Development and Validation of a New Qualitative Pain Assessment Model: Preliminary Results of the WORDSforPAIN Project

Pain is a leading cause of disability. Currently no tool returns information on pain based on patients’ narrative. This study aims to develop and validate a pain assessment model based on Artificial Intelligence (AI). A prospective study, funded by Tuscany Health Ecosystem (PNRR – NewGenerationEU), is in progress. Narrative on pain and biopsychosocial variables are collected form 200 adults with spinal pain twice at baseline, at discharge, and some outcomes are also collected at a 6-months phone follow-up. The ontological analysis of the patients' narrative is conducted manually to obtain enough elements to train an AI algorithm. The new model reliability is tested by kappa statistic along with observed agreement, whereas construct validity is explored correlating pain dimensions with biopsychosocial variables. This is a preliminary analysis on 65 subjects. By texts manual annotation, 6 categories, 13 subclasses (second classification level), and 63 topics (third classification level) emerged. Test-retest reliability analysis on 34 stable subjects declaring no/very few changes in their condition showed poor (k = −0,045) to fair (k = 0,388) concordance, but generally thigh observed agreement (55,88% - 94,12%); often, a net imbalance was found between observed agreement on dimensions presence/absence, leading to underestimation of reliability (kappa paradox). Pain dimensions correlate (absolute Kendall’s tau from 0.203 to 0.561) with all explored biopsychosocial variables related to pain, e.g. severity, medication use, disability, and catastrophizing. In conclusion, this approach matches narrative medicine and AI to personalise care, many construct-consistent correlations emerged, but results must be considered carefully because of still limited data.