Subtyping Atopic Dermatitis Trajectories Using Digital Patient-Reported Outcomes and Machine Learning
Author(s)
Ying-Ming Chiu, MD.
attending, Tungs' Taichung MetroHarbor Hospital, Taichung, Taiwan.
attending, Tungs' Taichung MetroHarbor Hospital, Taichung, Taiwan.
OBJECTIVES: The clinical course of Atopic Dermatitis (AD) is highly heterogeneous. This pilot study aimed to explore short-term AD trajectories in patients initiating JAK inhibitor therapy, using rich digital patient-reported outcomes (dPROs) and machine learning to uncover predictive lifestyle and psychosocial factors.
METHODS: Although only 26 patients were enrolled, each contributed a dense and multidimensional dataset, including 90 days of repeated dPROs (via the Atopic Dermatitis Control Tool) and extensive baseline information on lifestyle, psychological status, and social context. Hierarchical clustering was used to subtype disease trajectories, and a CatBoost machine learning model was developed to predict subtype membership. An explainable AI (XAI) framework combining SHAP and a large language model (LLM) was applied for interpretation.
RESULTS: Three distinct trajectory subtypes were identified: (i) Severe with Marked Improvement (n=11), (ii) Moderately Severe with Marked Improvement (n=10), and (iii) Moderately Severe with Moderate Improvement (n=5). The CatBoost model achieved 75.0% accuracy and AUC of 0.828 on the test set. Key predictors included stress and mood fluctuations, marital and family status, and exercise frequency. The XAI-LLM framework successfully translated these findings into clinically interpretable insights.
CONCLUSIONS: This study illustrates that even with a small patient cohort, the collection of rich, high-frequency, and multidimensional dPROs enables "big data" analysis and meaningful patient stratification. Our findings support the feasibility of using AI-enhanced analytics to identify lifestyle-linked treatment response patterns in AD, offering a scalable methodology for future personalized care research.
METHODS: Although only 26 patients were enrolled, each contributed a dense and multidimensional dataset, including 90 days of repeated dPROs (via the Atopic Dermatitis Control Tool) and extensive baseline information on lifestyle, psychological status, and social context. Hierarchical clustering was used to subtype disease trajectories, and a CatBoost machine learning model was developed to predict subtype membership. An explainable AI (XAI) framework combining SHAP and a large language model (LLM) was applied for interpretation.
RESULTS: Three distinct trajectory subtypes were identified: (i) Severe with Marked Improvement (n=11), (ii) Moderately Severe with Marked Improvement (n=10), and (iii) Moderately Severe with Moderate Improvement (n=5). The CatBoost model achieved 75.0% accuracy and AUC of 0.828 on the test set. Key predictors included stress and mood fluctuations, marital and family status, and exercise frequency. The XAI-LLM framework successfully translated these findings into clinically interpretable insights.
CONCLUSIONS: This study illustrates that even with a small patient cohort, the collection of rich, high-frequency, and multidimensional dPROs enables "big data" analysis and meaningful patient stratification. Our findings support the feasibility of using AI-enhanced analytics to identify lifestyle-linked treatment response patterns in AD, offering a scalable methodology for future personalized care research.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
CO220
Topic
Clinical Outcomes, Methodological & Statistical Research, Study Approaches
Topic Subcategory
Clinician Reported Outcomes
Disease
Sensory System Disorders (Ear, Eye, Dental, Skin), Systemic Disorders/Conditions (Anesthesia, Auto-Immune Disorders (n.e.c.), Hematological Disorders (non-oncologic), Pain)