使用機器學習分析醫療紀錄以預測心房震顫之發生

TMUIR > College of Medicine > Professional Master Program in Artificial Intelligence in Medicine > Dissertations/Theses > Item 987654321/61662

Please use this identifier to cite or link to this item: http://libir.tmu.edu.tw/handle/987654321/61662

Title:	使用機器學習分析醫療紀錄以預測心房震顫之發生 Assessment of Machine Learning Using Medical Records to Develop a Prediction Model for Atrial Fibrillation
Authors:	高永達 KAO, YUNG-TA
Contributors:	醫學院人工智慧醫療碩士在職專班張資昊
Keywords:	心房震顫;臨床資料庫;機器學習;預測工具 Atrial fibrillation;Electronic medical record;Machine learning;Predictive model
Date:	2021-06-30
Issue Date:	2022-04-28 23:10:02 (UTC+8)
Abstract:	一、前言心房震顫是腦中風的獨立危險因子，它增加了五倍的腦中風機會。早期偵測到心房震顫的發生，並且適當地給予口服抗凝血劑，可以改善臨床預後。本研究的目的是運用機器學習，根據多面向的醫療資訊（非心電圖訊號），研發出新生心房震顫的預測工具。二、方法研究分析自西元2008年1月1日開始，到2016年12月31日為止的臺北醫學大學臨床研究資料庫，約有兩百萬人的門急診與住院資料檔。依據連續三年的醫療資訊（含：診斷、用藥，與檢驗報告等），我們嘗試建立心房震顫的預測工具，以預測接續的一年是否會新發生心房震顫。我們計算此預測工具的特異度、靈敏度、跟接收者操作特徵曲線下的面積，以推估其表現。三、結果本研究總共收錄2138位心房震顫的病患（其中有1028位女性，占比48.1%）以及經過年齡性別配對的8552位無心房震顫診斷的對照組病患（其中有4112位女性，占比48.1%），結果發現當使用隨機森林綜合分析病患過往三年的診斷碼、藥物碼、以及檢驗報告等資訊來預測接下來的一年是否會新發生心房震顫時，能達到最佳的接收者操作特徵曲線下面積（0.74）以及最佳的精確率召回率曲線下面積（0.89）。大多數的機器學習模組都能達到正確預測率80%左右，唯有決策樹除外（約70%）。四、結論依此研究，我們得到一個能適度預測一年後有較高可能性產生心房震顫的高風險個案之預測工具。對於模組預測為陽性的個案，我們應及早安排適切的檢查工具以確診心房震顫（如：24小時心電圖檢查或心臟事件記錄器），希望藉由此預測工具能達到早期診斷（心房震顫）以及儘早施予治療的目的，改善病患預後。 Background: Atrial fibrillation (AF) itself is an independent risk factor for stroke. It increases the risk of stroke five-fold. Early detection of AF and judicious prescription of oral anticoagulants could improve clinical outcomes. Targeted screening should be performed on specific individuals. The purpose of our present study is developing an AF predictive model by machine learning (ML), based on non-electrocardiogram and three-year medical information in our database to identify AF risk. Methods: This study analyzed clinical information of 2 million patients including outpatient and in-hospital files from the Taipei Medical University Clinical Research Database (which comprises 3 hospitals in Taipei City and New Taipei City) from January 1, 2008 to December 31, 2016. We exposed all the electronic medical records in the database, including sex, age, diagnostic codes, medications, and laboratory data. The model was judged by sensitivity, specificity, F1 score, area under the receiver operating characteristic curve (AUROC), and area under the precision-recall curve (AUPRC). Results: A total of 2138 participants (1028 female [48.1%]; mean age 78.8 years) with AF and 8552 random controls (after matching process) without AF (4112 female [48.1%]; age 78.8 years) were studied in the model. The 1-year new-onset AF risk predictive model based on random forest algorithm using medication and diagnostic information along with specific laboratory data attained an AUROC of 0.74 and an AUPRC of 0.89. Most ML models could achieve an accuracy of 0.8, except the decision tree algorithm. Conclusion: The finding of the study suggests our AF predictive model could offer acceptable discrimination in differentiating risk of incident AF in the following year. Based on this study, healthcare professionals could focus on predict-positive patients and apply further feasible examinations such as Holter monitoring or event recorder for these specific individuals. A targeted screening approach for AF could result in a clinical choice with efficacy to identify patients who are at risk and improve outcome.
Description:	碩士指導教授：張資昊委員：吳育瑋委員：蘇家玉
Data Type:	thesis
Appears in Collections:	[Professional Master Program in Artificial Intelligence in Medicine] Dissertations/Theses

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	149	View/Open

View Licence

Loading...