Related Experiment Video
Updated: May 3, 2026
Identification of Disease-related Spatial Covariance Patterns using Neuroimaging Data
Published on: June 26, 2013
Disease Phenotypes in Refractory Musculoskeletal Pain Syndromes Identified by Unsupervised Machine Learning
Thomas Hügle1, Tiffany Prétat1, Marc Suter1
1University Hospital Lausanne and University of Lausanne, Lausanne, Switzerland.
Objective:
Overlapping chronic pain syndromes, including fibromyalgia, are heterogeneous and often treatment-resistant entities carrying significant socioeconomic burdens. Individualized treatment approaches from both a somatic and psychological side are necessary to improve patient care. The objective of this study was to identify and visualize patient clusters in refractory musculoskeletal pain syndromes through an extensive set of clinical variables, including immunologic, psychosomatic, wearable, and sleep biomarkers.
Methods:
Data were collected during a multimodal pain program involving 202 patients. Seventy-eight percent of the patients fulfilled the criteria for fibromyalgia, 77% had a concomitant psychiatric-mediated disorder, and 22% a concomitant rheumatic immune-mediated disorder. Five patient phenotypes were identified by hierarchical agglomerative clustering as a form of unsupervised learning, and a predictive model for the Brief Pain Inventory (BPI) response was generated. Based on the clustering data, digital personas were created with DALL-E (OpenAI).
Results:
The most relevant distinguishing factors among clusters were living alone, body mass index, peripheral joint pain, alexithymia, psychiatric comorbidity, childhood pain, neuroleptic or benzodiazepine medication, and response to virtual reality. Having an immune-mediated disorder was not discriminatory. Three of five clusters responded to the multimodal treatment in terms of pain (BPI intensity), one cluster responded in terms of functional improvement (BPI interference), and one cluster notably responded to the virtual reality intervention. The independent predictive model confirmed strong opioids, trazodone, neuroleptic treatment, and living alone as the most important negative predictive factors for reduced pain after the program.
Conclusion:
Our model identified and visualized clinically relevant chronic musculoskeletal pain subtypes and predicted their response to multimodal treatment. Such digital personas and avatars may play a future role in the design of personalized therapeutic modalities and clinical trials.
