Internal Validation
Validation Studies
Internal concordance data for each biomarker, with pathologist-annotated reference sets. We believe every AI pathology tool should show its work.
Study Design
Validation Methodology
All internal validation studies used a prospectively defined concordance protocol. Whole-slide IHC images were split from the training corpus — no overlap between training and validation sets.
Each WSI in the validation set received three independent reads from board-certified anatomic pathologists with subspecialty oncologic expertise. The three reads were used to establish a consensus reference (majority vote for categorical outcomes; mean for continuous scores). The AI algorithm was blinded to pathologist reads; pathologists were blinded to AI output.
Concordance was measured using intraclass correlation coefficient (ICC, two-way mixed effects, absolute agreement) for continuous scores, and linearly weighted Cohen's kappa for categorical classification. Sensitivity and specificity were computed at clinically relevant thresholds.
Validation Set Summary
Results
Per-Biomarker Concordance Tables
| Metric | Value | 95% CI | n |
|---|---|---|---|
| ICC (H-score, continuous) | 0.94 | 0.92–0.96 | 380 |
| Weighted Kappa (categorical) | 0.89 | 0.86–0.92 | 380 |
| Overall concordance (0/1+/2+/3+) | 91.8% | 89.2–94.1% | 380 |
| Sensitivity — 3+ positive | 94.2% | 90.6–96.8% | 121 |
| Specificity — 3+ positive | 97.1% | 94.7–98.7% | 259 |
| Test-retest ICC | 0.997 | 0.995–0.999 | 50 |
| Metric | Value | 95% CI | n |
|---|---|---|---|
| Weighted Kappa (TPS categories) | 0.88 | 0.84–0.91 | 290 |
| ICC (continuous TPS %) | 0.92 | 0.89–0.94 | 290 |
| Sensitivity — TPS ≥ 50% | 91.4% | 86.8–94.8% | 105 |
| Specificity — TPS ≥ 50% | 94.1% | 90.2–96.8% | 185 |
| CPS Pearson correlation | 0.89 | 0.86–0.92 | 290 |
| Test-retest ICC | 0.994 | 0.990–0.997 | 50 |
| Metric | Value | 95% CI | n |
|---|---|---|---|
| ICC (global % positive) | 0.91 | 0.88–0.94 | 265 |
| ICC (H-score) | 0.93 | 0.91–0.95 | 265 |
| Categorical concordance (<14% / 14–20% / >20%) | 89.4% | 85.4–92.6% | 265 |
| Hotspot localization concordance | 86.8% | 82.0–90.8% | 265 |
| Nucleus detection recall | 94.2% | 92.5–95.7% | 5,000 cells |
| Test-retest ICC | 0.996 | 0.994–0.998 | 50 |
Validation Disclaimer
Internal validation studies. Data on file. Synthia is not FDA cleared or approved. For research use only / investigational use only. Performance in your laboratory may vary from internal validation results. Results are not intended to replace pathologist review. Clinical use of these scores requires appropriate validation in your specific laboratory environment, patient population, and clinical indication.