<p>Uncertainty-Aware Wound Localization and First-Aid Decision Support From Smartphone Images Using an Explainable Multimodal AI Framework</p>

Purpose Automated wound assessment from unconstrained smartphone images remains challenging due to limited annotated data, heterogeneous imaging conditions, and the lack of trustworthy AI mechanisms such as uncertainty handling and interpretable decision support. This study evaluates the feasibility of a safety-aware multimodal framework that integrates computer vision and language-based reasoning to provide structured first-aid guidance under open-world imaging conditions. Methods Two custom datasets were constructed from openly available images for wound segmentation and coarse anatomical region classification. The proposed pipeline integrates YOLOv11-Segmentation for wound localization, a ResNet-50 classifier with Gradient-weighted Class Activation Map visualization for anatomical context, and a lightweight offline Mistral large language model to generate structured first-aid recommendations. An uncertainty-aware routing mechanism implements selective prediction, allowing the system to abstain from automated guidance and request clarification when prediction confidence falls below a predefined threshold. Generated responses are constrained through structured prompting to avoid diagnostic claims and emphasize general first-aid principles. Confidence scores used for routing were not explicitly calibrated. Results On heterogeneous smartphone imagery, YOLOv11-Segmentation achieved an overall mean average [email protected] of 0.771 and mean average [email protected] of 0.685. A U-Net baseline trained on the same dataset achieved a pixel accuracy of 0.9623 and a mean intersection-over-union of 0.7004, providing contextual segmentation benchmarking. For anatomical classification, ResNet-50 achieved 82.11% accuracy and outperformed an EfficientNet-B0 baseline (74.42%). Sensitivity analysis of the routing threshold showed that increasing the confidence threshold reduced automated guidance and increased clarification requests, demonstrating how uncertainty-aware routing regulates language-model usage under ambiguous visual conditions. Evaluation of language-model outputs was limited to structured qualitative analysis due to the lack of standardized quantitative metrics for first-aid guidance. Conclusion These findings demonstrate the feasibility of combining wound localization, anatomical context modeling, and uncertainty-aware language generation within a transparent multimodal workflow for open-world smartphone imagery. The system is not intended for clinical use and has not been validated with clinical experts or external datasets. While the proposed framework illustrates a safety-oriented approach to integrating perception and controlled reasoning, further work is required to evaluate generalizability, calibrate uncertainty estimates, and establish rigorous quantitative and expert-based evaluation of generated recommendations before real-world deployment.