Computer Vision for Classification of Pain Sketch for Amputees Using Custom Convolutional Neural Networks Presented During: View Presentation

PURPOSE: Pain sketches can qualitatively depict neuropathic pain and predict surgical outcomes after targeted muscle reinnervation (TMR) in lower extremity (LE) amputees (1-5). This study aims to develop a machine learning (ML) tool to automate the analysis of these drawings. METHODS: Pain sketches were obtained from adult transtibial or transfemoral amputees who underwent unilateral primary or secondary TMR at a tertiary care center (2021-2024). Sketches were resized, grayscaled, and mirrored. Clinical experts labeled the images based on pain categories: focal (FP), radiating (RP), diffuse (DP), or no pain (NP). A convolutional neural network (CNN) was trained to classify sketches into the categories. Data was split into training, validation, and test sets (70%, 15%, 15%). Binary classification models were built and evaluated using precision, recall, F1-score, and area under the receiver operator characteristic (AUROC). RESULTS: We included 588 sketches from 68 primary and 138 secondary LE TMR patients. The NP vs. DP classification model achieved the highest AUROC (0.799), while other models ranged from 0.587 to 0.760. Unsupervised clustering revealed distinct pain patterns not captured by human categorization. CONCLUSIONS: Computer vision and ML techniques can be used to successfully classify pain sketches in LE amputees, automating the identification of common pain distributions. This tool has potential clinical utility in assisting surgeons with preoperative planning for neuroma-related surgical interventions, particularly for patients undergoing TMR.