[Methods 2024]Automatic ICD-10-CM coding via Lambda-Scaled attention based deep learning model

论文网址：Automatic ICD-10-CM coding via Lambda-Scaled attention based deep learning model - ScienceDirect

英文是纯手打的！论文原文的summarizing and paraphrasing。可能会出现难以避免的拼写错误和语法错误，若有发现欢迎评论指正！文章偏向于笔记，谨慎食用

目录

1. 心得

2. 论文逐段精读

2.1. Abstract

2.2. Introduction

2.3. Methodology

2.3.1. Overview of workflow

2.3.2. Creating clinical Pool of liver transplant (CPLT) database

2.3.3. Model architecture

2.3.4. Web application deployment

2.4. Experiments

2.4.1. Datasets

2.4.2. Evaluation metrics

2.4.3. Parameter setting

2.5. Experimental results

2.5.1. Baseline models

2.5.2. Results

2.6. Discussion

2.7. Conclusion

1. 心得

（1）设计的比较简单

2. 论文逐段精读

2.1. Abstract

        ①Task: automatic International Classification of Diseases (ICD) (ICD-10-CM) coding

2.2. Introduction

        ①Version of ICD: ICD-9, ICD-10, ICD-11 etc.

        ②Challenge: lack of kownledge system will caused suboptimal diagnosis result

        ③Introduced relevant works

2.3. Methodology

2.3.1. Overview of workflow

        ①Pipeline:

2.3.2. Creating clinical Pool of liver transplant (CPLT) database

        ①The authors annotated MIMLT with 1380 ICD-9-CM samples then named it as “Clinical Pool of Liver Transplant” (CPLT)

        ②They transfer ICD-9-CM to ICD-10-CM by https://www.aapc.com/icd-10/codes/（啊，这可靠吗）

        ③They directly accepted one-to-one transfer but found experts to precisely classify the cases of one-to-many

2.3.3. Model architecture

（1）Embedding layer

        ①Employ Word2Vec to mapping original code/words  in one clinical text  to  vectors  with dimension of  

（2）Deep bi-directional LSTM layer

       ①The processes of bi-LSTM:

where （是根据时间步自适应的？）

（3）Multi-scale CNN layer

        ①This MS-CNN is constructed by a max pooling layer and mean pooling layer after CNN 

        ②Concatenate all output from MS-CNN they get :

（4）Lambda-Scaled attention layer

        ①They further scale features:

（5）Classification

        ①Classification by fully connected layers with Sigmoid 

        ②Optimizer: Adam

        ③Binary cross entropy loss:

2.3.4. Web application deployment

        ①They build a web to predict ICD code

2.4. Experiments

2.4.1. Datasets

        ①Pretraining on MIMIC III and test on MIMIC III-Top 50 and CPLT

        ②Statistics:

（1）MIMIC III

        ①Sample: 53423

        ②Pre-training: stopword removal, tokenization, lowercase conversion, and removal of numbers, punctuation, and symbols by the Natural Language Toolkit (NLTK) library

        ③Limited record length: 2500

（2）CPLT

        ①Samples: 1380

        ②Preview: 

        ③Data split: 1104 for training, 138 for testing and 138 for val

        ④Max record length: 150

2.4.2. Evaluation metrics

        ①Micro F1 and Macro F1 for imbalanced data

2.4.3. Parameter setting

        ①Hidden dim of bi-LSTM: 64 for CPLT and 128 for MIMC III-Top 50

        ②Batch size: 8

        ③Epoch: 50

        ④Dropout rate: 0.5

2.5. Experimental results

2.5.1. Baseline models

        ~

        

2.5.2. Results

（1）Complete distribution of ICD-10-CM codes in the CPLT database

        ①34 ICD code in CPLT dataset:

（2）Comparison of DRCNN-ATT model with baselines on CPLT database

        ①Performance comparison table on CPLT:

（3）Comparison of DRCNN-ATT model with baselines on MIMIC III-Top 50 database

        ①Performance on MIMIC III-Top 50:

（4）Medical code Predictor web application

        ①Application preview:

（5）Ablation study

        ①Attention module ablation:

2.6. Discussion

        ~

2.7. Conclusion

        ~