Options
使用臨床電子病歷記錄來預測診斷疾病代碼
Other Title
Using Clinical Note to Predict Diagnosis Codes
Type
thesis
Date Issued
2023-07-19
Author(s)
JAKIR HOSSAIN BHUIYAN MASUD
Advisor
林明錦
Subjects
系所名稱：醫學資訊研究所博士班
Description
學位別：博士
語文別：英文
口試委員：林嶔 LIN, CHIN;蘇家玉 SU, EMILY CHIA-YU ;邱泓文 CHIU, HUNG-WEN ;徐建業 HSU, CHIEN-YEH ;林明錦 LIN,MING-CHIN
授權範圍：網際網路,開放日期為2023-07-26
語文別：英文
口試委員：林嶔 LIN, CHIN;蘇家玉 SU, EMILY CHIA-YU ;邱泓文 CHIU, HUNG-WEN ;徐建業 HSU, CHIEN-YEH ;林明錦 LIN,MING-CHIN
授權範圍：網際網路,開放日期為2023-07-26
Abstract
Background
Automatic coding and classification systems play a significant part in healthcare. Deep learning can make efficient automated diagnosis coding for quality of care. Our two objectives are: for Paper 1, i) to create and confirm a deep-learning-based model to support physicians in choosing the proper ICD codes when ordering drugs, for Paper 2, ii) to generate a deep-learning-based model to support physicians in choosing the suitable ICD codes.
Methods
This dataset was used from a Hospital, Taipei, Taiwan. The dataset, which we compiled from clinical notes (drug lists were used to predict diagnosis codes for Paper 1 and SOAP notes with drug lists were used to predict diagnosis codes for Paper 2) from several departments, covered the months of January through December of 2016. For training and testing, we divided the dataset 90/10. After the data preprocessing, this study developed a CNN model to investigate the data.
Results
For paper 1, the best model attained the performance for cardiology with precision (0.69), recall (0.89), and an F-score (0.78).
For paper 2, the best model attained the performance for cardiology with precision (0.96), recall (0.99), and an F-score (0.98).
Conclusion
Our results achieved higher performance compared to the previous studies using CNN for automatic ICD-10 coding.
Automatic coding and classification systems play a significant part in healthcare. Deep learning can make efficient automated diagnosis coding for quality of care. Our two objectives are: for Paper 1, i) to create and confirm a deep-learning-based model to support physicians in choosing the proper ICD codes when ordering drugs, for Paper 2, ii) to generate a deep-learning-based model to support physicians in choosing the suitable ICD codes.
Methods
This dataset was used from a Hospital, Taipei, Taiwan. The dataset, which we compiled from clinical notes (drug lists were used to predict diagnosis codes for Paper 1 and SOAP notes with drug lists were used to predict diagnosis codes for Paper 2) from several departments, covered the months of January through December of 2016. For training and testing, we divided the dataset 90/10. After the data preprocessing, this study developed a CNN model to investigate the data.
Results
For paper 1, the best model attained the performance for cardiology with precision (0.69), recall (0.89), and an F-score (0.78).
For paper 2, the best model attained the performance for cardiology with precision (0.96), recall (0.99), and an F-score (0.98).
Conclusion
Our results achieved higher performance compared to the previous studies using CNN for automatic ICD-10 coding.