Telepharmacy Data Reveal Hidden Minority of Psoriasis Patients With Persistent Quality-of-Life Burden

Modern systemic and biologic therapies drive substantial improvement in patient-reported quality of life in psoriasis, and telepharmacy programmes that collect electronic patient-reported outcome measures can document those gains at scale. Yet a new analysis of a large Spanish telepharmacy cohort warns that the average improvement tells only part of the story. Beneath a favourable cohort mean, roughly one in six patients followed a persistently impaired course, remaining markedly burdened throughout a full year of monitoring while their peers converged toward near-remission.

The study, published in Frontiers in Digital Health, analysed data from the NAVETA telepharmacy programme, a multicenter infrastructure operating across 20 hospital pharmacy centres in the Spanish public healthcare system. The researchers assembled an analytic cohort of 219 adults with moderate-to-severe psoriasis who contributed 831 quality-of-life observations recorded between December 2021 and March 2026. Their instrument of choice was the Dermatology Life Quality Index, or DLQI, a validated 0-to-30 scale in which higher scores indicate greater impairment of daily life by skin disease, and a score of 1 or less denotes essentially no effect of the disease on a patient’s life.

On average, the cohort improved substantially. Mean DLQI fell from 7.78 at baseline to 2.89 at twelve months, and the proportion of patients reaching the no-effect threshold of DLQI 1 or less rose from 16.9 percent to 58.3 percent. A linear mixed-effects model confirmed that every follow-up visit, at months 1, 3, 6 and 12, was significantly better than baseline, with the largest gap appearing at one year. By conventional reporting standards, this would be counted as a clear success story for remote pharmaceutical follow-up.

But averages can be deceptive, and the research team deliberately set out to test whether this one was. Using unsupervised hierarchical clustering with Ward linkage on complete individual trajectories, they grouped 133 patients with full baseline-to-month-six DLQI data according to the actual shape of their recovery. The analysis resolved three distinct patterns: a low-burden early control group of 90 patients who started mild and improved quickly; a high-burden responder group of 20 patients who began severely impaired but recovered to near-remission by month six; and a persistent-burden group of 23 patients, 17.3 percent of the clustered sample, whose scores stayed elevated throughout follow-up. The same 23 patients separated as a distinct group across two-, four- and five-cluster solutions, suggesting the finding was not an artefact of the chosen number of clusters.

The contrast between the two high-burden groups is the study’s most striking finding. At baseline they were almost indistinguishable, with mean DLQI scores of 15.35 and 15.04, yet they separated sharply from the first month onward, with the responders falling to 5.20 while the persistent group remained at 12.61. In an exploratory extension to twelve months among 88 patients with available data, the persistent-burden group still averaged a DLQI of 7.31, compared with 1.91 and 1.80 in the other two patterns, and only 31.2 percent of them had reached the no-effect threshold, versus roughly two-thirds of the responder groups. The differences were statistically significant, although the researchers stress that these between-pattern comparisons are exploratory and were not corrected for multiple testing. Contextual variables such as obesity, drug allergy and age did not differ significantly across the patterns, though the proportions of each were numerically higher in the persistent-burden group.

Identifying this hidden minority raises an obvious practical question: could patients on divergent courses have been flagged earlier, and would sophisticated algorithms have helped? To answer this, the team ran a deliberately sceptical benchmark. They predicted six-month clinically meaningful improvement, defined as a reduction of at least four DLQI points, using a ladder of seven models of increasing complexity, from a simple clinical rule through a one-predictor logistic regression to four tree-based ensemble learners including XGBoost and LightGBM. All models were evaluated under identical conditions across 100 repeated stratified 80/20 train-test splits, with no hyperparameter tuning, and calibration was reported alongside discrimination, in line with the TRIPOD+AI reporting standards for clinical prediction models.

