A Dermatology Model for Predicting Atopic Dermatitis Flares Using Multi-Modal Data

Thesis Abstract

This study addresses the persistent clinical challenge of unpredictable flares in atopic dermatitis (AD) and the need for a robust, data-driven predictive framework that integrates multi-modal data to anticipate flare events. The aim is to develop and validate a dermatology-specific predictive model that synthesizes physiological, environmental, subjective, and genomic signals to forecast AD flares with clinically actionable accuracy. Specific objectives include (1) identifying salient multi-modal predictors of flares from skin biomarker profiles, patient-reported outcomes, lifestyle factors, environmental exposures, and wearable-derived physiological signals; (2) constructing a theoretical framework that integrates these predictors within an evidence-based predictive model grounded in temporal, machine learning, and systems biology perspectives; (3) evaluating the predictive performance of the model against established clinical risk scores using longitudinal real-world data; and (4) assessing model interpretability and potential for clinical deployment through decision-support simulations. A prospective longitudinal design will be employed, enrolling 300 adults with physician-diagnosed AD across two tertiary dermatology centers over 18 months. Data collection will occur at baseline and biweekly intervals, with additional flare-triggered assessments as needed. Instruments include multiplex skin biomarker assays (filaggrin, IL-4, IL-13, TSLP, eosinophil cationic protein), a validated environmental exposure panel (pollutants, humidity, temperature), wearable sensors capturing sleep, skin temperature, transepidermal water loss, and activity, plus daily electronic patient-reported outcome measures (POEM, SCORAD components) and genotyping for AD-associated loci (e.g., FLG variants). Data will be augmented by clinical records of flare episodes, treatment changes, and healthcare utilization. The analytical framework will integrate time-series methods, regularized regression (elastic net) for feature selection, and a composite deep learning model (temporal convolutional networks with attention) to capture non-linear interactions across modalities. Model performance will be evaluated with time-to-flare analyses, area under the receiver operating characteristic curve (AUC-PR), calibration plots, and decision-curve analysis to determine clinical utility. Explainability will be pursued through SHAP value analyses and a post-hoc interpretation of temporal feature contributions. The study expects to identify a parsimonious set of multi-modal predictors—comprising skin biomarker trajectories, environmental exposure indices, and wearable-derived physiological signals—that collectively yield superior flare prediction (target AUC ? 0.82, Brier score ? 0.12) compared with current clinical risk scores. It is anticipated that temporal patterns, such as rising TSLP and IL-4/IL-13 signaling preceding clinical exacerbations by 5–14 days, combined with increases in transepidermal water loss and disrupted sleep, will emerge as robust early indicators. The model is expected to demonstrate good calibration across subgroups defined by age, disease severity, and FLG genotype, and to maintain performance in external validation cohorts. Contribution to knowledge includes (i) the first integrated multi-modal predictive framework for AD flare forecasting that combines molecular, environmental, behavioral, and physiological data within a coherent temporal modeling approach; (ii) empirical elucidation of interaction effects among skin immunology, barrier integrity, and environmental triggers in driving flares; and (iii) practical guidelines for deploying a dermatology-focused decision-support system, including thresholds for alerting patients and clinicians and considerations for data privacy, workflow integration, and cost-effectiveness. The study concludes that multi-modal data fusion within a temporally aware predictive engine can meaningfully improve early flare detection and personalized management in atopic dermatitis, thereby enabling timely therapeutic adjustments and potentially reducing disease burden. Recommendations include developing user-friendly clinician dashboards, integrating model outputs with electronic health records for real-time decision support, exploring individualized intervention studies guided by predicted risk, and conducting external validations across diverse populations to ensure generalizability.