Multimodal Integration of Gait, Balance, and Infrared Thermography Enhances Machine-Learning Classification of Knee Osteoarthritis: A Cross-Sectional Study.

Abstract

To investigate whether combining gait, balance, and infrared thermography improves discrimination between individuals with knee osteoarthritis (KOA) and healthy controls (HC) compared with single-domain models. Cross-sectional diagnostic study with supervised machine-learning analysis using repeated 10 × 5 cross-validation and paired ROC comparisons. Instrumented assessment at Research Center on Motor Activities (CRAM) University of Catania, in collaboration with the Orthopaedics and Traumatology and the Physical Medicine and Rehabilitation Units of AOU Policlinico "Rodolico - San Marco," University of Catania. Fifty-five adults aged 45 to 80 years were included: 28 participants with clinically and radiographically confirmed KOA and 27 HC. KOA severity was graded using the Kellgren-Lawrence scale. HC had no history of knee pain, injury, or surgery. Not applicable. Spatiotemporal gait parameters, postural stability measures, knee surface temperatures, and classification performance of gait-only, balance-only, movement-only, thermal-only, and multimodal SVM models, assessed by area under the curve (AUC), sensitivity, and specificity. Compared with HC, KOA participants showed slower walking speed (adjusted p = 0.030), reduced gait symmetry (adjusted p = 0.018), and larger sway ellipse area, especially under eyes-closed conditions (adjusted p = 0.003). Thermography showed consistently higher temperatures in KOA, particularly in the lower popliteal fossa (31.11 ± 0.99°C). Among the top-10 models, the thermal-only model achieved perfect discrimination (AUC = 1.000), whereas the multimodal model also showed excellent performance (AUC = 0.988; sensitivity = 0.937; specificity = 0.887). Paired ROC comparisons showed that the multimodal model significantly outperformed gait-only, balance-only, and movement-only models, while not differing from the thermal-only model. KOA is characterized by concurrent gait, balance, and thermal alterations. Although thermography alone showed the highest discriminative performance in this sample, multimodal integration provided a broader and clinically coherent framework for KOA classification.