Identification and longitudinal evaluation of inhaled medication adherence barriers using clinical natural language processing in asthma: a real-world cohort study

Abstract

Background To develop a rule-enhanced clinical natural language processing (clinical NLP) framework for the automated identification of inhaled medication adherence barriers in asthma, characterize their longitudinal trajectories, and evaluate their associations with longitudinal medication adherence. Methods This real-world longitudinal cohort study included patients with asthma receiving standardized management at a tertiary hospital. Free-text telephone follow-up records from three visits were analyzed using a rule-enhanced clinical NLP framework to identify five adherence barriers: intentional treatment discontinuation, concerns about side effects, medication access barriers, unintentional forgetfulness, and disease denial. Generalized estimating equations (GEE) were used to assess longitudinal associations between adherence barriers and high medication adherence. Complete-case analysis and Firth penalized logistic regression were performed as sensitivity analyses. Results A total of 216 patients were included; evaluable adherence data were available for 209, 194, and 176 patients at Waves 1, 2, and 3, respectively. The primary GEE analysis included 579 observations from 216 patients. The NLP framework achieved an overall accuracy of 97.4% and an F1 score of 0.883. Intentional treatment discontinuation (OR = 0.143, 95% CI: 0.080–0.258), concerns about side effects (OR = 0.330, 95% CI: 0.163–0.666), and unintentional forgetfulness (OR = 0.059, 95% CI: 0.012–0.278) were associated with lower odds of high adherence. Medication access barriers and disease denial were not statistically significant. Intentional treatment discontinuation increased across follow-up waves, whereas concerns about side effects increased mainly at the final wave. The proportion of low adherence increased from 11.5% among patients without identified barriers to 68.8% among those with three or more barriers (p for trend <0.001). Sensitivity analyses supported the associations for intentional treatment discontinuation and unintentional forgetfulness, whereas the association for concerns about side effects was less stable. Conclusion Rule-enhanced clinical NLP may help extract adherence-related information from routine asthma follow-up records. Intentional treatment discontinuation and unintentional forgetfulness were relatively consistent markers of low adherence, whereas findings for concerns about side effects require cautious interpretation. External validation in larger, multi-center populations is needed before broader clinical application.