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

Human accuracy in diagnosing psychiatric disorders is still low. Even though digitizing health care leads to more and more data, the successful adoption of AI-based digital decision support (DDSS) is rare. One reason is that AI algorithms are often not evaluated based on large, real-world data. This research shows the potential of using deep learning on the medical claims data of 812,853 people between 2018 and 2022, with 26, 973, 943 ICD-10-coded diseases, to predict depression (F32 and F33 ICD-10 codes). The dataset used represents almost the entire adult population of Estonia. Based on these data, to show the critical importance of the underlying temporal properties of the data for the detection of depression, we evaluate the performance of non-sequential models (LR, FNN), sequential models (LSTM, CNN-LSTM) and the sequential model with a decay factor (GRU-At, GRU-decay). Furthermore, since explainability is necessary for the medical domain, we combine a self-attention model with the GRU decay and evaluate its performance. We named this combination Att-GRU-decay. After extensive empirical experimentation, our model (Att-GRU-decay), with an AUC score of 0.990, an AUPRC score of 0.974, a specificity of 0.999 and a sensitivity of 0.944, proved to be the most accurate. The results of our novel Att-GRU-decay model outperform the current state of the art, demonstrating the potential usefulness of deep learning algorithms for DDSS development. We further expand this by describing a possible application scenario of the proposed algorithm for depression screening in a general practitioner (GP) setting-not only to decrease healthcare costs, but also to improve the quality of care and ultimately decrease people's suffering.