I started college as a math major at a time so long ago that was how one learned computer science and programming. Ah, those bygone days of batch processing punch cards overnight in hopes they would eventually work out your problem on a DEC-10. But I wound up shifting majors and getting into healthcare via graduate school in clinical psychology.
Once I had my license and started my private practice, I fretted that I may misdiagnose an endocrine or other purely biological disorder as depression or something with a psychogenic etiology. So, in response, I went back to my computing-roots and built a very rudimentary algorithm/program that could help me make a more informed differential diagnosis.
That may have been the start of my combining technology with medicine, healthcare, and Personalized practice.
I have been on faculty at two well known medical schools and I can attest that while there was plenty of technology and computing afoot, there was an almost odd absence of machine learning or artificial intelligence lending a hand in clinical practice. In fact, I found no mention of machine learning before 2016 in either The New England Journal of Medicine or the Journal of the American Medical Association, arguably two of top American medical journals.
Of course, these are more so the nascent days of ML/AI in medicine rather than halcyon. Dr. Eric Topol is one of the most well-versed physicians writing in the areas of what artificial intelligence can and cannot (yet) do in healthcare and medical practice. Reviews (“How Algorithms Could Bring Empathy Back To Medicine” in Nature, “How AI Is Humanizing Healthcare” in Mission and “Making Health Care Human Again” in Fortune) of Topol’s latest book, “Deep Medicine” and his related piece on high-performance medicine in Nature Medicine provide poignant examples.
And it seems like the startup landscape is in hyper-growth for AI/ML approaches to medical services. As I have noted in the past, “Rock Health found that funding in just US-based companies in this space have seen a 2010% increase in total investment between 2011 and 2017, to the tune of $98.4 million and via risk capital, aggregated investments reached $7 billion in 2018. CB Insights found that healthcare AI had more than 300 first equity rounds since 2016. Indeed, even non-per se medical companies are also heavily wading into the depths of healthcare and technology—between 2013 and 2017, Apple had filed 54 healthcare related patents while Microsoft filed for 73, and Alphabet submitted a whopping 186 healthcare patents.”
Not without side-effects
While many may think that AI is an impartial, agnostic, bias-free and data-centric tool. It is not, but it may be hard to know. The work of Heather Dewey-Hagborg highlights the situation of hard-to-see, programmed-in biases of the coders and/or in the case of AI and facial recognition algorithms—based on the availability of data/images. Topol himself noted Cathy O’Neil’s finding in her book Weapons of Math Destruction that “many of these models encoded human prejudice, misunderstanding, and bias.” Readers may recall not that long ago IBM’s Dr. Watson’s patients were not doing well as the algorithm was found to have “recommended ‘unsafe and incorrect’ cancer treatments” to the oncologists using it.
When a promising new drug is being developed, it must run the gauntlet of approvals managed by the Food and Drug Association to be sure it is not harmful and that it is effective. The same is not the case for algorithms or even apps. Just because something appears in the App Store is no guarantee that it works, or that it is harm-free for that matter.
Case in point, the standard approach to evaluate predictive algorithm accuracy is via cross-validation, but not all meet statistical significance. Researchers found that “…record-wise, cross-validation often massively overestimates the prediction accuracy of the algorithms… (and) …that this erroneous method is used by almost half of the retrieved studies… (to) predict clinical outcomes.” Needless to say, that is not a good way to practice medicine.
However, there is new breed of apps known as “digital therapeutics “ which “…deliver evidence-based therapeutic interventions to patients that are driven by high quality software programs to prevent, manage, or treat a broad spectrum of physical, mental, and behavioral conditions. Digital therapeutics form an independent category of evidence-based products within the broader digital health landscape, and are distinct from pure-play adherence, diagnostic, and telehealth products.”
Companies like Happify Health or the doctor’s prescription required approach of Pear Therapeutics are pioneering innovation in the blurry areas of apps and evidence-based, third party conducted randomized control trials. And I suspect that is a good thing.
Personalized Precision Medicine
I have written about the Precision Medicine Initiative® (PMI) that former President Obama instituted during his tenure. Through advances in research, technology and policies that empower patients, the goal is to enable a new era of Personalized medicine in which researchers, providers, and patients work together to develop individualized care. It’s a bit “moonshot-ish,” which I like. It’s also very integrative, which I also like. Results may not be as immediate as anyone would prefer, but I think the sorting and sifting of the resulting big data will be aided by AI. I would like to see the combination of large patient outcome registries to be combined with the Federal findings in order to have the best of both worlds, or rather truly personalized medicine.
Our old friend Topol is advising Tempus, a startup that uses big data to accelerate cancer research within a context precision medicine and revolutionize how unstructured data sets (clinical notes, lab reports, pathology images, and radiology scans to capture phenotypic, therapeutic, and outcomes data) are used to personalize and optimize treatments, and include genomic tests to analyze DNA, RNA, and proteomic data to understand a patient’s tumor at the molecular level so they can identify treatment options tailored to each and every patient.
In my current work, I am an advisor and Chief Clinical Officer for a pre-money startup called YubiHealth that holds a primary focus in behavioral health with a personalized medicine approach. We work to reconcile and synthesize a person’s diverse datasets of blood and lab tests, genetics, microbiome, psychological test results, and other inputs that then inform individualized specific treatment decisions on what medications should be optimal, and educate on diet choices, lifestyle change and tracked via an accountability all with measured outcomes and specificity to the individual. It is my belief that there is much promise in utilizing decision making algorithms that can learn from the available big data, but be delivered in a hyper-personalized manner.