A Candidate Verification Gate for Reducing False-Positive Staging in a Pressure Ulcer AI Pipeline

Background/Objectives: A four-class stage classifier can assign a pressure ulcer stage to an erroneous detection. We evaluated supervised candidate verification and its rejection costs. Methods: This retrospective challenge-set study used a frozen detector and stage classifier with an existing binary gate threshold of 0.70. Negative evaluation data comprised 1138 Normal-labeled skin images from 347 source photographs and 328 non-pressure-wound images. A stronger source-family audit reduced the original 112-image positive subset to 94 images from 93 families; these still overlapped detector development. Additional experiments compared three source-grouped binary classifiers and paired detector fine-tuning with or without mined negative images. Results: The historical gate reduced skin images with false-positive outputs from 269/1138 (23.6%) to 54/1138 (4.7%). In the corrected positive subset, 106/107 detected lesions were retained; overall lesion recall changed from 107/110 to 106/110. Both historical false-negative lesions remain in the safety review. Newly trained You Only Look Once (YOLO) 26s, ResNet18, and MobileNetV3-small models achieved internal test-positive retention of 97.4–98.5% after separate calibration. Adding mined negatives directly to detector fine-tuning also reduced false-positive outputs. The historical gate accepted 122/147 matched non-pressure-wound candidates. Conclusions: Negative supervision reduced false-positive outputs through both modular verification and detector retraining. The findings do not establish superiority of a separate gate, patient-level generalization, clinical acceptability of missed lesions, or pressure ulcer etiology discrimination.