ClarusC64/clinical-escalation-discipline-v0.1

What this dataset does

This dataset tests whether a model can decide when a patient should be escalated rather than simply monitored.

The task is not to identify the sickest patient by a single score.

The task is to decide whether the current pattern requires escalation.

Core stability idea

Escalation depends on more than visible severity.

A patient with a moderate score may need escalation if the trajectory is worsening and treatment response is poor.

A patient with a similar score may be safe to monitor if treatment response is improving and buffer signals are stable.

Prediction target

The label column is binary.

Label 1 means escalate.

Label 0 means monitor.

Row structure

Each row contains:

• —scenario_id
• —news_score
• —heart_rate
• —resp_rate
• —map
• —lactate_trend
• —oxygen_requirement
• —urineoutputtrend
• —treatment_response
• —mentalstatuschange
• —label

treatment_response uses:

• —improving
• —partial
• —poor
• —none

oxygen_requirement uses:

• —0 = room air or minimal support
• —1 = low-flow oxygen
• —2 = high oxygen requirement

mentalstatuschange uses:

• —none
• —mild
• —moderate

Evaluation

Submissions must contain:

scenario_id,prediction
test_001,0
test_002,0
test_003,1

Run:

python scorer.py predictions.csv

Optional truth path:

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

The scorer reports:

Accuracy
Precision
Recall
F1
Confusion matrix
Structural Note

This dataset tests escalation discipline under uncertainty.

It is designed to prevent a simple rule such as "high NEWS score equals escalate" or "normal MAP equals monitor."

The intended reasoning requires attention to trajectory, treatment response, reserve capacity, and emerging instability.

The dataset does not expose the hidden rationale behind each label.

License

MIT