Machine learning clinical decision support for interdisciplinary multimodal chronic musculoskeletal pain treatment: prospective pilot study of patient assessment and prognostic profile validation (Preprint)

BACKGROUND Chronic musculoskeletal pain impacts around 20% of people globally; resulting in patients living with pain, fatigue, restricted social and employment capacity, and reduced quality of life. Interdisciplinary multimodal pain treatment (IMPT) programs have been shown to provide positive outcomes and machine learning predictive patient profiles provide promising positive contribution for clinicians and patients to assist with personalized assessment. OBJECTIVE This prospective pilot study aimed to; externally validate machine learning prognostic patient profiles and examine how they may assist clinicians and patients during IMPT assessment; review and consolidate our IMPT outcome framework; assess presentation and interpretation of the patient profiles and develop new profile summary indicators. METHODS Using our previously developed framework of 13 multidimensional machine learning models across 5 clinically relevant domains including activity/disability, pain, fatigue, coping and quality of life; we developed patient prognostic profiles for new IMPT patient assessments in the Netherlands during November 2023 (N=17). The patient profile was consolidated from 13 to 7 outcome measures. New summary indicators were developed including redefined categories to articulate positive, negative and mixed prognostic profiles; an accuracy indicator based on weighted true positive and true negative values, defined as high, medium and low accuracy; and an indicator to highlight profiles with consistently positive or negative outcome measures. The consolidated reporting guidelines for prognostic machine learning modelling studies checklist has been completed for reporting of patient profile development and validation. RESULTS The pilot prognostic patient profiles were assessed from 3 perspectives; 1) using the initially developed 13 outcome framework; indicating 16 positive and 1 negative prognostic profile; consistent with clinician assessment for 13 of 17 new patients (76.5%); 2) using a consolidated profile of 7 outcome measures; achieving profiles consistent with clinician opinion for 12 of 17 (70.1%) patients; 3) using the consolidated 7 outcome framework with extended new prognostic summary indicators; showing 13 positive and 1 negative profile; with 13/14 (92.9%) consistent with clinician assessment; this articulated the remaining 3 patients with mixed positive and negative outcome measures; the prognostic profile was not used in 2 of these cases due to rescheduling and preliminary review showing no need for help with pain rehabilitation; and was consistent with clinician opinion for the remaining patient; providing a total 14/17 (82.4%) prognostic patient profiles consistent with clinician assessment. CONCLUSIONS This prospective machine learning pilot study has provided; further positive external validation of our prognostic patient profile; shown improved clinician consistent validation using a consolidated 7 outcome profile; and provided new summary indicators to better assist clinician and patient interpretation and IMPT decision support.