--- language: en license: mit task_categories: - text-classification tags: - clinical-trials - five-node-cascade - clarus - clarus-v0.9 - intervention-competition - clinical - mof size_categories: - 1K 38.5, WBC > 12 | | hemodynamic_coupling | MAP drift, vasopressor need, capillary refill, shock burden | Hypotension, rising pressor need, poor perfusion | | organ_coupling | Creatinine drift, bilirubin drift, ABG deterioration, urine output | AKI progression, bilirubin rising, oliguria, gas exchange worsening | These five variables define the local stability geometry of the multiorgan failure cascade. # Example intervention competition scenario A typical v0.9 row captures a situation where two or more rescue paths compete. Example: Primary intervention path `source_control_then_perfusion_support_then_dialysis` Secondary intervention path `vasopressors_then_perfusion_support_then_dialysis` Example signals - intervention_alignment_score ≈ 0.82 - intervention_leverage_score ≈ 0.79 - pathway_divergence_margin ≈ 0.07 - intervention_competition_ratio ≈ 0.91 - rescue_window_width ≈ 0.24 Interpretation - multiple plausible rescue paths exist - the primary path has slightly stronger structural leverage - the rescue window is narrowing - selecting the wrong path risks losing recoverability This geometry is exactly what the v0.9 benchmark tests. # Prediction target Label column `label_mof_cascade_boundary_stabilization` Default rule ```python label = 1 if stabilization_success = 1 AND trajectory_shift < -0.10 Relaxed variant label = 1 if stabilization_success = 1 Preferred v0.9 rule label = 1 if stabilization_success = 1 \ AND trajectory_shift < -0.10 \ AND intervention_alignment_score >= 0.60 The v0.9 rule rewards correctly targeted rescue, not merely recovery. Row structure Each dataset row describes system geometry. State variables metabolic_stress physiologic_buffer inflammatory_lag hemodynamic_coupling organ_coupling Trajectory signals drift_gradient drift_velocity drift_acceleration Boundary signals boundary_distance secondary_boundary_distance boundary_competition_ratio Uncertainty signals boundary_uncertainty trajectory_uncertainty intervention_uncertainty Regime transition signals regime_transition_score transition_direction regime_separation_margin transition_uncertainty transition_velocity Intervention competition signals intervention_leverage_score intervention_alignment_score rescue_window_width pathway_divergence_margin intervention_competition_ratio primary_intervention_path secondary_intervention_path intervention_uncertainty pathway_switch_velocity Recovery signals recovery_distance recovery_gradient return_feasibility Perturbation signals perturbation_radius collapse_trigger Delta signals delta_metabolic_stress delta_physiologic_buffer delta_inflammatory_lag delta_hemodynamic_coupling delta_organ_coupling Outcome fields trajectory_shift minimal_intervention_path stabilization_success Label label_mof_cascade_boundary_stabilization Files data/train.csv Training rows including: stabilization_success label column data/tester.csv Evaluation rows excluding: stabilization_success label column scorer.py The scorer evaluates both classification accuracy and intervention reasoning. Binary metrics accuracy precision recall f1 confusion_matrix Primary v0.9 metrics recall_correct_intervention_selection false_effective_intervention_rate Intervention competition diagnostics primary_intervention_path_accuracy secondary_intervention_path_accuracy intervention_alignment_accuracy high_uncertainty_intervention_miss_rate rescue_window_miss_rate intervention_competition_error pathway_switch_misread_rate benchmark_spec.json Machine-readable dataset specification including: schema thresholds signal groups evaluation interface Dataset construction Dataset rows are generated in stages. Stage 1 — baseline cascade state Five-node variables are sampled to represent plausible multiorgan failure states. Stage 2 — trajectory dynamics Drift signals describe movement toward or away from instability. Stage 3 — boundary and regime detection Boundary and transition signals describe proximity to instability basins across the cascade. Stage 4 — intervention competition modelling Multiple intervention paths are generated. For each candidate path: intervention_alignment_score measures regime compatibility intervention_leverage_score estimates trajectory impact pathway_divergence_margin measures separation between paths intervention_competition_ratio captures relative leverage rescue_window_width represents remaining recoverability The best path becomes: primary_intervention_path The second-best becomes: secondary_intervention_path Dataset limitations This dataset models structural intervention geometry, not detailed intervention pharmacology. Limitations include: intervention_alignment_score represents structural compatibility rather than exact treatment efficacy rescue_window_width represents geometric recoverability rather than clock time intervention paths represent simplified bundles of actions cascade coupling is structural and does not capture every bedside timing dependency These abstractions allow intervention competition geometry to generalize across domains. Intended use This dataset is intended for: research on intervention selection under instability evaluating regime-aware reasoning systems testing models under narrowing rescue windows benchmarking control-style decision models studying cascade-aware action selection in multiorgan failure This dataset is not intended for: clinical decision support in patient care treatment recommendation systems medical device deployment regulatory evaluation It is a research benchmark. Clarus ladder context Clarus datasets follow an instability-geometry ladder. v0.1 cascade detection v0.2 trajectory awareness v0.3 cascade forecasting v0.4 boundary discovery v0.5 recovery geometry v0.6 intervention reasoning v0.7 uncertainty-aware intervention v0.8 regime transition geometry v0.9 intervention competition geometry v0.9 marks the transition toward control-layer reasoning benchmarks. Structural note Clarus benchmarks represent instability geometry directly. This dataset frames collapse as competition between: failure pressure intervention leverage rescue window narrowing In a five-node cascade this produces a richer control problem than a simple classifier. Production deployment This benchmark structure helps models distinguish between: a plausible intervention the best intervention a nearly tied but weaker intervention an intervention that becomes obsolete as the rescue window closes This geometry appears in many domains including: multiorgan failure monitoring ICU escalation logic infrastructure failure mitigation supply chain disruption recovery AI system stabilization protocols Enterprise and research collaboration The v0.9 intervention competition structure generalizes across domains where multiple rescue paths compete under instability pressure. Possible extensions include: multiorgan cascade control septic shock escalation monitoring respiratory deterioration modelling infrastructure recovery planning financial crisis intervention sequencing AI system resilience benchmarking