A new research framework wants health AI to stop just snapshotting your vitals and start modeling how they change after something happens to you.
Researchers propose the Physiological World Model, or PWM, an event-conditioned system that links a person's physiological state before an event to the event itself, its context, any intervention, and the trajectory afterward. The core unit is what they call a HumanState Transition Token, a quality-scored record meant to capture that whole before-and-after arc. The paper lays out four capability levels, from basic state representation up to bounded intervention planning, plus four protocols for collecting and validating the data, and six benchmark tasks spanning forecasting, individualized response prediction, and simulating alternative interventions.
Most consumer and clinical health AI today is built to recognize a current state or flag a risk from a single biomarker. PWM's pitch is different: model the physiology like a system that responds to inputs, so a clinician or an app could eventually simulate what happens if you skip a dose, change a workout, or get eight hours less sleep. That reframes wearable data from a dashboard into something closer to a forecasting tool.
It is worth noting this is a framework paper, not a working product; the real test is whether bounded intervention planning holds up once actual event and outcome data get messy. The authors do explicitly flag that prediction is not causal inference, which is the right caveat and also the hardest part to get right in practice.