Multi-task deep learning and interpretable non-linear neural interaction modeling for personalized skin concern prediction

We develop a multi-modal predictive framework that integrates genetic and lifestyle factors to model six dermatological phenotypes—Pigmentation, Dryness, Sensitivity, Scarring, Acne, and Redness—using a cohort of 5254 adult women (aged 30–70, mostly from Belgium) from the Nomige platform. We implement a multi-task deep neural network (MT-DNN) with shared layers and dual outputs for continuous severity prediction (evaluated by mean absolute error, MAE) and ordinal classification (evaluated by quadratic weighted kappa, QWK). To model potential non-linear gene–lifestyle (G × L) interactions, we incorporate a Neural Interaction Network that explicitly represents pairwise interactions between six skin-related genes (MMP1, MMP3, SOD2, GPX1, AQP3, FLG) and 22 lifestyle factors. The MT-DNN achieves robust internal performance with strong ordinal agreement across phenotypes. PFI, saliency, and SHAP identified model-attributed predictors, with lifestyle variables such as stress, sleep, and hydration-related behaviors carrying substantial predictive information in this cohort, while genetic profiles such as GPX1, SOD2, and AQP3 contributed in phenotype-specific ways. SHAP interaction analysis highlighted candidate gene–lifestyle patterns, including AQP3 × stress for Redness and GPX1 × outdoor exposure for Acne. These patterns were further examined using a post hoc ordinal-logistic sensitivity analysis with FDR correction. Collectively, the results support multi-task interaction modeling as an internally validated, hypothesis-generating approach for studying cohort-specific gene–lifestyle associations in personalized skin-health prediction.