acadtags / awesome-medical-coding-nlp Goto Github PK

Our paper has been accepted to SIGIR'23. We present the first results on MIMIC-IV and demonstrate several weaknesses of previous work. I hope you find it relevant to your list!

We have a new paper on Arxiv: An Unsupervised Approach to Achieve Supervised-Level Explainability in Healthcare Records

In this paper, we propose a new explanation method, AttInGrad, that outperforms previous methods by a large margin. Furthermore, we show that training our models to be adversarial robust also improves their explanations. By combining both contributions, we show that we can perform similarly and sometimes better than a supervised approach that is trained on evidence span annotations.

In addition to the paper, we also released a new GitHub repository. Our previous repository medical-coding-reproducibility has become quite popular. This repository introduces several improvements and new features:

• Explainability: We implemented multiple feature attribution methods and metrics.
• Implementation of a modified PLM-ICD: In our previous paper, we had issues with PLM-ICD that occasionally collapsed during training. We have fixed this problem.
• Huggingface Datasets: we implemented MIMIC-III, IV, and MDACE as HuggingFace datasets.
• More efficient pre-processing: Our data pre-processing is faster and uses less memory thanks to Polars.
• Inference code: you can now easily use the models for inference without access to the original dataset. This was a desired feature in our previous repository.