Predicting Nonresponse to Multicomponent Treatment in Fibromyalgia: Development and Validation of a Machine Learning
Rodrigo López-García1,2, Mayte Serrat3, Juan P Sanabria-Mazo4,5
1Escoles Universitaries Gimbernat, Autonomous University of Barcelona, Sant Cugat del Vallès.
Background:
Multicomponent programs combining therapeutic exercise, cognitive-behavioral therapy, and pain neuroscience education demonstrate overall efficacy for fibromyalgia (FM). However, a substantial proportion of patients do not achieve clinically meaningful improvement. This study aimed to identify predictors of nonresponse and to develop a prognostic classifier model.
Methods:
Participants (n=788) from multiple randomized controlled trials received a standardized 12-week multicomponent intervention. This secondary analysis defined nonresponse as a <20% reduction in Fibromyalgia Impact Questionnaire Revised (FIQR) scores. A machine learning approach called least absolute shrinkage and selection operator (LASSO) regularization was used to train a classifier. Model performance was assessed by external validation in a holdout sample, with the area under the curve (AUC) and mean squared error as evaluation indices.
Results:
Higher baseline anxiety (B=0.010), depression (B=0.007), kinesiophobia (B=0.005), and FM severity (B=-0.004), along with lower physical function (B=-0.014), younger age (B=-0.005), and lower body mass index (B=-0.010), were associated with nonresponse. The model demonstrated adequate classification accuracy out-of-sample (AUC=0.657; 95% CI: 0.586-0.728). A prototype calculator incorporating these predictors was developed.
Conclusion:
Psychological, functional, and demographic factors were linked to nonresponse to multicomponent treatment in FM. Although predictive accuracy was limited, these findings support further validation of stratification approaches to inform treatment planning.
