Phenotyping antidepressant treatment response with deep learning in electronic health records

Efficient, accurate phenotyping for antidepressant treatment response in electronic health records (EHRs) could facilitate precision psychiatry applications but remains a challenge. Increasingly, artificial intelligence methods using deep learning applied to clinical data have shown promise in complex classification problems. Here, we systematically evaluate the performance of eight deep learning-based language models in classifying response to antidepressants in a large real-world healthcare setting. We obtained data spanning 1990-2018 for adults with depression and a co-occurring antidepressant prescription from the EHR data warehouse of the Mass General Brigham healthcare system (n = 111,563). Clinical note sets (n = 3996) were collected for 4-12 weeks after antidepressant initiation, and were manually reviewed to classify response status as "improved" or "no evidence of improvement" in depression symptoms for model development and evaluation. The phenotyping models performed well, with the majority having areas under the receiver operator curve (AUROC) exceeding 0.80. For example, BERT-large with sliding window had an AUROC = 0.84 and PPV = 0.80 at a specificity of 0.87. Our results indicate that deep learning language models applied to EHR data can accurately classify antidepressant response in a real-world healthcare setting.