of the room. So not so great for podcasting but great for walking. Now I walk at least an hour a day, and I really enjoy it. It’s really fun.
So maybe sometimes the answer is not to track the number of steps or track whatever it is you’re interested in, but instead insert some interrupt into your life that’s going to lead to the behavior change that gets the desired outcome.
(1:03:56) [Damien Blenkinsopp]: Yeah, that’s really clever, changing your environment like that. Definitely one of the most effective things I’ve found is changing your environment.
So, you’ve had a lot of insight, and it sounds like you’ve made a lot of changes over time. Are there any other more recent changes or things you’re thinking about based on any of these things that you’ve tracked or that have come up? Or are you basically optimized now and you’re quite happy?
[Christopher Kelly]: I am very happy, actually, but I am worried.
One of the reasons I’ve been so good at doing–I say so good. People have told me I’m quite good at doing the podcast, and then also the client calls. It’s because I was so able to relate to the other person’s specific situation, because I had also been through that same situation. I’m worried that I’m losing that now.
But, I do continue to think about it, and maybe there’s something I’m missing, but one of the things I advocate is that it really isn’t that hard. There aren’t that many things to think about.
Diet, which we talked a lot about on the podcast. There’s appropriate management of stress; you’re never going to get away from stress, but you need some way to appropriately manage it. Whatever you do, don’t be lonely. That’s like smoking 15 cigarettes a day. Just because you live in London doesn’t mean you’re not lonely; it’s perfectly easy to be lonely even though…
[Damien Blenkinsopp]: Well especially with all of our devices these days. I think a lot of people choose that route rather than…
[Christopher Kelly]: Exactly. Yeah. So these real relationships are being replaced by Facebook and Twitter and all the rest of this stuff.
And then what else is there, there’s movement, appropriate movement. You need to walk, I think, and you need to occasionally lift heavy things and maybe sprint. And that’s really all there is to it. It’s not that complicated. Or at least it didn’t take me long to say it.
(1:05:35) [Damien Blenkinsopp]: Yeah, excellent.
If you were to recommend just one experiment that someone should try to improve their body – and it could be to improve health, performance, longevity, whatever they’re after or whatever you think is most important – with the biggest payoff, what would that be?
[Christopher Kelly]: Definitely monitoring blood glucose, without a question.
[Damien Blenkinsopp]: Okay. Is that with the CGM or a blood meter? How should they do that, and how…
[Christopher Kelly]: I think most people are not going to have access to the CGM.
I have worn one, personally. Somebody sent me one from New Zealand. So I’m in Santa Cruz in California and I believe that you still need a prescription from a doctor to get one, which is unfortunate, and I’m sure that will change in the future. Somebody sent me one from New Zealand and I did learn a couple of things.
The first was that when I walk, my blood glucose goes down quite surprisingly rapidly, even walking with a three year old girl. So most people would think that’s nothing, that’s not enough exercise to have any impact on anything. But it turns out that it is.
So I can prevent postprandial glucose spikes just by going for a walk with my three year old daughter. And I never would have known that without the continuous monitor. You just wouldn’t know to stick your finger to see that it’s happening.
And then the other thing I found out from the CGM was that intense exercise really, really raises my blood glucose. I don’t know whether its cortisol or what, I haven’t done continuous cortisol monitoring. But when I do intense exercise I can get my blood glucose up to 180 milligrams/deciliter, no trouble at all. So, that kind of makes you aware of the fact that it’s not just the food you put into your mouth that can raise blood glucose.
The place to start is with the finger stick test that everybody has access to. You can go to your local drug store, anywhere in the world, and pick up one of these finger stick tests that I know you’ve talked about on the podcast before. Then just check your blood glucose first thing in the morning.
Optimal evidence based reference range for fasting blood glucose is 83 to 88 milligrams/deciliter. So that’s what one of our models is trying to predict, being out of range. So that’s where I think you should be, and that’s where I am now. It took me a while to get there, but I am in that optimal range.
[Damien Blenkinsopp]: Is that throughout the day? So they can check it at any time away from food or…
[Christopher Kelly]: I’m not sure.
I wouldn’t want to see excursions too far out of that reference range. Once you go above 120 it becomes questionable whether you’re doing yourself any good. Obviously there’s going to be some excursions. If you’re eating any carbohydrate at all then it’s going to go above 100, I would expect.
