ClarusC64/clinical-five-node-sepsis-cascade-boundary-v0.3

What this repo does

This dataset models sepsis cascade boundary approach using a Clarus five-node coupling framework combined with trajectory and system dynamics.

The goal is to predict whether a patient is approaching the sepsis cascade boundary.

The dataset introduces a dynamic forecasting layer that allows models to reason about motion through the stability manifold rather than relying only on static physiological snapshots.

Core five-node cascade

infectionpressure immunebuffer treatmentlag organcoupling perfusion_stability

These five variables represent the interacting physiological state controlling sepsis stability.

infection_pressure Represents infectious burden such as pathogen load, inflammatory activation, toxin pressure, or escalating host-pathogen stress.

immune_buffer Represents patient reserve including immune resilience, metabolic compensation, and tolerance to rising infectious insult.

treatment_lag Represents the delay between deterioration onset and effective clinical response.

organ_coupling Represents cross-system interaction such as inflammatory spillover, renal involvement, circulatory destabilization, or multi-organ linkage.

perfusion_stability Represents the remaining stability of tissue perfusion, circulatory adequacy, and microvascular resilience under sepsis stress.

The five-node structure models how these variables interact to produce either recoverable dynamics or cascade deterioration.

Trajectory layer

drift_gradient represents the direction of motion in the system state space.

Values near +1 indicate motion toward instability.

Values near −1 indicate motion toward recovery.

This variable captures trajectory alignment with the instability boundary.

Dynamic forecasting layer

Three additional variables describe how the system moves through the stability manifold.

driftvelocity — speed of motion through state space driftacceleration — change in velocity across consecutive time steps boundary_distance — proximity to the instability boundary

Together these variables allow models to estimate how rapidly instability is approaching rather than simply identifying its direction.

This converts the dataset from trajectory detection into dynamic cascade forecasting.

Dynamic variable definitions

drift_velocity

Magnitude of state change between consecutive time steps.

Definition

drift_velocity(t) = ||x(t) − x(t−1)||

Interpretation

Higher values indicate faster movement through the stability manifold.

Lower values indicate slower system evolution.

drift_acceleration

Rate of change of drift velocity across three consecutive snapshots.

Definition

driftacceleration(t) = driftvelocity(t) − drift_velocity(t−1)

where

drift_velocity(t) = ||x(t) − x(t−1)||

Interpretation

Positive values indicate accelerating movement toward instability.

Negative values indicate deceleration or stabilization.

boundary_distance

Weighted metric distance between the current system state and the instability boundary.

Definition

Computed as weighted Euclidean distance from the current state vector to the nearest point on the instability boundary, normalized to the range 0 to 1.

Interpretation

0 indicates the system has reached the cascade boundary.

Lower values indicate minimal remaining stability margin.

Higher values indicate greater separation from collapse.

Prediction target

labelsepsiscascade_boundary

Binary classification.

1 indicates the system is entering the sepsis cascade boundary regime.

0 indicates the system remains recoverable.

Binary simplification note

Real sepsis deterioration unfolds as a continuous physiological process.

This dataset encodes boundary approach as a binary classification problem to simplify model evaluation and benchmarking.

False stability example

The central challenge in this dataset is detecting cases that appear stable when viewed only through the five-node state variables.

Example

infectionpressure 0.44 immunebuffer 0.72 treatmentlag 0.21 organcoupling 0.27 perfusionstability 0.70 driftgradient +0.66 driftvelocity 0.18 driftacceleration +0.08 boundarydistance 0.07 labelsepsiscascadeboundary 1

This row appears relatively safe if only the static state variables are considered.

However:

driftgradient shows motion toward deterioration driftvelocity shows active movement through state space driftacceleration shows increasing momentum boundarydistance shows very little remaining stability margin

This is a false stability case.

The patient appears stable in a static snapshot but is dynamically close to deterioration.

Row structure

scenarioid infectionpressure immunebuffer treatmentlag organcoupling perfusionstability driftgradient driftvelocity driftacceleration boundarydistance labelsepsiscascade_boundary

Files

data/train.csv data/tester.csv scorer.py readme.md

Evaluation

Models are evaluated using binary classification metrics.

accuracy precision recallcascadedetection falsesaferate f1 confusion_matrix

Primary metric

recallcascadedetection

Secondary diagnostic metric

falsesaferate

The primary goal is detecting cascade onset rather than maximizing overall accuracy.

License

MIT

Structural Note

Clarus datasets encode cascade instability through interacting system variables rather than isolated predictors.

Collapse emerges from coupled system dynamics rather than from any single measurement crossing a threshold.

Production Deployment

These datasets support early warning models designed to detect deterioration trajectories before irreversible cascade occurs.

Such models may assist clinical monitoring systems by identifying dynamic instability patterns earlier than threshold-based alerts.

Enterprise & Research Collaboration

The Clarus dataset framework can be applied across multiple domains including clinical medicine, infrastructure monitoring, complex AI systems, and other environments where cascade instability must be detected before boundary crossing.

For dataset expansion, custom coherence scorers, or deployment architecture: team@clarusinvariant.com

Instability is detectable. Governance determines whether it propagates.