CoolFace
Datasetpublic

ClarusC64/clinical-quad-oxygen-demand-buffer-lag-coupling-respiratory-collapse-v1.1

Clinical Quad Oxygen Demand Buffer Lag Coupling Respiratory Collapse v1.1 What this repo does This dataset evaluates whether a model can select the correct control policy when: multiple respiratory interventions appear viable early signals suggest improvement alternative policies produce better long-term outcomes The task is not prediction. The task is selecting the correct action under uncertainty, feedback, and misleading signal structure. Core… See the full description on the dataset page: https://huggingface.co/datasets/ClarusC64/clinical-quad-oxygen-demand-buffer-lag-coupling-respiratory-collapse-v1.1.

sourceHugging Facemitupdated 6mo agoView on Hugging Face
1likes34downloads
Dataset Card

Clinical Quad Oxygen Demand Buffer Lag Coupling Respiratory Collapse v1.1

What this repo does

This dataset evaluates whether a model can select the correct control policy when:

  • —multiple respiratory interventions appear viable
  • —early signals suggest improvement
  • —alternative policies produce better long-term outcomes

The task is not prediction.

The task is selecting the correct action under uncertainty, feedback, and misleading signal structure.

Core quad

The system is defined by four interacting variables.

  • —oxygen_demand
  • —buffer_capacity
  • —lag_burden
  • —coupling_stress

These variables define system state.

All signals describe how this state evolves under pressure and intervention.

Clinical variable mapping

Quad VariableClinical MeasurementsTypical Indicators
oxygen_demandoxygen requirement, FiO2 burden, work of breathing pressureescalating oxygen need, desaturation pressure, rising ventilatory demand
buffer_capacityphysiological reserve, lung reserve, hemodynamic tolerancepreserved reserve, compensatory space, tolerance to escalation
lag_burdendelayed correction load, unresolved instability debtdelayed intervention, prolonged hypoxemia, accumulated respiratory debt
coupling_stresscross-system destabilization linking lungs, perfusion, and downstream organsV/Q mismatch spillover, hemodynamic stress, organ strain

Prediction target

label_respiratory_collapse

Binary classification.

1 = optimal robust control policy selected 0 = suboptimal or deceptive policy selected

Label logic

text
label = 1 if:
    stabilization_success = 1
    AND trajectory_shift < -0.10
    AND intervention_alignment_score >= 0.60
    AND control_sequence_alignment_score >= 0.60
    AND recovery_consistency_score >= 0.60
    AND policy_regret <= 0.10
    AND policy_robustness >= 0.60
    AND counterfactual_failure_risk <= 0.30
    AND deceptive_signal_score <= 0.50
What v1.1 adds

Earlier versions answer:

where the system is

where it is moving

which intervention may help

whether a control sequence stabilizes it

v1.1 adds:

is this the best available policy

This introduces:

counterfactual comparison

robustness under perturbation

resistance to deceptive signals

Example scenario (realistic numeric row)
scenario_id: resp002

oxygen_demand: 0.89
buffer_capacity: 0.24
lag_burden: 0.73
coupling_stress: 0.77

drift_gradient: 0.62
drift_velocity: 0.79
trajectory_shift: -0.04

intervention_alignment_score: 0.72
control_sequence_alignment_score: 0.66
recovery_consistency_score: 0.54

optimal_policy_score: 0.91
selected_policy_score: 0.59

policy_regret: 0.32
policy_robustness: 0.47
policy_stability_delta: 0.23

local_improvement_score: 0.83
delayed_failure_risk: 0.78
deceptive_signal_score: 0.69
signal_conflict_score: 0.65

short_term_gain_long_term_loss_flag: 1

stabilization_success: 0
label_respiratory_collapse: 0

Interpretation:

the intervention produces early improvement

alignment scores appear acceptable

long-term respiratory control is weak

a superior counterfactual policy exists

deceptive signals mask later collapse

The correct decision is to reject this policy.

Row structure

Each row includes:

system state

trajectory and boundary signals

intervention candidates

control sequence behavior

counterfactual comparisons

policy diagnostics

adversarial signals

outcome fields

Dataset construction

Scenarios are generated by:

sampling respiratory system states across the quad

generating multiple intervention pathways

simulating outcome trajectories

For each scenario:

optimal_policy_score is assigned from the best trajectory

selected_policy_score is assigned from a candidate path

Derived signals:

policy_regret = optimal − selected

policy_robustness = stability under perturbation

policy_stability_delta = terminal outcome difference

Adversarial structure is introduced by:

high short-term improvement

delayed instability

conflicting subsystem signals

misleading alignment patterns

Counterfactual fields are computed during construction.

Files

data/train.csv
Full dataset with labels

data/tester.csv
Same schema without:

stabilization_success

label_respiratory_collapse

scorer.py
Evaluation script

benchmark_spec.json
Formal benchmark definition

dataset_schema.json
Full schema with types and ranges

README.md
This file

Evaluation

Primary metric:

recall_optimal_policy_selection

Secondary metric:

false_robust_policy_rate

Additional diagnostics:

policy_regret_error

policy_robustness_error

policy_stability_delta_error

counterfactual_miss_rate

deceptive_policy_selection_rate

control_sequence_alignment_accuracy

Running the scorer
python scorer.py data/train.csv predictions.csv
Dataset limitations

deceptive_signal_score reflects structural scenario design, not adversarial model input

counterfactual policies are constructed, not exhaustively searched

robustness is simulated

clinical abstraction may omit domain-specific nuance

Intended use

Use for:

benchmarking control decision systems

evaluating policy selection

stress testing models under uncertainty

Not for:

direct clinical decision making

patient-level diagnosis

deployment without validation

standalone decision systems

Structural note

Each version adds one capability:

v0.x: detection and trajectory

v0.6–v0.9: intervention and competition

v1.0: control

v1.1: counterfactual and adversarial evaluation

v1.1 introduces:

multiple valid actions

only one optimal

failure modes that look correct

Production deployment

Applicable to:

clinical systems

infrastructure control

autonomous systems

financial risk

Enterprise and research collaboration

Clarus evaluates system stability.

The focus is:

not what happens next

but whether the chosen action stabilizes the system

License

MIT