Update benchmark_matrix.md

benchmark_matrix.md

@@ -1,68 +1,90 @@
 # Clarus Clinical Stability Benchmark Matrix
 
-This document defines the evaluation structure for the Clarus Clinical Stability Benchmark.
+The Clarus Clinical Stability Benchmark evaluates whether machine learning models can detect **latent stability dynamics** rather than relying on simple feature correlations.
 
-The benchmark evaluates whether models can detect **latent instability across multiple physiological regimes**.
+Each dataset represents a specific **clinical system domain** and **instability mechanism**.
+
+The benchmark also includes **robustness variants** that test whether models remain reliable under incomplete observations and class imbalance.
 
 ---
 
-# Regime Overview
-
-| Stability Regime | Dataset |
-|---|---|
-| Hemodynamic collapse | clinical-hemodynamic-collapse-v0.1 |
-| Sepsis trajectory instability | clinical-sepsis-trajectory-instability-v0.1 |
-| Intervention delay failure | clinical-intervention-delay-failure-v0.1 |
-| Organ coupling cascade | clinical-organ-coupling-cascade-v0.1 |
-| Recovery window detection | clinical-recovery-window-detection-v0.1 |
-| Ventilation–Perfusion instability | clinical-ventilation-perfusion-instability-v0.1 |
-| Hemorrhage compensation collapse | clinical-hemorrhage-compensation-collapse-v0.1 |
-| Electrolyte instability | clinical-electrolyte-instability-v0.1 |
-| Microcirculation instability | clinical-microcirculation-instability-v0.1 |
-| Endocrine instability | clinical-endocrine-instability-v0.1 |
-| Thermoregulation instability | clinical-thermoregulation-instability-v0.1 |
-| Cellular energy instability | clinical-cellular-energy-instability-v0.1 |
-| Respiratory drive instability | clinical-respiratory-drive-instability-v0.1 |
-| Coagulation instability | clinical-coagulation-instability-v0.1 |
-| Hospital operational collapse | clinical-hospital-operational-collapse-v0.1 |
+# Benchmark Matrix
+
+| Dataset | Domain | Instability Mechanism | Missing Data Variant | Class Imbalance Variant |
+|---|---|---|---|---|
+| clinical-hemodynamic-collapse | circulation | pressure collapse | ✓ | ✓ |
+| clinical-perfusion-instability | circulation | microvascular perfusion failure | ✓ | ✓ |
+| clinical-oxygen-transport-instability | circulation | oxygen delivery failure | ✓ | ✓ |
+| clinical-respiratory-drive-instability | respiration | ventilatory control failure | ✓ | ✓ |
+| clinical-acid-base-instability | physiology | buffering collapse | ✓ | ✓ |
+| clinical-renal-filtration-instability | renal | filtration failure | ✓ | ✓ |
+| clinical-fluid-balance-instability | renal | volume dysregulation | ✓ | ✓ |
+| clinical-electrolyte-instability | renal/metabolic | electrolyte imbalance | ✓ | ✓ |
+| clinical-glucose-regulation-instability | metabolic | glucose feedback instability | ✓ | ✓ |
+| clinical-nutritional-metabolic-instability | metabolic | metabolic supply failure | ✓ | ✓ |
+| clinical-hormonal-feedback-instability | endocrine | endocrine feedback instability | ✓ | ✓ |
+| clinical-drug-toxicity-instability | pharmacology | toxic accumulation | ✓ | ✓ |
+| clinical-hemostasis-instability | hematology | coagulation imbalance | ✓ | ✓ |
+| clinical-microvascular-instability | microcirculation | capillary flow heterogeneity | ✓ | ✓ |
+| clinical-neurologic-deterioration-instability | neurology | intracranial perfusion instability | ✓ | ✓ |
+
+---
+
+# Robustness Evaluation
+
+The Clarus benchmark includes robustness variants designed to test whether models truly learn stability dynamics.
 
 ---
 
-# Benchmark Tasks
+## Missing Data Variants
 
-The benchmark supports three evaluation tasks.
+Clinical observations are often incomplete.
 
-## Task 1 — Single-Regime Performance
+To simulate this, trajectory datasets may include variants where observations are missing.
 
-Train and test on the same dataset.
+Supported variants:
 
-Purpose:
+- **missing t0** — initial observation removed  
+- **missing t1** — intermediate observation removed  
+- **missing t2** — final observation removed  
+- **random missing** — one or more values randomly removed  
 
-Evaluate baseline classification ability.
+These datasets evaluate whether models can infer stability dynamics from **partial trajectories**.
 
 ---
 
-## Task 2 — Cross-Regime Transfer
+## Class Imbalance Variants
 
-Train on one regime and test on another.
+Many real-world systems exhibit **rare instability events**.
 
-Example:
+To test robustness to prevalence shifts, datasets may include variants with altered class distributions.
 
-Train → clinical-hemodynamic-collapse-v0.1  
-Test → clinical-microcirculation-instability-v0.1
+Supported regimes:
 
-Purpose:
+- **balanced (50 / 50)**  
+- **mild imbalance (70 / 30)**  
+- **severe imbalance (90 / 10)**  
+- **extreme imbalance (99 / 1)**  
 
-Measure whether models learn **general stability reasoning**.
+These variants test whether models rely on **true stability reasoning** rather than prevalence heuristics.
 
 ---
 
-## Task 3 — Multi-Regime Learning
+# Benchmark Objective
 
-Train on multiple datasets simultaneously.
+A robust stability model should remain reliable across:
+
+- multiple instability mechanisms  
+- different clinical domains  
+- incomplete observations  
+- rare-event prevalence conditions  
+
+Models that rely on shallow correlations or class frequency will degrade under these evaluation regimes.
+
+---
 
-Evaluate on all datasets.
+# Structural Note
 
-Purpose:
+Each dataset reflects **latent stability geometry expressed through observable clinical proxies**.
 
-Test whether models can learn **shared stability representations across physiological systems**.
+The dataset generator and latent stability rules are not included in the benchmark repositories.