Monday, May 15, 2017

Science Fiction Meets Science Fact: Machine Learning and Precognition

Many science fiction and fantasy stories feature characters who can predict events before they occur. In Minority Report (both the short story and film),"precogs" are able to foresee crime, allowing the police to arrest the perpetrator before the event actually occurs. As of yet, we have no evidence of people with precognition abilities. But the real-life equivalent may already be here.

A study published just over a month ago in Clinical Psychological Science used emergency room data and machine learning to predict suicide attempts:
Due to convention and the limitations of most traditional statistical approaches, clinical psychological science has often attempted to use simple algorithms to solve complex classification problems. This approach can produce statistical significance but has a limited ability to produce clinical significance. For example, as noted earlier, recent meta-analyses on hundreds of studies from the past 50 years indicate that the ability to predict future suicide attempts has always been at near chance levels. The primary reason for this lack of progress is that researchers have almost always used a single factor (i.e., a simple algorithm) to predict future suicide attempts.
They used claims data to identify people who had attempted (but not succeeded in committing) suicide, by specifically looking for self-harm diagnostic codes. Two experts reviewed the resulting 5,543 records to identify actual suicide attempts. Two control groups were used: a random sample of 12,695 patients from the same database who had no history of suicide attempts, and 1,917 individuals from the 5,543 self-harm records that did not appear to have a suicide attempt. They used Python and R to conduct their analyses.

Machine learning outperformed traditional prediction methods (such as regression), and though predictive ability was better closer to the actual attempt (e.g., 7 days before rather than 720 days before), ML still performed far better than chance at all time points. In the table below, you can see the frequencies of true positives (correctly identified as a suicide attempt), false positives (incorrectly identified a control case as a suicide attempt), true negatives (correctly identified as control case), and false negatives (incorrectly identified suicide attempt as control case):

ML correctly classified (as either suicide attempt or control) between 82% and 86% of cases overall, and between 94% and 98% of suicide attempts specifically. And keep in mind, this algorithm used information available in patients' charts, rather than the more in-depth information one could get by speaking to a patient one-on-one. This information could be useful for health care systems, who could routinely use machine learning on patient records to identify patients who may need follow-up or intervention.

No comments:

Post a Comment