Survival Analysis of Pressure Injury Images Using Deep Learning: Validating the Potential for Predicting Survival Probabilities

Abstract Purpose: This study aimed to analyze pressure injury images and predict hazard rates using deep learning. It offers a fresh perspective in an underexplored area within pressure injury research. Methods: This study included 958 patients, and pressure injury images were obtained at the first visit. To use meaningful image features, pre-trained models were employed during the feature ex-traction phase. In the survival prediction phase, these image features were used as inputs for two deep learning-based survival analysis models (DeepSurv, PC-Hazard) and two statistics-based sur-vival analysis models (CoxPH, Piecewise Exponential). Results: Using the Concordance index to assess survival model performances, we found that the deep learning-based survival analysis models outperformed the statistics-based survival analysis models. Notably, the 'ResNet50 + PC-Hazard' model scored the highest at 0.7636. Conclusion: We confirmed that the deep learning-based survival analysis models performed better in most cases. This indicates that pressure injury images could act as biomarkers, reinforcing the need for more extensive research.