Researchers systematically compared eight deep learning architectures for phenotyping antidepressant treatment response from electronic health records, reporting that transformer-based models significantly outperformed traditional machine learning approaches in a study published in Nature Translational Psychiatry.
Accurately determining whether a patient is responding to antidepressant therapy from EHR data is notoriously difficult. Clinical notes contain rich but unstructured information that standard billing codes miss factors like side effect tolerability, partial response, and functional improvement that determine real-world treatment success.
The study evaluated models across multiple academic medical center EHR systems, comparing performance on tasks ranging from binary response classification to granular symptom trajectory prediction. Transformer models incorporating both structured data and unstructured clinical text achieved the highest accuracy, particularly when pretrained on large medical corpora before fine-tuning on the antidepressant response task.
The findings have immediate implications for precision psychiatry. Reliable automated phenotyping could enable large-scale observational studies of treatment effectiveness, power clinical trial recruitment by identifying appropriate patient populations, and eventually support clinical decision support systems that help match patients to the right antidepressant earlier in their treatment journey.
The authors emphasize that model performance varied significantly across patient demographics and clinical settings, underscoring the need for diverse training data and rigorous validation before deployment in real-world mental health care.
