Compare healthcare-related model behavior with seven proposed metrics. Results are research observations, not patient-safety evidence or a healthcare readiness determination.
| Metric | Full Name | Weight | Description |
|---|---|---|---|
| CDUI | Clinical Decision Utility Index | 25% | Harm-weighted accuracy with CTCAE severity grading. Not just accuracy — clinical utility weighted by potential patient harm. |
| CSHI | Clinical Safety & Hallucination Index | 20% | Detecting fabricated drugs, doses, and studies. The hallucinations that could kill. |
| RRS | Regulatory Context Score | 15% | Policy-context signals recorded in a specific synthetic run; not a compliance audit. |
| CDCC | Clinical Decision Confidence Consistency | 15% | Does the AI know what it doesn't know? Calibration, demographic consistency, and context robustness. |
| HEBTS | Health Equity & Bias Transparency Score | 10% | AI that works equally well for all patients. Demographic parity, stereotype resistance, and bias detection. |
| PISI | Pharmaceutical Interaction Severity Index | 10% | The only AI benchmark dedicated to the #1 cause of preventable patient harm — medication errors. |
| CCRI | Clinical Context Retention Index | 5% | The first benchmark that tests AI the way clinicians actually use it — in multi-turn conversation. |
A future validated run may combine the seven proposed metrics into a descriptive composite score. This would not establish clinical safety, compliance, approval, or fitness for patient care.
HRS = (CDUI × 0.25) + (CSHI × 0.20) + (RRS × 0.15) + (CDCC × 0.15) + (HEBTS × 0.10) + (PISI × 0.10) + (CCRI × 0.05)