ClarusC64/clinical-stability-benchmark

Benchmark Documentation

Core

• benchmark_structure.md
• benchmark_matrix.md
• datasets.md

Evaluation

• evaluation_framework.md
• transfer_matrix.md
• clarus_score.md

Robustness

• missing_data_protocol.md
• imbalance_protocol.md
• robustness_suite.md

Theory

• stability_manifold.md
• stability_topology.md
• stability_mechanisms.md

Results

• baseline_results.md
• leaderboard.md

Clarus Clinical Stability Benchmark

The Clarus Clinical Stability Benchmark evaluates whether machine learning models can detect latent instability in complex clinical systems.

Most tabular benchmarks reward models for learning correlations within a single dataset.
The Clarus benchmark instead evaluates whether models can infer instability from interacting proxy signals across multiple physiological and operational regimes.

Each dataset represents a simplified regime in which instability emerges from multi-variable interaction rather than single-variable thresholds.

---

Benchmark Concept

In real clinical systems, deterioration rarely occurs because one measurement crosses a threshold.

Instead, instability emerges when several components drift simultaneously.

Examples include:

• circulatory compensation failure
• microvascular perfusion loss
• metabolic energy collapse
• respiratory control failure
• endocrine dysregulation
• thermoregulatory breakdown
• coagulation instability
• hospital operational overload

Each dataset exposes a different regime while keeping the underlying structure similar:
instability arises from interacting system signals.

The generative rules that determine the labels are intentionally not published.

Models must infer instability from observable proxies.

---

Included Datasets

Stability Regime	Dataset
Hemodynamic collapse	ClarusC64/clinical-hemodynamic-collapse-v0.1
Sepsis trajectory instability	ClarusC64/clinical-sepsis-trajectory-instability-v0.1
Intervention delay failure	ClarusC64/clinical-intervention-delay-failure-v0.1
Organ coupling cascade	ClarusC64/clinical-organ-coupling-cascade-v0.1
Recovery window detection	ClarusC64/clinical-recovery-window-detection-v0.1
Ventilation–Perfusion instability	ClarusC64/clinical-ventilation-perfusion-instability-v0.1
Hemorrhage compensation collapse	ClarusC64/clinical-hemorrhage-compensation-collapse-v0.1
Electrolyte instability	ClarusC64/clinical-electrolyte-instability-v0.1
Microcirculation instability	ClarusC64/clinical-microcirculation-instability-v0.1
Endocrine instability	ClarusC64/clinical-endocrine-instability-v0.1
Thermoregulation instability	ClarusC64/clinical-thermoregulation-instability-v0.1
Cellular energy instability	ClarusC64/clinical-cellular-energy-instability-v0.1
Respiratory drive instability	ClarusC64/clinical-respiratory-drive-instability-v0.1
Coagulation instability	ClarusC64/clinical-coagulation-instability-v0.1
Hospital operational collapse	ClarusC64/clinical-hospital-operational-collapse-v0.1

Each dataset repository contains: data/train.csv data/test.csv scorer.py README.md

---

Evaluation Protocol

Predictions must follow the format:

scenario_id,prediction

Example:

MC101,0 MC102,1

Evaluation is performed using the scorer located in the dataset repository.

Example:

python scorer.py --predictions predictions.csv --truth data/test.csv --output metrics.json

The --truth path refers to the dataset's local data/test.csv file.

Metrics reported include:

• accuracy
• precision
• recall
• f1
• confusion matrix

---

Benchmark Tasks

The benchmark supports three evaluation settings.

1 Single-Dataset Evaluation

Train and test on the same dataset.

Purpose:

Measure baseline performance within a single stability regime.

---

2 Cross-Regime Transfer

Train on one regime and test on another.

Example:

Train → clinical-hemodynamic-collapse-v0.1
Test → clinical-microcirculation-instability-v0.1

Purpose:

Determine whether models learn general stability reasoning rather than dataset-specific correlations.

---

3 Multi-Regime Training

Train on multiple datasets simultaneously.

Evaluate performance across all regimes.

Purpose:

Test whether models can learn shared stability representations across physiological systems.

---

Dataset Design Principles

The Clarus datasets follow several explicit design rules.

No Single-Feature Dominance

No observable variable strongly predicts the label independently.

Target:

|correlation| < 0.30

---

Interaction-Based Labels

Instability emerges from interactions between multiple variables rather than isolated thresholds.

---

Adversarial Symmetry

Rows with nearly identical values may produce opposite labels.

This prevents trivial heuristics.

---

Decoy Variables

Some variables appear meaningful but do not determine the label independently.

---

Hidden Generative Logic

The dataset generator and rule equations are intentionally not published.

Models must infer instability from proxy signals.

---

Baseline Results

Reference baseline experiments are provided in:

baseline_results.md

These establish approximate difficulty levels for common tabular models.

---

Benchmark Architecture

The benchmark can be interpreted as observing a shared stability manifold through different clinical regimes.

Each dataset exposes a different control system while preserving the underlying concept of instability emerging from interacting signals.

Additional details are provided in:

stability_manifold.md

---

Research Applications

The benchmark supports research into:

• system stability reasoning
• interaction-based tabular learning
• cross-domain generalization
• clinical early warning modeling
• infrastructure and system risk detection

---

Quick Start

Quick Start

This example demonstrates how to evaluate a simple model on one Clarus dataset.

---

1 Install dependencies

Example environment:

pip install pandas scikit-learn

---

2 Load the dataset

train = data/train.csv test = data/test.csv

---

3 Train a simple baseline model

Example using logistic regression:

import pandas as pd from sklearn.linear_model import LogisticRegression

train = pd.read_csv("data/train.csv")

X = train.drop(columns=["scenario_id","label"]) y = train["label"]

model = LogisticRegression() model.fit(X, y)

---

4 Generate predictions

test = pd.read_csv("data/test.csv")

X_test = test.drop(columns=["scenario_id","label"])

pred = model.predict(X_test)

out = pd.DataFrame({ "scenario_id": test["scenario_id"], "prediction": pred })

out.to_csv("predictions.csv", index=False)

---

5 Evaluate predictions

Run the official scorer:

python scorer.py --predictions predictions.csv --truth data/test.csv

The scorer returns:

• accuracy
• precision
• recall
• f1
• confusion matrix

License

MIT