Healthcare has a pattern recognition problem

$2.6 trillion was spent on healthcare in the US in 2015. 2016 will see a rise in total spending. We all know that healthcare in America is shockingly expensive — relative to other developed nations and expensive relative to pretty much anything. Why?

It’s expensive because healthcare has a pattern recognition problem. Healthcare is arguably the most information intensive industry in the world yet doctors and nurses have no effective way to find the information they need. This affects the quality of the care and means much of the care that is delivered is unnecessary.

Unnecessary care costs $750 billion each year.
There are 18 million diagnostic errors every year. “Nearly every person will experience a diagnostic error in their lifetime” — Institute of Medicine, 2015.

The most medically complex patients are the most expensive. Nearly half (117 million) of US adults have at least one chronic condition.[1] One in four US adults has multiple chronic conditions. 71% of total health care spending in the United States is associated with care for these Americans with more than one chronic condition.[2] Physicians rarely have the medical equivalent of applicable “manuals” for patients like this.[3]

The disproportionate cost of providing care to these patients is not just because they are more ill, it’s because they are significantly more complex as patients. US Department of Health and Human Services describes the health care system as “primarily organized to provide care on a disease-by-disease basis. So when individuals see a number of specialists, the opportunity for confusion escalates.”[1] To up the complexity, patients with multiple chronic conditions are also typically prescribed multiple medications at once.

Caring for these patients can be like completing four jigsaw puzzles layered each on top of the other. And each puzzle is 3-dimensional. It is extremely difficult for a provider to determine if the constellation of care is the right approach for a patient like their patient.

It can also be difficult to answer the questions of a patient with a single disease. I work with a leading oncologist who treats patients with prostate cancer. Many of his patients expect that he will save their life. They are concerned about the side-effects of the treatment, which can be life-changing and depend on a variety of factors some known (has the patient had pelvic surgery, his age, does he have a history of alcohol abuse) and some not. It’s very difficult for our colleague to answer that question.

Or the common — and pressing — set of questions asked each day in clinics and hospitals across the United States: is this procedure, this medication worth the cost? Is it worth the cost for me? What are my alternatives? What have people specifically like me tried? What worked for them? What didn’t?

I have two friends who are different ages with different medical histories and different needs. They both have multiple sclerosis and were both initially prescribed medication typically prescribed to women decades older, the “average patient” with this disease. One friend just gave birth to a daughter. She’s asked her doctor about the side effects of different treatments and what has been tried for patients like her. She wants to know how her MS will evolve. He’s struggling to find an answer.

When patient data is analyzed, too often the objective is to take a very large group of patients who may have little in common but for a narrowly defined clinical element, measure an “average patient” and then use that average as the model for treating all patients who have the same disease. But this approach doesn’t always work. Not all patients who have multiple sclerosis are the same; some also have hypertension and some have arthritis, some just gave birth to healthy children, some are on an unwieldy regimen of medications. What works for each of these patients is not what will work for every patient.

Every clinic and hospital in the United States collects and stores data that can be used to answer these questions — yet patient-relevant, predictive data is not accessible at the point of care.

Physicians can provider better, faster care when they know what’s been tried with like patients
Physicians can better determine whether their patient is at risk — and for what — when they know what’s happened to similar patients

We started macro-eyes to solve this problem. We’re building a similarity search engine for physicians to use to find information on patients like their patient. Physicians can search to see how similar scenarios have been resolved and determine the most effective course of care for their patient (based on what works for like patients). Seeing data on patients like you means an increased chance that your doctor will be able to determine the best course of care for you - even if he’s not seen someone like you before.

This should not be a revolutionary concept. Similarity search is the very essence of the practice of medicine. When a physician sets out to diagnose or treat the patient in front of them, they run through a database in their head to find whether they’ve seen or read about a patient like their patient. Every patient is unique, but there are recurring patterns. The practice of medicine is the practice of picking up on those patterns.

As a team, macro-eyes is made up of talented people who have solved different kinds of problems: in medicine (at a regional health system) and in healthcare finance and strategy (NYU Langone Medical Center); designing how critical information should be seen and experienced and the systems to work with it (American Express, IMF); large-scale statistical machine learning (MIT, Carnegie Mellon, Microsoft Research). macro-eyes is beginning work and launching implementations at three major academic medical centres.

I studied the history of art at Wesleyan; the foundational problems in the history of art — transformation, style, influence and how meaning is constructed visually — are grounded in deciphering meaningful similarity. I then spent several years trading derivatives, discerning actionable patterns in vast amounts of rapidly-changing data. This is where I first had experience leading the development of technology that identified similar patterns, built to insure that the traders I managed act on trends that we might not have seen. Pattern recognition is a constant across my career.

Pattern recognition is a part of the daily practice of healthcare, but the complexity of the problems to be solved means physicians needs tools to help them find what works.


(1) A chronic condition is a disease like diabetes or coronary heart disease that requires ongoing medical attention and/or limit activities of daily living. Examples include arthritis, diabetes, heart disease and hypertension.

(2) Gerteis J, Izrael D, Deitz D, LeRoy L, Ricciardi R, Miller T, Basu J. Multiple Chronic Conditions Chartbook. AHRQ Publications No, Q14–0038. Rockville, MD: Agency for Healthcare Research and Quality; 2014.

(3) The gold standard of evidence is the care guideline derived from a randomized controlled trial (RCT). RCTs employ narrow patient-inclusion criteria and rarely include patients with multiple chronic conditions. It can be difficult to generalize RCT-derived evidence to real, multidimensionally complex clinical situations. It is widely acknowledged that RCT-derived care guidelines are only applicable to a [very] small proportion of patient encounters. Stanford’s Nigam Shah puts this issue in harsh terms: “only about 4 percent of the time have you got a clinical-trial-based guideline applicable to the patient facing you right now.”

(4) “The Challenge of Managing Multiple Chronic Conditions”