A machine learning-based predictive model for stem cell therapy outcomes in plastic surgery.

Objective: Stem cell therapy has emerged as a promising approach in plastic surgery, yet its efficacy varies markedly among individuals and lacks reliable predictive assessment tools. This study aimed to construct and validate a predictive model for assessing the therapeutic efficacy of stem cell therapy in plastic surgery by identifying key influencing factors through clinical data analysis and machine learning. Methods: Patients who underwent stem cell therapy in the Department of Plastic Surgery from June 2021 to July 2024 were retrospectively included and randomly divided into a training set and a validation set at a 7:3 ratio. Baseline clinical data were collected, and independent influencing factors were screened via univariate analysis, followed by multivariate logistic regression and LASSO feature selection in the training set. Three machine learning models-random forest (RF), support vector machine (SVM), and K-nearest neighbors (KNN)-were constructed using Python 3.8.5 and the scikit-learn library, followed by performance validation in the validation set. Results: A total of 620 patients who underwent stem cell therapy were included. In the training set (n = 434), 262 cases (60.37%) showed effective treatment outcomes, while 112 cases (60.23%) were effective in the validation set (n = 186). Multivariate logistic regression revealed that age, disease duration, diabetes history, and cell passage number were independent risk factors for therapeutic efficacy (all P < 0.05), whereas baseline skin score, stem cell dosage, and injection frequency were independent protective factors (all P < 0.05). The AUC values of the RF, SVM, and KNN models were 0.798, 0.770 and 0.723 in training set, and 0.787, 0.761 and 0.708 in validation set, respectively, with the RF model demonstrating superior performance. Conclusion: The machine learning-based predictive model for stem cell therapy efficacy in plastic surgery, constructed through clinical data analysis, exhibits moderate predictive accuracy and may serve as a reference for clinical personalized treatment.