The result was a humbling one for machine learning. Because the improvement endpoint is mathematically anchored to baseline severity, a single baseline DLQI value already carried nearly all the predictive signal. The one-predictor logistic model achieved the highest mean AUROC of 0.923, matching or exceeding the ensembles, with XGBoost reaching 0.903. Feature-attribution analysis using SHAP values confirmed that baseline DLQI dominated the importance ranking, with sociodemographic and lifestyle variables trailing far behind. When the researchers removed baseline DLQI and all DLQI-derived quantities entirely, discrimination collapsed to an AUROC of 0.577, barely better than chance. In a prespecified sensitivity analysis restricted to patients with a baseline DLQI of 4 or greater, which reduces the deterministic component of the endpoint, the one-predictor model still reached an AUROC of 0.806. The lesson, the authors argue, is not that machine learning failed, but that for outcomes governed by baseline burden there is little room for complex algorithms to add value.

The picture changed, however, when the researchers turned to longitudinal information. For the more demanding endpoint of six-month DLQI remission, a baseline-only model achieved an AUROC of just 0.641. But a parsimonious model that simply incorporated the month-one DLQI measurement, the earliest reassessment available through telepharmacy, lifted discrimination to 0.753, a mean gain of 0.113 AUROC units, with good calibration: an intercept of −0.02, a slope of 1.04 and an expected calibration error of only 0.021. Notably, the broader RandomForest model that added sociodemographic covariates performed worse than the simple month-one update. In other words, the distinctive value of digital monitoring lay not in elaborate baseline modelling but in the repeated early measurement itself, which plausibly reflects early pharmacological response, adherence and clinical stabilisation, without implying causality.

The authors are careful about what these findings do and do not establish. This was a retrospective secondary analysis of a curated dataset in which treatment-line and medication-level fields were largely missing, so the persistent-burden pattern cannot be attributed to treatment failure, non-adherence or unmet treatment need; the study should be read as a monitoring study rather than a treatment-response study. The clustering solution showed only moderate internal separation, with a silhouette coefficient of 0.43, and the patterns are data-driven groupings rather than established clinical entities. Requiring complete trajectories reduced the sample from 219 to 133 patients, and although completers and non-completers did not differ significantly at baseline, selection bias related to differential follow-up cannot be excluded. Under extreme assumptions about which patients disengaged, the prevalence of the persistent-burden pattern could range from 10.5 percent upward, so the 17.3 percent figure should be read as a cohort-specific internal estimate. All prediction models were internally validated only, and the researchers did not test whether acting on the early-risk signal actually changes outcomes.

Even with those caveats, the study points toward a concrete and testable shift in how remote monitoring is organised. Two DLQI measurements, one at programme entry and one at the first month, were sufficient to update each patient’s probability of six-month remission with useful discrimination and good calibration. That early, well-calibrated estimate could allow telepharmacy services to stratify their caseload, keeping patients on a favourable course under standard remote follow-up while concentrating intensified pharmacotherapeutic attention, adherence verification and earlier therapeutic reassessment on the minority flagged as unlikely to improve. The authors call for external and prospective validation of the trajectory patterns, the month-one updating model and, crucially, the clinical value of acting on the early-risk signal, before any of this reaches routine practice. Pending that validation, the message is clear: report trajectories, not just averages, and let the first follow-up do the predictive work that ever-larger algorithms cannot.

Subject of Research: Longitudinal quality-of-life trajectory patterns and early risk updating in telepharmacy-monitored psoriasis

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

Article References: Mercadal-Orfila, G., Salvador-Pareja, S., Sotos, J. M., Torres Aranda, A. M., Pinteño-Blanco, M., Maestre-Fullana, M. A., Ríos, S., Riera, P. T., & Herrera-Pérez, S. (2026). Longitudinal DLQI trajectory patterns and early risk updating in telepharmacy-monitored psoriasis. Frontiers in Digital Health, 8, Article 1961097. https://doi.org/10.3389/fdgth.2026.1961097

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1961097

Keywords: psoriasis, telepharmacy, DLQI, patient-reported outcomes, machine learning, trajectory analysis, risk prediction, digital health, quality of life, NAVETA, longitudinal monitoring, calibration