Treating Probabilities

Less than a week ago, I wrote that a medical treatment »might eventually be the result not of one experiment performed on you, but of a million experiments performed on a model of you«.¹ Now biotechnology company Moderna has announced encouraging results from a personalised cancer vaccine. The treatment sounds like a step in that direction: take information from a particular patient, analyse it, then manufacture an intervention specifically for that individual.²

»You’d pay the medical system to keep your probabilities below some threshold«

This isn’t really a vaccine in the familiar sense. After a tumour has been removed and its mutations analysed, the therapy, intismeran autogene, is designed around the specific immune targets (neoantigens) found in that tumour. It uses a temporary genetic message (mRNA) to teach the immune system to recognise and attack those targets. The treatment sits somewhere between a vaccine, a drug, and a piece of personalised computation.

But notice what’s still missing. We need a tumour as the starting point. The treatment is personalised in how it’s made, but not yet in when it intervenes. What if we took the same idea of continuous measurement and personalised intervention and moved it backwards along the disease’s trajectory?

Medicine is typically organised around events. A tumour appears, or a heart attack occurs. We diagnose the event and treat it. But the event is usually the end of a much longer process. The tumour was preceded by a molecular state in which its emergence was becoming more likely. The heart attack was preceded by a cardiovascular state in which it was becoming more likely.

A medical system that continuously measured your biology could update a model of your current state, estimate the probability of different future states, and intervene when one became sufficiently likely. It wouldn’t ask whether you’re ill. It would ask, continually, whether this is the right moment to intervene.

That’s an old mathematical problem. In »optimal stopping«, an agent observes a changing system and decides whether to continue waiting or act. The decision depends on the probability of future states, the cost of intervention, the cost of waiting, and the consequences of being wrong.³

Imagine being tested every day to receive not a diagnosis but an updated estimate of where your biology is heading. Most days, nothing happens. Occasionally, a trajectory crosses a threshold. An intervention is made. The next measurement tells the system what happened, the model changes, the loop continues. Measure. Model. Predict. Intervene. Measure again.

That’s effectively control theory applied to the human body. Instead of treating a fully formed condition, the system would continually adjust the probabilities of conditions occurring. You’d pay the medical system to keep your probabilities below some threshold.

The incentives are obvious. The patient wants to remain healthy. The state wants to reduce the cost of disease. Insurers want to reduce expensive future events. Pharmaceutical companies might eventually discover that manufacturing lifelong maintenance is a larger market than treating the diseases that maintenance prevents.

The relationship between patient, doctor, and drug would change, too. Gregory Bateson once described a lumberjack, his axe, and a tree as a single system of feedback: the lumberjack acts on the tree through the axe, the tree’s response changes the lumberjack’s next action. The intelligence lies not in the lumberjack, but in the feedback loop.⁴

A continuously managed probabilistic medical system would work in a similar way. The patient generates measurements. The measurements update the model. The model changes the intervention. The intervention changes the patient. The changed patient generates new measurements. The doctor, the algorithm, the pharmaceutical company, and the patient would no longer be separate actors. They would become components of a system trying to maintain a state.

We already think this way about machines. We don’t wait for an aircraft engine to fail before servicing it. We monitor its condition and intervene when the probability of failure becomes unacceptable. We do the same with buildings, reading stains, cracks, leaks, and other small deviations as evidence of a system moving towards failure.

Perhaps medicine is heading in the same direction. Its future may be maintenance: keeping people within a healthy region of their biological state space.

We don’t wait for the roof to collapse before repairing it.

References