PG-GECR: A Provenance-Gated Hybrid Expert System for Auditable Otoscopic Classification and Human-Supervised Referral

This is the authors' abstract. We don't add key points for this paper.

Otoscopic artificial-intelligence studies can be distorted by duplicate images, acquisition-domain shortcuts, small minority classes, and overconfident predictions. This minor revision evaluates PG-GECR as a provenance-gated hybrid expert system for auditable image-level classification and an exploratory selective-referral rule. To align the evidentiary claim with provenance governance, the expert-only ROI branch is now designated the primary classifier; the 37 pseudo-labelled additions and their ROI fusion are retained only as exploratory training sensitivity because their incremental classification benefit was not statistically established. On 104 expert-labelled test images, the expert-only ROI classifier achieved 0.788 accuracy, 0.688 balanced accuracy, 0.677 macro F1, 0.862 macro AUC, and 0.598 MCC. CSOM recall remained 0.286 (2/7; exact 95% CI 0.037-0.710), while high ASOM discrimination remained inseparable from JPEG-ASOM acquisition confounding. A new post-review referral ablation isolated the conditional-latent branch: at the locked ROI threshold, ROI-only confidence/entropy gating retained 70/104 images at 88.6% accuracy and captured 63.6% of observed errors; adding latent-branch agreement/confidence/entropy retained 54/104 images at 94.4% accuracy and captured 86.4% of errors. At matched 51.9% coverage, ROI-only retention achieved 88.9% accuracy and 72.7% error capture versus 94.4% and 86.4% with latent consensus; these post hoc image-level findings are exploratory and do not establish clinical safety, particularly because neither of the two retained CSOM images under the full gate was correctly classified. A global 64-bit perceptual-hash audit of all 2,922 exact-unique auxiliary images found no cross-label near-duplicate component at Hamming distance <=4, so the prior group-aware split required no conflict exclusion. Repeated balanced learning-curve sampling (five independent subset seeds per training size, fixed validation/test sets and model hyperparameters) yielded mean held-out macro F1 from 0.761 +/- 0.021 at 25 images/class to 0.937 +/- 0.007 at 350/class. The principal contribution remains governance and transparent uncertainty; multicentre, device-diverse, patient-level validation is required before clinical use.