Insights from deep learning models on new-onset anxiety in patients following bariatric metabolic surgery.

Background: Due to its long-term effectiveness in weight control and cost-efficiency, bariatric metabolic surgery (BMS) has emerged as a promising treatment option for patients with severe obesity. However, its impact on certain mental health disorders remains unclear. Objective: This study aimed to utilize a deep learning (DL) model, DeepBiomarker2, which integrates social determinants of health (SDoH) and electronic health records (EHR), to identify clinical features associated with new-onset anxiety disorder following BMS. Methods: We conducted a case-control study using longitudinal EHR data from the University of Pittsburgh Medical Center (Jan 2004-Oct 2019) on patients who underwent bariatric surgery. DeepBiomarker2, a DL model integrating diagnoses, medications, lab tests, and neighborhood socioeconomic status, predicted new-onset anxiety. Perturbation-based contribution analysis identified key predictive features. Results: A total of 14,856 eligible patients who underwent BMS without a prior history of anxiety disorder were identified. DL models outperformed traditional logistic regression in predicting post-BMS anxiety, yielding area under the curve (AUC) values exceeding 0.89. Key features associated with post-BMS anxiety included abnormal urine and blood lab results, opioid and psychiatric medication use, frequent emergency department (ED) visits, and pre-existing mental health conditions. Potential protective indicators included omega-3 fatty acids, vitamin B12, calcium citrate, and pravastatin. Inclusion of nSES data led to marginal improvements in model performance. Conclusion: Our DL models successfully identified clinical features potentially associated with new-onset anxiety following BMS, offering valuable insights to support early intervention and personalized mental health strategies for postoperative care.