ClarusC64/clinical-fluid-balance-instability-v0.1

clinical-fluid-balance-instability-v0.1

What this dataset does

This dataset evaluates whether models can detect instability in fluid balance dynamics.

Each row represents a simplified clinical fluid-management scenario observed across three time points.

The task is to determine whether the system remains volume-stable or is moving toward fluid overload instability.

Core stability idea

Fluid instability does not depend on fluid input alone.

A patient may receive significant fluid while remaining stable if renal clearance and diuretic response remain effective.

Conversely, moderate fluid input may produce instability when urine output declines, pulmonary fluid markers rise, renal function weakens, and intervention is delayed.

The dataset tests interaction reasoning across:

• —fluid input trajectory
• —urine output trajectory
• —body weight proxy trajectory
• —lung fluid marker trajectory
• —renal function proxy
• —diuretic response
• —intervention delay

Prediction target

label = 1 → fluid balance instability label = 0 → stable volume trajectory

Row structure

Each row includes:

• —fluid input trajectory
• —urine output trajectory
• —body weight proxy
• —lung fluid marker
• —renal function proxy
• —diuretic response
• —intervention delay

Decoy variables:

• —monitor_noise
• —chart_noise

These variables appear meaningful but do not determine the label alone.

Evaluation

Predictions must use:

scenario_id,prediction

Run:

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

Metrics returned:

• —accuracy
• —precision
• —recall
• —f1
• —confusion matrix

Structural Note

This dataset reflects latent stability geometry through observable proxies.

The generator and latent rule structure are not included.

This dataset is part of the Clarus Stability Reasoning Benchmark.

License

MIT