Patient-reported outcome measures (PROMs) are routinely collected in Swedish specialized pain rehabilitation and capture the multidimensional impact of chronic pain, enabling clinical decision support systems (CDSS) that predict post-treatment outcomes from baseline profiles. Although unsupervised clustering has identified prognostically distinct phenotypes, how such approaches compare with fully supervised machine-learning models remains unclear.
Using registry data from Swedish specialized pain clinics, we compared three prediction strategies for 14 discharge PROMs: a global-mean benchmark, phenotype-mean prediction, and supervised elastic-net and random-forest models. Analyses used repeated five-fold cross-validation with fold-wise imputation and preprocessing to prevent leakage. Performance was quantified using standardized root-mean-square error (sRMSE) with bootstrap confidence intervals, and heterogeneity across phenotypes was examined using per-patient errors.
Phenotype-mean prediction improved over the global-mean benchmark (mean sRMSE 0.898 vs 1.000), while supervised models achieved the lowest error (elastic net 0.820; random forest 0.823). Predictability varied by domain, with physical functioning most predictable and pain intensity and mental health outcomes more challenging.
Baseline PROM clusters capture clinically meaningful heterogeneity but do not replace supervised models. A hybrid CDSS strategy combining clustering for contextualization with supervised predictions appears most promising.
