Evaluation Metrics
AUROC evaluates sensitivity across all decision thresholds, avoiding threshold-tuning bias. Quadratic weighted kappa accounts for ordinal distances between diabetic retinopathy severity stages (0–4).
Review reported results, evaluation context, and the limits of each model.
Team-reported metrics from the September 2026 model overview. See each model card for source limitations and evaluation scope.
| Model | Family | Evaluation Metric | Reported Performance | Clinical Status |
|---|---|---|---|---|
| Wrist Fracture Detection | Onyx | Overall diagnosis AUROC | 0.974 | PREVIEW |
| ChestVision | Onyx | Mean AUROC across 14 conditions | 0.805 | PREVIEW |
| SightDx | Opal | Quadratic weighted kappa | 0.8965 | PILOT |
| BoneScan X | Onyx | Internal validation AUROC | 0.878 | PREVIEW |
| TB Audio Screening | Sona | Preliminary research accuracy | ~84% | RESEARCH |
| Brain MRI Reconstruction | Flux | Not published | Not published | RESEARCH |
AUROC and kappa measure different tasks. These results are not directly comparable.
AUROC evaluates sensitivity across all decision thresholds, avoiding threshold-tuning bias. Quadratic weighted kappa accounts for ordinal distances between diabetic retinopathy severity stages (0–4).
SightDx was evaluated in a team-reported multi-reader pilot at Sachdeva Netralaya, Jammu, measuring agreement with experienced ophthalmologists on fundus photographs.
BoneScan X was evaluated on Stanford's MURA musculoskeletal dataset, achieving 0.878 AUROC in team validation compared to the published 0.929 baseline. Independent multi-center validation details will be published with our technical report.
Original Stanford MURA Research Paper ↗Three research priorities for model development and future healthcare deployment.
Each model has defined input eligibility criteria. The planned platform is designed to reject out-of-scope studies and return them for clinician review.
Planned outputs include calibrated confidence alongside findings, so ambiguous cases can be flagged for senior review.
Attention heatmaps are designed to show which image regions contributed to a finding, such as a fracture line or haemorrhage. They support review and do not prove causation.