A central problem in medicine is identifying patients who are at risk of developing a certain disease. By identifying high-risk patients, steps can be taken to hinder or delay the onset of the disease or prevent it altogether. This is an example of predictive analytics at work–using information from previous events to make predictions about the future. There are certain diseases that are particularly popular for prediction research: congestive heart failure, myocardial infarction, pneumonia, and chronic obstructive pulmonary disease are just a few examples of high-mortality, high-cost diseases that benefit from early identification of high-risk patients.

Not only do we care about what diseases will occur in the future, we are also interested in identifying patients who are at risk of requiring high-cost treatments, such as hospital readmissions and doctor visits. By identifying these patients, we can take money-saving steps proactively to reduce the risk of these high-risk treatments, and we can also reward healthcare organizations that do a good job.

This is a broad example with several unknowns to consider. First: what specific event (or disease) are we interested in predicting? Second: what data will we use to make our predictions? Structured clinical data (data organized as tables) drawn from electronic medical records is currently the most popular data source; other possibilities include unstructured data (medical text), medical or x-ray images, biosignals (EEG, EKG), data recorded from devices, or even data from social media. Third: what machine learning algorithm will we use?