But I think the fasting value is really interesting because we have some epidemiological data that shows hazard ratios go up significantly once you get above – or below actually – that 83 to 88 milligrams/deciliter in fasting blood glucose.
[Damien Blenkinsopp]: Right. So is that first thing in the morning then?
[Christopher Kelly]: Exactly. As soon as you get up, you stick your finger before you’ve had a chance to move around too much or eat anything.
[Damien Blenkinsopp]: Yeah. I do think sometimes that’s a tricky one for some people, like me, because with my CGM I’ve seen over time really quickly after I wake up I start to get a rise from cortisol.
[Christopher Kelly]: Okay.
[Damien Blenkinsopp]: And so it’s always made me wonder.
I’ve been going to get my blood’s fasting glucose for years, and it didn’t necessarily come back ideal. But then if I look at the whole day, I’m basically in the optimum range all the time. And it’s just this one little spike when I wake up in the morning.
I think I do have some cortisol dysregulation, but I think it’s relatively common as well.
(1:09:11) [Damien Blenkinsopp]: Just on your exercise thing you were talking about. I maybe have a little bit of information for you there. I’ve been testing the Wim Hof method recently.
[Christopher Kelly]: Oh yeah. I am familiar, I’ve tried it.
[Damien Blenkinsopp]: Have you? Oh cool. Well how are you finding it?
[Christopher Kelly]: It’s hyperventilating, and it made my face tingle, and I felt kind of funny. I could see that it was doing something. But yeah I’m not sure what else to say about it.
[Damien Blenkinsopp]: I’ve been through the whole program and taking it pretty seriously. It’s actually helping me with some things which I’ll cover in a later episode.
But I’ve tracked it extensively as well with CGM and things like that, and the breathing, the hyperventilation, whacks up your blood sugar every single time.
Yeah, with a cortisol response. So it makes you wonder if when we get an exercise response, is that due to the breathing? Because when we’re exercising hard, we’re actually breathing really hard as well. Or is it the actual exercise as the actual trigger?
[Christopher Kelly]: It’s a hormetic stressor, is what it is. It has to be.
So some part of your brain thinks that you’re being chased by a tiger, so it’s trying to liberate energy. It’s trying to liquidate your assets. Let’s just get some glucose moving. I bet if you were to measure blood levels of fatty acids, you’d see the same thing; that energy is going up too.
So you’re just liberating your assets so that you can escape from whatever this danger is. But your brain doesn’t know that it’s not really a tiger that’s chasing you, your just doing the Wim Hof thing.
But eventually it leads to you getting stronger. So the same thing happens when I do kettle bell swings, or if I go in the sauna, or if I ride my mountain bike for long enough. And so it’s a hormetic stressor; eventually, hopefully, you get stronger.
[Damien Blenkinsopp]: You do feel it as well, when you first start. You feel this slight anxiety when you’re doing the hyperventilation, and over time that goes away, which fits with your explanation there as well.
(1:11:01) [Damien Blenkinsopp]: But anyway just to come back to your fasting glucose thing, there is that slight variation you have to be aware of, but overall the morning is probably the best time. Is that what you’d advise?
[Christopher Kelly]: Yeah, I think so. It’s overall stability that you’re probably shooting for.
Just because you see, maybe you’re eating a ketogenic diet. And we nearly always see elevated fasting blood glucose with someone who’s been eating a ketogenic diet for a while. But does it mean anything anymore based on my evidence based reference range, because that’s epidemiological data and you can bet your bottom dollar that those people that were in that data set were not eating a ketogenic diet.
So at that point, all bets are off. But for the ketogenic dieter, they’re still achieving overall stability, which may be the most important thing.
[Damien Blenkinsopp]: Right.
Chris, it’s been a really interesting episode. Thanks for all your thoughts and for building your little tool here, which is a great first in functional medicine, I think. So congrats on that, and of course I’ll give the link and everything in the show notes for everyone to follow up with. It’ll be interesting to see what everyone gets from it.
[Christopher Kelly]: Yeah. I’m very excited to know what people think. If you think I’m an idiot and I should stop doing this, please tell me, because otherwise I won’t know.
[Damien Blenkinsopp]: Great Chris. Talk to you again soon.
[Christopher Kelly]: Thanks Damien.