Part 2: Predicting heterogeneity of treatment effects to transcranial direct current stimulation for knee osteoarthritis pain and symptoms

BackgroundAssessing the heterogeneity of treatment effects (HTE) is a fundamental aspect of precision medicine, which aims to predict the most optimal treatments based on participant-specific characteristics. This study seeks to identify key predictors of the HTE of transcranial direct current stimulation (tDCS) in individuals with symptomatic knee osteoarthritis (KOA) using machine-learning approaches. MethodsWe performed a secondary analysis of a randomized clinical trial involving 60 participants with symptomatic KOA. These participants underwent 15 daily sessions of 2-mA active tDCS (each session lasting 20 minutes) over a period of three weeks. Initially, we applied group-based trajectory modeling to classify participants into distinct subgroups based on longitudinal KOA pain and symptom patterns from baseline to three months post-intervention to examine differential responses to tDCS. A multi-layer perceptron classifier was then trained to predict the trajectory subgroups using demographic, clinical, and quantitative sensory testing data collected during baseline visits. Feature selection methods, including f-regression, r-regression, and SHapley Additive Explanations (SHAP), were employed to identify the influential features. Additionally, SHAP was used to analyze the correlation and impact of each feature on classification. ResultsParticipants exhibited distinct response patterns to tDCS: high responders (individuals with low initial symptoms showing significant improvement, n = 28) and low responders (individuals with high initial symptoms showing minimal improvement, n = 32) to tDCS. The influential features included conditioned pain modulation (CPM), cold pain intensity, pressure pain thresholds (PPTh) at the medial knee and trapezius, and pain catastrophizing. SHAP analysis revealed that pain catastrophizing was the most influential feature. Additionally, lower CPM, higher cold pain intensity, lower PPTh, and greater pain catastrophizing were associated with a higher likelihood of being classified as low responders. ConclusionOur results contribute to the existing literature, suggesting that factors such as pain catastrophizing, peripheral and central pain sensitization, and individuals endogenous pain-inhibitory capacity should be carefully considered in future tDCS trials.