ClarusC64/clinical-hysteresis-sepsis-v1

Clinical Hysteresis Sepsis Detection

Overview

This dataset tests whether a model can detect hysteresis in a sepsis-like clinical system.

Hysteresis occurs when the future behavior of a system depends not only on its current state, but also on the path it took to get there. Two systems may appear similar at the present moment while having very different stability properties because one carries unresolved effects from prior stress.

The goal of this benchmark is to determine whether the system exhibits path dependence that materially alters present stability.

Prediction target

label_hysteresis_effect

0 = no meaningful hysteresis effect 1 = system exhibits hysteresis / path dependence

The task is to determine whether prior stress continues to shape current system behavior beyond what is visible in present state variables alone.

Row structure

Each row represents a synthetic clinical scenario.

Columns:

scenarioid pressure buffercapacity couplingstrength trajectorydrift historicalstress recoveryvelocity statereversalgap

Training rows include the label. Tester rows omit the label.

Evaluation

The scoring script reports:

accuracy precision recall f1 specificity negative predictive value (npv)

Primary metric recall

Secondary metric f1

Recall is prioritized because correctly identifying path-dependent fragility is critical for preventing premature assumptions of recovery.

Why this benchmark matters

Clinical systems do not always reset cleanly when surface indicators improve.

A patient may return to values that resemble a healthier state while remaining structurally altered by prior stress. This creates a gap between apparent recovery and true reset.

This benchmark tests whether models can reason about system memory and path dependence in dynamical systems.

Structural note

This dataset exposes system geometry while keeping the generator used to produce the scenarios private.

The goal is to evaluate whether models can detect hysteresis effects rather than memorizing static patterns.

Clarus Stability Geometry Benchmarks

This dataset is part of a broader benchmark family exploring instability and recovery in complex systems.

Related probes include:

clinical-compensation-collapse-sepsis-v1 clinical-fork-point-sepsis-transition-v1 clinical-organ-failure-cascade-v1 clinical-recovery-window-sepsis-v1 clinical-intervention-alignment-sepsis-v1 clinical-recovery-stability-sepsis-v1 clinical-false-stability-sepsis-v1 clinical-instability-margin-sepsis-v1 clinical-intervention-competition-sepsis-v1 clinical-oscillatory-instability-sepsis-v1 clinical-counterfactual-intervention-sepsis-v1 clinical-intervention-timing-sepsis-v1 clinical-perturbation-resilience-sepsis-v1

Together these benchmarks map the lifecycle of instability and recovery in clinical dynamical systems.

License

MIT