Predicting depression and anxiety in diabetes with limited primary care data
We pre-trained a model on good-quality hospital records and fine-tuned it on primary care data, which predicted post-discharge depression and anxiety much better than baselines.
The question
Primary care records in China are varied and limited, which makes it hard to assess the psychological risk of people with type 2 diabetes there. Hospitals hold better-quality data, but it is not the same setting.
What we did
We built a deep learning framework called DAP. First we pre-trained a foundation model, using contrastive learning, a way for a model to learn patterns from unlabelled records, on good-quality records from 85,085 type 2 diabetes in-patients at a large university hospital. Then we fine-tuned it to predict depression using records from 149,596 patients on the Nanjing Health Information Platform.
What we found
The approach was useful for predicting depression and anxiety after discharge among type 2 diabetes patients in primary care, with much higher performance than baseline models.
Why it matters
Knowledge learnt in data-rich hospitals can be transferred to data-poor primary care, where most patients are seen.