|
Marcel Haas@harcel.bsky.social |
This can be very useful, as the structured data in Electronic Health Records are often very incomplete in these features, so this is a promising way to enhance the completeness and quality of such EHRs. This was achieved by training and fine-tuning BERT models to identify the best strategy. (3/N)
1 replies 0 reposts 0 likes
|
|
Marcel Haas
@harcel.bsky.social
[ View ] |
A model pre-trained on the corpus itself strongly outperforms string matching (the conventional method). A great strategy: translate to English and use an English clinical language model, fine-tuned for the prediction task. Promising results, also for other (perhaps less popular) languages! (N/N)
0 replies 0 reposts 0 likes