DEformer: A Doctor’s Diagnosis Experience Enhanced Transformer Model for Automatic Diagnosis

The DEformer: A Doctor’s Diagnosis Experience Enhanced Transformer Model for Automatic Diagnosis is built based on DxFormer's decoder-encoder framework. The repo can be used to reproduce the results in the paper:

Overview

In this paper, A Doctor’s Diagnosis Experience Enhanced Transformer Model for Automatic Diagnosis model is proposed to learn more implicit experience of doctors. On the Dxy、MZ-4 and MZ-10 dataset, our model outperforms in the core metrics diagnosis accuracy in lower inquiry rounds from 0.7% to 2.0% compared to the most advanced models. In addition, on the MZ-10 dataset our model's symptom recall rate metric improve 9.4% compared to the previous state-of-the-art model.

Setup

The repo mainly requires the following packages.

• nltk 3.3
• python 3.8
• torch 1.7.0+cu110
• torchvision 0.8.1
• scikit-learn 0.20.0

Full packages are listed in requirements.txt.

1. Download data

The dataset can be downloaded as following links:

• Dxy dataset
• MZ-4 dataset
• MZ-10 dataset

2. Preprocess

python preprocess.py

3. Pre-training

python pretrain.py

4. Training

python train.py

5. Inference

python early_stop.py

Acknowledgement

Many thanks to the open source repositories and libraries to speed up our coding progress.

• DxFormer https://github.com/lemuria-wchen/DxFormer