Prognostics: The Science of Prophecy

In the opening scenes of the sci-fi thriller Gattaca, no sooner isour doomed protagonist, Vincent Freeman, born than we hear his life’s outlook icily summarized: “Neurological condition — 60% probability, manic depression — 42% probability, attention deficit disorder — 89% probability, heart disorder — 99% probability. Early fatal potential, life expectancy — 30.2 years.”

Perhaps it’s fortunate that we’ve yet to achieve this level of predictive precision. But medical prognostics is developing at an intrepid pace.

Prognostics is the art and science of predicting the course of disease.

In medicine, accurately anticipating the course of illness is often as important as arriving at a diagnosis or devising a treatment plan.

For patients, being forewarned of the turns their health may take offers a degree of control and relief from the pains of uncertainty. For doctors, prognoses influence our treatment plans and help us better allocate limited resources (this idea was best captured by Hippocrates the father of medicine, “And he will manage the cure best who has foreseen what is to happen from the present state of matters.”)

So how do doctors predict the future?

Doctors make predictions in broad strokes. We understood in general terms the typical course of a disease or the common effects of a treatment and expect our patients to follow suit.

For instance, a patient with glioblastoma multiforme (GBM), a particularly treacherous brain tumor, may be expected to survive fourteen months. That is, on average, how long other GBM patients have survived.

This approach is notoriously imprecise.

For one, patients and their diseases are infinitely diverse; no two cases are ever identical. Prognoses based on “average outcomes” are therefore, almost by definition, inaccurate for a large portion of patients.

Second, prognostics, more so than any other aspect of medical care, is subject to the personal disposition and character of a physician. “Physicians, like other people, can often be classified as either optimists or pessimists.” wrote Dr. George Dock, one of the first professors of medicine in the United States, “But the practice of prognosis should be cultivated as objectively as any other part of medicine.”

Perhaps most limiting of all is the fact that medical outcomes are shaped by myriad factors, many of which are unknown, unpredictable, and interdependent.

More recently, simple scoring systems, such as the Palliative Performance Scale, in which patients are assessed against a set of simple criteria (for instance the ability to care for oneself), have been used to offer a degree of order to the prognostication process. Nonetheless, these suffer from many of the same limitations and in practice have found modest success.

What then is the alternative?

Recent years have seen the advent of novel approaches to medical prognostication. Tools borrowed from the fields of machine learning and artificial intelligence are enabling unprecedented foresight into the course of disease.

One such tool is Predictive Analytics, a conceptually simple technique to anticipate future events.

In it, computers are used to search historic records for patterns that precede a given event. Recognizing these patterns in the future suggest that the event will recur.

For instance, in an effort dubbed Predictive Policing, the Santa Cruz Police Department used predictive analytics to cut property thefts by 20%. By analyzing thousands of historic police records the department identified common patterns in city crime and was able to deploy officers to the right place, before thefts were committed.

The implication for medical prognostics is obvious.

By using computers to rifle through thousands of patient records, patterns could be found that predict a given clinical event, for instance a heart attack. Recognizing these patterns in future patients can help us anticipate and hopefully preempt such an event.

Researchers in the department of surgery at the University of Louisville, School of Medicine did just that. The team used predictive algorithms to scan thousands of records of patients with melanoma, an aggressive form of skin cancer, and identify attributes predictive of survival. They used these findings to build “Melanoma Calculator” (http://melanomacalculator.com/) a tool that calculates the probability of 5 year survival for any given patient.

Scientists used predictive analytics at the Sylvia Lawry Centre for Multiple Sclerosis (MS) Research to identify patients at risk of disease complications. By comparing new MS patients to a database of over a thousand previous patients, they were able to provide individualized predictions far more relevant and accurate than previously possible.

These examples demonstrate how predictive analytics addresses the shortcomings of traditional prognostics. Individual patient traits are directly taken into account, doctor disposition is irrelevant because the process is automated, and the myriad relevant factors are identified computationally.

Today, these tools are far from perfect. For one, they rely on the presence of large and reliable data sets. These are often unavailable particularly for diseases that are rare or difficult to diagnose. Moreover, the algorithms require that patient information is organized in a structured and computable form. At present, the vast majority of such information is buried deep within loosely organized doctor notes. Lastly, the algorithms themselves are imperfect and suffer from technical shortcomings that limit their accuracy and reliability.

These issues notwithstanding, machine learning is poised to transform medical prognostics and with it the decisions that patients and doctors make.

It might be some time before lifelong prognoses can be doled out in the delivery room. But a Gattacan scene is closer to reality than we might think.

This article was originally written in 2012.

Mutaz Musa, MD, MSc, MBA is a physician, health care consultant, and developer living in NYC.