A convolutional neural network (InceptionV3 with transfer learning) was developed and externally validated to detect incisional surgical site infections (SSIs) from wound photographs submitted via telemedicine, across multiple surgical specialties.
The model achieved an AUC of 0.91 (95% CI, 0.88–0.93) on internal testing (trained on 4,978 images) and an AUC of 0.82 (95% CI, 0.75–0.90) on external validation (407 images, 95 patients). Decision curve analysis showed net clinical benefit.
- External validation was limited to a single academic center (95 patients, 407 images), which may not reflect diverse real-world settings. - Image labeling by multiple physicians introduced potential interobserver variability in the ground-truth labels. - The study does not report sensitivity/specificity at specific operating thresholds, limiting direct clinical applicability.
A deep learning model can flag wound photos for SSI with good-to-strong discrimination, suggesting a viable automated triage tool for telemedicine wound monitoring programs. Clinical integration should await prospective trials defining the optimal decision threshold before deployment.
Explore related topics
