MACHINE LEARNING–BASED PREDICTION OF GRAFT SURVIVAL AND HAIR DENSITY AFTER HAIR TRANSPLANTATION USING MULTIMODAL CLINICAL AND TRICHOSCOPIC DATA

Abstract

Background: Accurate prediction of follicular graft survival and postoperative hair density remains a major challenge in hair transplantation because of interpatient variability, donor follicle characteristics, recipient-site conditions, surgical parameters, and postoperative biological responses. Conventional predictive approaches may inadequately capture nonlinear relationships between clinical variables and trichoscopic features, limiting individualized outcome estimation. This study investigates a machine learning–based framework designed to integrate multimodal clinical and trichoscopic information for predicting hair-transplantation outcomes.

Objective: The primary objective is to develop and evaluate predictive models for graft survival and postoperative hair density and to investigate the potential contribution of integrating clinical characteristics, surgical variables, and trichoscopic features. The study also evaluates model performance using complementary regression and classification metrics to establish a quantitative foundation for personalized outcome assessment.

Methods: The analytical framework included 500 patient records and 23 variables, covering clinical characteristics, transplantation-related parameters, and postoperative outcome measures. Ridge regression and random forest algorithms were evaluated alongside clinical-only and trichoscopy-only predictive approaches. Model performance was assessed using mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R2R^2), accuracy, sensitivity, specificity, F1-score, and the area under the receiver operating characteristic curve (ROC-AUC). For graft-survival classification, a threshold of 85% graft survival was used to distinguish the outcome categories. Regression models were evaluated for continuous graft-survival and hair-density prediction, while classification metrics characterized the ability to identify patients meeting the specified graft-survival threshold.

Results: The ridge regression baseline achieved an MAE of 5.61 percentage points and an RMSE of 7.66 percentage points for graft-survival prediction, with an R2R^2 of 0.295. For postoperative hair-density prediction, the model achieved an MAE of 6.55 hairs/cm², an RMSE of 8.60 hairs/cm², and an R2R^2 of 0.758. The graft-survival classification model achieved a ROC-AUC of 0.776, an accuracy of 72.0%, a sensitivity of 86.2%, a specificity of 52.4%, an F1-score of 0.781, and a Brier score of 0.184. Comparative evaluation showed that the multimodal random forest achieved an MAE of 6.09 percentage points for graft survival and 8.10 hairs/cm² for hair density, whereas the clinical-only random forest achieved corresponding MAEs of 6.14 percentage points and 7.77 hairs/cm². These comparisons indicate that predictive performance differed by outcome and model configuration, emphasizing the importance of rigorous comparative evaluation rather than assuming that multimodal integration necessarily improves every endpoint.

Conclusion: The findings demonstrate the potential of machine learning to quantify postoperative hair-transplantation outcomes and provide a foundation for individualized graft-survival and hair-density prediction. The observed R2R^2 of 0.758 for hair-density prediction indicates stronger model fit than the corresponding graft-survival R2R^2 of 0.295, highlighting differences in the predictability of the two outcomes within the evaluated analysis. The classification ROC-AUC of 0.776 further suggests useful discrimination potential, although the specificity of 52.4% indicates that classification performance requires improvement before clinical application. Further validation using real-world patient cohorts, independent external datasets, standardized postoperative measurements, and prospective multicentre studies is necessary to establish clinical reliability and generalizability. This research provides a quantitative starting point for developing interpretable, patient-centred decision-support systems in precision hair restoration medicine.