Machine Learning


by Guy McKhann, M.D.

January 31, 2017

This is a column from Dana's print publication, Brain in the News.

A recent article on the Vox website addresses “machine learning.” There are many examples of the use of machine learning to handle massive amounts of data (that is, after all, its prime role). In the clinical area, it is hoped that machine learning can decode huge subject databases to determine patterns that might lead to diagnoses or responses to therapy. For example, in analyzing the fMRI scans of subjects with schizophrenia, the findings from 10 patients may not provide useful data. The findings from 1,000 or 10,000 patients, however, might yield specific diagnostic patterns.

To me, as a clinical neurologist, the use of machine learning for prediction of disease activity is an exciting possibility. An obvious use would be in epilepsy. Could we predict when a patient is likely to have his or her next seizure? The data for this prediction could come from an electroencephalogram (EEG)—not the full multi-lead machine in use today, but perhaps one using only two leads (sources of data). Data could be obtained for roughly 30 minutes, three times a day, and used to determine pre-seizure brain patterns. Repeating this process for multiple patients, we could determine whether there are common patterns. 

Clinical investigators are already studying versions of the approach. Brian Litt, at the University of Pennsylvania, has been interested in the spread of seizures: How does one transform from a discrete focus to a more generalized seizure, and can the progression be interrupted by electrical currents? This is the process cardiologists use to block the spread of abnormal electrical activity in the heart. Combining Litt’s approach to intervention with pre-seizure prediction would provide an entirely new approach to seizure management.

Stroke is another clinical problem where prediction would be of great help. For most patients, a stroke is quite an unexpected thing. Many go to bed and awaken with the disabilities of a beginning stroke. Unlike epilepsy, where an EEG can be a source of data, it is not clear to me how to gather information for stroke prediction. Perhaps imaging, such as magnetic resonance angiography, which would show the caliber and flow through arteries of the brain, could be helpful. But this approach would be expensive, and difficult to repeat at shorter intervals. My stroke expert colleagues have their work cut out for them.

The potential to predict seizures is especially meaningful to me because of an incident that occurred some years ago when I had as a patient a young lady who experienced occasional seizures. During a seizure, she would stop and stare, mute and motionless. The entire episode lasted less than a minute, and she was unaware that the seizure had occurred. Medication limited their frequency, according to her boyfriend, who was a medical student.

At one point, they asked if I thought skiing would be a safe enough activity. I hesitated, but gave permission provided she was very careful and wore a helmet. They skied several times without incident, until she ran into a tree. In a desperate attempt to save her, she was brought by helicopter to the Shock Trauma Center at the University of Maryland. I rushed over to see her, and it was clear that she had a devastating head injury—and that she had not been wearing her helmet. A young resident asked, “Who was the dumb jerk who said she could go skiing?” I quietly said, “I was.” She did not survive.