Longitudinal DLQI trajectory patterns and early risk updating in telepharmacy-monitored psoriasis

Background Telepharmacy platforms capture longitudinal patient-reported health-related quality-of-life (HRQoL) data in psoriasis, but reports have emphasized average improvement, which may conceal a minority with persistent burden. Whether machine learning adds value beyond baseline severity, and whether the first follow-up sharpens risk estimation, remain unclear. Objective To characterize 12-month Dermatology Life Quality Index (DLQI) trajectories in a telepharmacy psoriasis cohort, to identify a persistently impaired minority through trajectory-pattern analysis, and to determine where the added value of longitudinal monitoring lies. Methods After curation of the NAVETA telepharmacy dataset, the analytic cohort consisted of 219 patients contributing a total of 831 observations. As the primary analysis, unsupervised Ward clustering of DLQI trajectories (a complete-case subset of 133 of the 219 patients with complete baseline-to-month-6 DLQI) identified exploratory trajectory patterns. A leakage-aware benchmark compared a null rule, a one-predictor baseline-DLQI logistic model, multivariable logistic regression and four ensemble learners for 6-month clinically meaningful improvement (≥4-point reduction) across 100 repeated stratified 80/20 splits, with calibration. For 6-month remission (DLQI ≤ 1), a baseline-DLQI-only comparator and a month-1 updating model were evaluated in the same dynamic cohort. Results Mean DLQI improved from 7.78 to 2.89 over 12 months. Three exploratory trajectory patterns were identified—low-burden early control ( n = 90), high-burden responder ( n = 20) and persistent burden ( n = 23; 17.3%); the retained three-cluster solution showed a silhouette of 0.43, and the last pattern remained impaired at 12 months (mean DLQI 7.31 vs. 1.91 and 1.80; p = 0.009). For improvement, a one-predictor baseline-DLQI logistic model matched the leading ensembles (AUROC 0.923 vs. XGBoost 0.903), and feature-attribution analysis confirmed the dominance of baseline DLQI. For DLQI remission, the baseline-DLQI-only model reached AUROC 0.641, whereas the month-1 DLQI updating model reached AUROC 0.753 with good calibration (intercept −0.02, slope 1.04, ECE 0.021). Conclusions A favourable average conceals a clinically important minority with persistent HRQoL burden that trajectory-pattern analysis can identify. For outcomes anchored to baseline severity, machine learning adds little beyond a simple baseline-DLQI model; the added value of telepharmacy may instead emerge from longitudinal information, with the first follow-up providing an early, well-calibrated update to individual risk. External validation is required.