A University of Washington statistician, with colleagues from Massachusetts Institute of Technology and Columbia University, have devised a mathematical model that predicts a patient’s possible medical conditions in the future based on the patient’s current and past medical history. The work of Washington’s Tyler McCormick (pictured right), with MIT’s Cynthia Rudin and Columbia’s David Madigan, will appear in an upcoming issue of the journal Annals of Applied Statistics.
The model created by McCormick and colleagues uses an approach similar to Web retailers that mines an individual’s visiting and purchasing behavior, and those of similar visitors, to suggest other items in their inventories. Only in this case, the team mines medical histories rather than purchases or rentals of books, movies, or electronics.
McCormick said that the model is one of the first predictive algorithms of its kind to be used in a medical setting. The Hierarchical Association Rule Model, as his team calls it, shares information across patients who have similar health problems, which improves the predictions when details of a patient’s medical history are sparse. The model can be particularly helpful when records are incomplete, since it can query medical data from patients with similar demographic characteristics and medical histories.
The researchers used for their study the medical records from a multi-year clinical drug trial involving tens of thousands of patients aged 40 and older. The records included the patients’ histories of medical complaints and prescription medications, as well as demographic details, such as gender and ethnicity.
A key issue in constructing the model is the low occurrence of many medical conditions, even among a large set of data records. McCormick and colleagues found that of the 1,800 medical conditions in their data, 1,400 of them occurred fewer than 10 times. The team had to devise a method that did not overlook those 1,400 conditions, while alerting patients who might actually experience those rarer conditions.
The researchers’ statistical modeling technique is based on Bayesian methods, where predictions are revised and gain confidence as evidence is accumulated. “We’re looking at each sequence of symptoms,” says McCormick, “to try to predict the rest of the sequence for a different patient.” For example, if a patient experiences an upset stomach and upper abdominal pain, heartburn could well be the next disorder the patient would encounter.
McCormick says the model can provide physicians “with insights on what might be coming next for a patient, based on experiences of other patients,” and that “also gives a predication that is interpretable by patients.”
Read more:
- Computer Model to Predict Hip Fracture Risk
- Math Model Gauges Potential Tylenol Overdose, Liver Damage
* * *
You must be logged in to post a comment.