Like many professions, medical care is sufficiently complex that its practice as a science is also an art. Predicting the outcomes for recipients of medical care may seem like an impossible task, and it might well be for all but the most artful doctors and nurses. But increasingly, computers are helping to do it, accurately and consistently.


Predicting patient outcomes is based on the application of statistical methods to a population, in this case of patients receiving care. By correlating common outcomes with common patient characteristics, predictive algorithms learn to predict outcomes for new patients who share those characteristics.

As with other machine learning tasks, the key is acquiring the relevant characteristics (bits of information) that describe each patient. Previously, this alone would have been a challenging and contentious task, because of the methodological imperative to derive effects from hypothesized causes. Our scientific method has always been based on observation of a given phenomenon, positing a theory or hypothesis for its cause, making predictions based on the hypothesis, testing those predictions through controlled experimentation and observation, and then repeating until the hypothesis has been sufficiently refined to make accurate and reliable prediction.

But thanks again to Moore’s Law, predictive models often no longer need a causal foundation to be reliable or accurate. Without understanding the specific causal chain, algorithms can successfully correlate different characteristics or factors, be they cause, effect, or coincident, through brute force, elegant modelling, and the statistical certitude of large numbers. Predictive analytics generally benefits from increasing the number of records and the level of detail. The more the model knows about each patient, and the more patients it knows, the more accurate its predictions become over time. It’s an example of machine learning because each subsequent case or record improves the model’s predictive ability.

Predicting patient outcomes and diagnosing disease, is increasingly a data analytics challenge. And as with all other machine learning efforts, diminishing costs and improving computational power is enabling ever-higher data resolution, enlarging datasets and honing predictive accuracy. Healthcare as a Service (HaaS) is expected to form the largest segment of the Big Data as a Service market, and valued at $1.09B in 2015.

The potential benefits are huge and varied. From predicting which patients will be least and most likely to return for subsequent care, to uniting divergent symptoms into an accurate diagnosis, to the preventative management of risk factors to which a particular patient is uniquely vulnerable, to quantifying the insurance risk of individual and groups of employees, predictive analytics holds almost unbelievable promise - to improve outcomes and save money.

The capacity to consider finer details about each patient down to the genetic level enables earlier identification and diagnosis of disease, better focused and less destructive treatment. Predicting patient readmission and other outcomes is just the start.

See more examples of Machine Learning in our Everyday Encounters blog series >>