Assessment of the Risk Factors of MDD Recurrence Based on Deep Learning Approaches

Abstract

Objective: To explore the risk factors associated with of major depression disorder (MDD) recurrence using logistic regression, support vector machines (SVM) and a long-short term memory (LSTM) approaches based on two electronic health records (EHR) databases. Method: We extracted patients with MDD from two large, multi-center clinical datasets. The inpatient and outpatient datasets between Jan. 2010 and Dec. 2015 were collected. Eligible patients were 18-90 years-old and had a diagnosis of MDD. The major depression disorder (MDD) were identified based on the MDD-related International Classification of Diseases, 9th revision-Clinical Modification (ICD-9-CM) diagnosis codes (296.2x, 296.3x); and MDD-related ICD-10-CM diagnosis codes (F32.x, F33.x). The patients with less than three-time visits in a cohort during 5-year were followed up. Eventually, 140,497 patients were qualified for further analysis, including 69.2% female patients. Among of 140,497, 20, 078 patients (14.3%) had no complications. Logistic regression, SVM, and LSTM were employed to predict the key risk factors associated with MDD recurrence. Survival analysis was also conducted for analyzing the risk factors of MDD recurrence. Results: The MDD patients with married /life partners had a lower prevalence rate (9.2%) of MDD recurrence than the patients with single marital status (11.8%). The primary MDD patients had a higher MDD recurrence rate (11.7%) than secondary MDD patients (10.5%). Primary MDD was associated with MDD recurrence (OR 2.49, 95%CI 1.53-3.96) via logistic regression analysis. Insomnia, anxiety and single marital status were also top-ranked risk factors for the MDD recurrence. The prediction accuracy of logistic regression, SVM and LSTM were 0.736, 0.791 and 0.834, respectively. Conclusions: Building statistical models by mining existing EHR data can explore the risk factors associated with MDD recurrence. Our predictive results indicated that primary MDD, never married, anxiety symptoms, and insomnia were risk factors for MDD recurrence. The prediction accuracy of the LSTM model was higher than the other two approaches. Funding Statement: This work was support by UT SBMI. Declaration of Interests: The authors state that they have no potential conflict of interest. Ethics Approval Statement: This study was approved by the institutional review board (IRB).