Answers are grounded in this paper and cite their sources.
Evaluates a probabilistic timed automaton for predicting sepsis progression, indicating improved early detection.
Background. Every deployed machine-learning sepsis tool raises its alert at or after the moment a clinician would already suspect sepsis. The pre-clinicaltrajectory -- the path through physiological state space before a point-in-time score such as qSOFA or NEWS2 crosses threshold -- has never been given a formal,verifiable model. Point scores are memoryless thresholds; they cannot reason about the dynamics of a trajectory. Methods. We specify pre-clinical deterioration toward sepsis as a probabilistic timed automaton whose locations are coarse physiological regimes, whose clocksmeasure time-in-regime, and whose edges carry transition probabilities estimated from data. We derive a transition-probability estimator from first principles,state a bounded reachability property of the care pathway in probabilistic timed computation tree logic, and evaluate it on the fitted model. We demonstrate thespecification end to end -- regime fitting, transition estimation, reachability evaluation, and lead-time analysis -- on a model-based ICU cohort calibrated topublished Sepsis-3 and NEWS2 population parameters. The released pipeline substitutes credentialed patient records through a single loader without anychange to the model. Results. On the calibrated cohort of 1,200 ICU stays, model selection returned four regimes mapping onto a stable, compensating, early-dysfunction, and overtstructure. The automaton's clock-guarded critical-transition flag preceded NEWS2 threshold crossing by a median of 26.0 hours (95 percent confidence interval25.0 to 28.0; n equals 180 progressor stays with both markers). Discrimination for eventual progression, scored on sustained occupancy of pre-critical regimes,reached an area under the receiver-operating-characteristic curve of 0.91 and an area under the precision-recall curve of 0.88, against 0.35 and 0.42 respectivelyfor NEWS2 evaluated at the same six-hour pre-onset horizon. The timed automaton fit held-out trajectories far better than a memoryless multinomial baseline(held-out log-likelihood of minus 9.75 thousand versus minus 19.1 thousand), and the bounded-reachability probability of reaching an intervened state before overtsepsis rose from 0.15 to 0.43 as the clock bound was relaxed from 2 to 24 hours. The automaton flags transition toward the critical state substantially earlierthan the point-in-time score. Conclusions. A probabilistic timed automaton makes the pre-clinical sepsis trajectory an object that is at once interpretable (explicit regimes and clocks),formally verifiable (bounded reachability), and clinically testable (lead time against standard scores). The contribution is the formal specification, itsverification, and an in silico demonstration that the model behaves as designed. Validation on credentialed ICU data is the natural next step, and the releasedpipeline is built to carry the identical model directly onto such data.
No takes yet. Share an insight, caveat, or question.
Amit Walinjkar (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: