ClarusC64/clinical-recovery-window-sepsis-v1

Clinical Recovery Window Sepsis Detection Overview

This dataset tests whether a model can detect when a clinical system is still recoverable.

In severe infections such as sepsis, patients often pass through a finite period during which intervention can still reverse the trajectory toward collapse. This period is known as the recovery window.

Once the recovery window closes, physiological deterioration becomes increasingly difficult to reverse and the system moves toward irreversible failure.

The benchmark therefore evaluates whether models can detect the boundary between reversible and irreversible system trajectories.

The recovery window concept

A recovery window exists when:

physiological buffers still remain intervention lag has not exceeded system tolerance the trajectory has not crossed the instability boundary

Inside this region:

appropriate intervention can move the system back toward stability.

Outside this region:

even aggressive treatment may fail because the system has already progressed too far toward collapse.

Core system geometry

Each scenario represents a simplified clinical dynamical system described using structural variables.

pressure Overall physiological stress acting on the system.

buffer_capacity Remaining physiological reserve available to absorb stress.

intervention_lag Delay between deterioration and treatment.

coupling_strength Strength of interaction between physiological subsystems.

trajectory_drift Directional movement of the system state toward recovery or collapse.

recovery_distance Distance between the current system state and the recovery basin.

These variables together describe the geometry of recoverability.

Clinical geometry mapping

The structural variables correspond to simplified abstractions of real clinical measurements.

Structural Variable Clinical Interpretation Possible Real Signals pressure Overall physiological stress heart rate, mean arterial pressure, lactate buffer_capacity Remaining physiological reserve organ function markers, oxygen reserve intervention_lag Delay before treatment time to antibiotics, time to fluids coupling_strength Interaction between organ systems inflammatory signaling, organ cross-talk trajectory_drift Direction of physiological change trend in lactate, organ function trajectory recovery_distance Distance to recovery basin improvement across multiple clinical indicators

These variables allow the dataset to represent system-level clinical dynamics rather than isolated measurements.

Prediction target

label_recovery_window

0 = recovery window closed 1 = recovery still possible

The task is to determine whether the system remains inside the region where intervention can still change the trajectory.

Row structure

Each row represents a synthetic clinical scenario.

Columns:

scenario_id pressure buffer_capacity intervention_lag coupling_strength trajectory_drift recovery_distance

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

Evaluation

The scorer reports the following metrics.

accuracy precision recall f1 specificity negative predictive value (npv)

Primary metric recall

Secondary metric f1

Recall is prioritized because missing a recoverable patient is more costly than predicting recoverability too often.

Why this benchmark matters

Many predictive models focus primarily on identifying deterioration.

However, a central clinical question is:

Can this patient still be saved?

Detecting recovery windows enables:

earlier targeted intervention better allocation of clinical resources improved survival outcomes

The benchmark therefore tests whether models can reason about reversibility boundaries in dynamical systems.

Structural note

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

The benchmark is designed to test whether models can detect structural transitions rather than memorizing statistical patterns.

Clarus Stability Geometry Benchmarks

This dataset is part of a broader family of instability geometry probes.

Related datasets include:

clinical-compensation-collapse-sepsis-v1 clinical-fork-point-sepsis-transition-v1 clinical-organ-failure-cascade-v1 clinical-intervention-alignment-sepsis-v1 clinical-recovery-stability-sepsis-v1 clinical-false-stability-sepsis-v1

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

License

MIT