Chikitsa

Chest-radiograph AI — overview

An assistive, non-autonomous multi-reader system for adult PA/AP chest radiographs. Every finding ships with a measured operating point and a stated validation basis — no single hero number. Model and dataset details are held in a separate confidential record.

Retrospective validation only · not cleared for clinical use · decision support, confirmed by a qualified radiologist.

Featured panel — operating points

Measured leave-site-out. The shipped point targets ≈90% specificity; a high-sensitivity point is available per finding. PPV is stated at the cohort reference prevalence — these readers are strong rule-outs (high NPV).

FindingAUCSensSpecNPV
Pleural effusion0.910.760.900.95
Pneumothorax0.900.760.900.99
Cardiomegaly0.850.680.900.95
Atelectasis0.840.550.900.98
Normal/abnormal triage0.900.660.900.74

External validation — never-trained cohorts

The deployed models, unchanged, run against radiographs from institutions and scanners absent from all training data. Discrimination (AUC) transfers; thresholds are re-calibrated per site. The cohorts include a US community hospital and an independent multi-site benchmark of ~6,000 studies.

FindingLeave-site-outUS hospitalIndependent benchmark
Pleural effusion0.900.950.92
Cardiomegaly0.900.930.88
Pneumothorax0.880.950.87
Consolidation0.820.960.82
Atelectasis0.820.840.82

Breadth & provenance

Fifteen findings plus an advisory triage gate, validated across 15 public datasets spanning 6 countries and 2 labelling regimes. The encoder is commercially licensed, and a commercial-clean build — trained only on openly-licensed data — is available for customers requiring full data provenance; several readers already run on those clean weights in production. Localisation is provided as advisory overlays for structural findings plus a trained detector returning real bounding boxes for nodules.

Honest limitations

← Summary→ Full datasheet (all findings, breadth, deployment)