Eldan Goldenberg's lab notebook

John Guttag: Shortening the Control Loop in the Management of Chronic Disease

Last week I went to a CSE colloquium at the University of Washington. The speaker was John Guttag, and he came to talk about treating chronic disease as a control problem (abstract), and why this approach helps with the management of many conditions. He started with a general description and justification of his approach, but most of the talk was taken up by two concrete examples—epileptic seizure management and heart failure prediction—which demonstrate the value of the approach in practice.

Much of the general part of the talk was about the shortcomings of typical consultant-patient interactions. The major issues relate to the flow of information: consultants have no choice but to rely on patient self-reports for a lot of their data, and self-reports are infrequent, subjective and affected by a massive recency effect. When the consultant asks the patient how have you been since we last met? the patient typically gives an answer that was more appropriate to how have you been in the past 24-48 hours?, which can miss a lot of important information. It would be helpful to be able to two things in relation to Guttag's view of chronic illness as a control problem: shorten the control loop by making treatment responsive to ongoing feedback between consultant visits, and improve long-term data gathering by storing large amounts of information and using heuristics to direct consultants' attention to manageable subsets of that data.

To drive home the point about shortcomings of the typical medical consultation, Guttag drew an analogy to controlling the temperature of a shower. If my shower is too hot or too cold, I can adjust a control immediately, and get a change in the temperature pretty soon. By contrast, the equivalent to a patient-consultant relationship would if when my shower was too hot I had to call a plumber, who would turn a valve somewhere to reduce the temperature, and set an appointment at some time in the future to find out whether the problem had been solved or not.

The longest section of the talk was spent describing a system for managing epileptic seizures, which is an instance of shortening the control loop. About 35% of epilepsy patients are pharmacoresistant; i.e. the drugs that suppress seizures in the other 65% of patients leave them with frequent and/or severe enough seizures to cause major disruption to their lives. Many of these patients are cured by cerebral resection (either removing a part of the brain or disconnecting the two hemispheres), which is frankly one of the most terrifying medical treatments still in use today, and has many obvious drawbacks.

It turns out that in most of these pharmacoresistant patients, a seizure can be stopped by transient electrical stimulation of the vagus nerve. Patients can have an implant that provides this stimulation, but the problem is determining when stimulation should be applied. 1 in 8 patients can identify seizure symptoms far enough ahead of losing motor control to be able self-administer this treatment, but the rest don't know they're having a seizure until the seizure has progressed to the point that they can't trigger the treatment. For this group—roughly 30% of all epilepsy patients—the current way to prevent seizures is to give the vagus nerve a pulse of stimulation every 30 seconds, which works but involves administering a lot more treatment than is actually necessary.

Guttag's solution was to develop a system that takes continuous brain activity readings (from external electrodes), analyses them, and provides the vagus nerve stimulation when the signs of a seizure are present. An EEG can detect a seizure before the patient notices any symptoms, so to all intents and purposes this system prevents seizures from happening, while artificially stimulating the nervous system a lot less often than every 30 seconds. For most patients the system is triggered within 8 seconds of a seizure starting, and this is before any symptoms are reported.

This system does not work out of the box, because each person's EEG patterns are distinctive. However, in spite of the huge between-patients variation, the signs of a seizure are pretty consistent for any individual patient. The same goes for choice of electrodes: patients are given a standard set of 21 to begin with in, but in each individual only 4-6 of those provide useful information. This allows for a support vector machine to be trained for an individual patient by feeding it EEG data from past seizures. 3 seizures seems to be enough to train the classifier, and then there is a tunable trade-off in operation between speed of response and rate of false positives. Because the side-effects of the treatment [which were not detailed in the talk] are much milder than the effects of a seizure, classification boundaries are always set to allow far more false positives than false negatives, but allowing the system a longer response time still reduces the number of false positives.

Less time was given to the other example in the talk, probably because it's a system that has not been developed as far. It was a system for detecting heart failure, and in terms of Guttag's overall aims this one demonstrated the benefits and difficulties of gathering time-series data. Gathering data like heartbeat patterns is not a major technical problem—a patient can wear chest electrodes while going about normal day-to-day business—and neither is storing it, but the challenge is figuring out what to do with all that data. It quickly becomes too much for a human to look over, so what is needed is an automatic classifier system that highlights regions of interest, so that the specialist can ignore most of the data, but instead of only seeing an arbitrary snapshot of while the patient was in hospital (which is often unrepresentative because people feel uneasy in hospitals, among other things) they can only see the time periods in which something is happening that's worth their attention.

This project sounded very much like work in progress, but Guttag did report that initial tests of an ECG classifier had been able to pick out a 9-heartbeat pattern that preceded heart failure in the majority of patients. Potentially, a system that just watches for this pattern and calls an ambulance as soon as it shows up, could save lives.