SP49. Machine Learning-derived Calculator to Predict Reoperations in Limb-sparing Upper Extremity Soft Tissue Sarcoma Reconstructions

Abstract

PURPOSE: Soft tissue sarcomas (STSs) are rare but aggressive malignancies that pose significant challenges to the reconstructive hand surgeon. Despite advancements in radiotherapy and chemotherapy protocols that have made limb salvage with wide resection and reconstruction the standard of care, patients continue to experience high rates of surgical complications and reoperations. In the era of artificial intelligence and precision medicine, there is a critical need for personalized tools to predict outcomes and guide the reconstructive surgeon’s choices. METHODS: Data from 88 upper extremity STS reconstructions from 2016-2022 were collected, including patient demographics, oncological data, surgical data, and dimensions of each STS and the respective limb segment. The dimensions included length (superoinferior), width (mediolateral), thickness (anteroposterior), volume (length x width x thickness) and respective STS-to-limb segment ratio of each measurement. Variables were then ranked using Bayesian permutation factor importance to identify the top candidates for the classifier machine learning model. The most parsimonious Logistic regression and Naive Bayes machine learning models were then developed through forward-backward selection of the top-ranked variables. These models were continuously validated via 5-fold cross-validation tests of area under the curve (AUC), F1 statistic, classification accuracy (CA), precision, recall, and Matthews correlation coefficient (MCC). The best-performing model was used to create a nomogram for individualized risk prediction. RESULTS: The most parsimonious Logistic Regression model outperformed the counterpart naïve Bayes model in predicting post-reconstruction reoperations, achieving an AUC of 0.8, CA of 0.9, F1 statistic of 0.9, precision of 0.9, recall of 0.8, and MCC of 0.237 on 5-fold cross-validation. Based on this model, a nomogram was developed to aid in individualized risk prediction. For instance, a patient with a BMI of 30 (1 point), undergoing resection of a primary STS (30 points) measuring 12.5 cm in length (40 points) and 5 cm in thickness (2.5 points), with pedicled flap reconstruction (4.5 points), accumulates a total of 78 points. According to the nomogram, this score corresponds to an estimated 11% probability of a reoperation post-reconstruction. CONCLUSION: The Logistic regression-derived nomogram can facilitate the prediction of reoperations in upper extremity soft tissue sarcoma reconstruction. In this model, tumor length has the major impact on the predicted probability.

Ähnliche Arbeiten