Evaluation of deep learning-based depression detection using medical claims data.

Area of Science:

• Artificial Intelligence in Medicine
• Computational Psychiatry
• Health Informatics

Background:

• Human diagnostic accuracy for psychiatric disorders remains suboptimal.
• Despite increased healthcare digitization, AI-based digital decision support (DDSS) adoption is limited due to insufficient real-world data evaluation.
• This research addresses the need for robust AI model validation using extensive, real-world medical claims data.

Purpose of the Study:

• To evaluate the efficacy of deep learning models in predicting depression using longitudinal medical claims data.
• To investigate the impact of temporal data properties on depression detection accuracy.
• To develop and validate a highly accurate and explainable AI model for depression prediction.

Main Methods:

• Utilized a large dataset of 812,853 individuals' medical claims from 2018-2022, encompassing 26,973,943 ICD-10 coded diseases.
• Compared non-sequential (LR, FNN), sequential (LSTM, CNN-LSTM), and decay-factor sequential models (GRU-Δt, GRU-decay).
• Developed and evaluated a novel explainable model, Att-GRU-decay, combining self-attention with GRU decay.

Main Results:

• The Att-GRU-decay model achieved superior performance with an AUC of 0.990, AUPRC of 0.974, specificity of 0.999, and sensitivity of 0.944.
• Outperformed existing state-of-the-art methods in depression prediction.
• Demonstrated the critical importance of temporal data features for accurate depression detection.

Conclusions:

• The novel Att-GRU-decay model shows significant potential for accurate depression prediction using medical claims data.
• Validated deep learning's utility for developing effective AI-based digital decision support systems (DDSS).
• Proposed an application scenario for GP-based depression screening to improve care quality and reduce costs.