Unsupervised learning for prognostic validity in patients with chronic pain in transdisciplinary pain care

Abstract

Chronic pain is not a singular disorder and presents in various forms and phenotypes. Here we show data from a cohort of patients seeking treatment in a transdisciplinary pain clinic. Patients completed a multidimensional patient-reported battery as part of routine initial evaluation at baseline and at each of the four subsequent visits over 1-year follow-up (0, 1, 3, 6, 12 months). The goal of this work was to use unsupervised modeling approach to identify whether patients with chronic pain undergoing transdisciplinary intensive rehabilitation treatment: (1) can be derived based upon self-reported outcome measures at baseline (or before treatment initiation), (2) are clinically validated based on their clinical diagnosis and medication use, and (3) differ in treatment trajectories over 1 year of transdisciplinary treatment. We applied unsupervised clustering on baseline outcomes using nine patient-reported symptoms and examined treatment trajectories. The three-cluster solution was internally validated. Psychiatric diagnosis, chronic back pain-related disability and symptoms severity determined cluster assignment and treatment prognosis. Conversely, clinical pain severity had lesser effect. Furthermore, clusters showed stability over time despite symptoms improvement. The accurate and meaningful subgrouping of the underlying chronic pain phenotypes would greatly enhance treatment and provide personalized and effective pain management.