Deep Learning-based Time-to-event Analysis of Depression and Asthma using the All of Us Research Program

Area of Science:

• Medical Informatics
• Public Health
• Psychiatry

Background:

• The association between depression and asthma is increasingly recognized.
• Large-scale retrospective studies using advanced analytical models are limited.

Purpose of the Study:

• To analyze the association between depression and asthma using deep learning (DL) and traditional statistical models.
• To compare the performance of DL models (DeepSurv, DeepHit) against logistic regression and Cox Proportional Hazards (CoxPH) models.
• To identify key variables influencing depression in asthma patients using SHAP values.

Main Methods:

• Retrospective cohort study of 239,161 participants from the All of Us Research Program.
• Analysis employed DL-based models, logistic regression, and CoxPH models.
• SHAP values were used for DL model interpretability; c-index evaluated model performance.

Main Results:

• A significant odds ratio for depression in asthma patients was observed.
• CoxPH model achieved a c-index of 0.619, DeepSurv 0.625, and DeepHit 0.596.
• SHAP analysis revealed different important variables compared to the CoxPH model, highlighting sex at birth and income.

Conclusions:

• Strong evidence supports a positive relationship between depression and asthma.
• DL-based models did not outperform the CoxPH model in terms of predictive accuracy (c-index).
• Sex at birth and income are potentially significant factors in the occurrence of depression among asthma patients.