Development and validation of machine learning models for classifying cancer-related sarcopenia using Kinect-based mixed-reality exercises in breast cancer survivors.

Abstract

Background: Sarcopenia in cancer survivors is often underdiagnosed due to limited access to imaging-based diagnostic tools such as computed tomography (CT) or dual-energy X-ray absorptiometry (DXA). Indirect classification using movement data may offer a practical, scalable alternative. This study aimed to develop and validate machine learning (ML)-based classification models for cancer-related sarcopenia using joint angle data obtained from Kinect-based mixed-reality (KMR) devices, aiming to improve classification accuracy and identify key movement-related predictors.

Methods: Overall, 77 breast cancer survivors (mean age, 48.9±5.4 years) were included based on stage I-III diagnosis, treatment completion ≥6 months prior, no metastasis, low physical activity, and no major comorbidities. Sarcopenia was diagnosed using skeletal muscle index (SMI) (<5.7 kg/m²) and handgrip strength (HGS) (<18 kg). KMR device data were collected during 8 weeks of exercise. After preprocessing, the dataset was randomly split (8:2) for training and testing. Four ML models-support vector machine (SVM), K-nearest neighbor (KNN), random forest (RF), and XGBoost (XGB)-were trained. Five-fold cross-validation was used for tuning, and feature importance was analyzed.

Results: Of the 38 participants in the exercise group included in the final analysis, 12 (31.5%) were initially diagnosed with sarcopenia. After the 8-week KMR device exercise intervention, 3 participants showed recovery from sarcopenia, resulting in 9 (23.6%) remaining classified with the condition. In the test set, the XGB model demonstrated the highest performance, achieving 94.7% accuracy, 91.2% recall, 95.8% precision, 93.4% F1 score, and 96.2% area under the curve (AUC). Feature importance analysis using RF and XGB consistently identified right "knee flexion (right)" as the most influential predictor.

Conclusions: Among ML classification models trained on KMR device joint data, XGB demonstrated the best performance. Right knee flexion emerged as the most influential feature in sarcopenia classification. These findings suggest that KMR device movement analysis may serve as a practical, non-invasive screening tool for sarcopenia, enabling early detection and personalized intervention strategies for breast cancer survivors in both clinical and remote settings.