A selective deep learning framework for pressure injury staging with calibrated confidence and automated clinical documentation

Objectives Pressure injuries are common chronic wounds that require accurate staging to guide management. Deep learning has shown promise for automated staging from clinical photographs, but most prior studies report classification accuracy alone without accounting for prediction uncertainty. We developed and evaluated a selective deep learning framework that calibrates prediction confidence, defers low-confidence cases for clinician review, and generates structured clinical documentation. Methods We trained a ConvNeXt-V2-Tiny classifier on 763 images labeled across four National Pressure Injury Advisory Panel (NPIAP) pressure injury stages. Post-hoc temperature scaling was applied to improve probability calibration. A selective prediction framework deferred low-confidence cases to clinician review. Structured wound documentation was generated by a large language model using NPIAP guideline-based staging definitions. External validation was performed on an independent dataset of 159 pressure injury images. Results The classifier achieved 74.4% overall accuracy, 73.7% balanced accuracy, and a macro-AUC-ROC of 0.929. Temperature scaling reduced test expected calibration error from 0.101 to 0.063. At an 85% confidence threshold, classification accuracy on retained predictions reached 90.8%, while deferral captured 83.3% of all baseline errors. The documentation module generated structured wound notes for all test images. External validation yielded macro-AUC-ROC of 0.809 and macro F1 score of 0.590. At the same 85% threshold, accuracy on retained predictions was 72.7%. Conclusions Selective deep learning with calibrated confidence improved the reliability of automated pressure injury staging by deferring uncertain predictions while generating structured wound documentation. Future work should evaluate prospective clinical deployment to better assess generalizability and workflow integration. Implications for Clinical Practice A selective deep learning system that reports calibrated confidence and defers uncertain cases may support clinicians in pressure injury staging while maintaining human oversight. Automated generation of structured wound documentation may also help standardize reporting and reduce documentation burden.