自動化醫學人工智慧用藥安全系統之跨國確效研究

Taipei Medical University Institutional Repository > 醫學科技學院 > 醫學資訊研究所 > 博碩士論文 > Item 987654321/63367

請使用永久網址來引用或連結此文件: http://libir.tmu.edu.tw/handle/987654321/63367

題名:	自動化醫學人工智慧用藥安全系統之跨國確效研究 Assessing the Cross-national Transferability of an Automatic Medical Artificial Intelligence-based Prescription Safety System
作者:	靳嚴博 CHIN, YEN-PO
貢獻者:	醫學資訊研究所博士班李友專
關鍵詞:	機器學習;病人安全;警示疲勞;臨床決策支持系統;用藥安全系統;醫學人工智慧;聯盟式學習 Machine learning;Patient safety;Alert fatigue;Clinical decision support system;Prescription safety system;Medical artificial intelligence;Federated learning
日期:	2022-12-26
上傳時間:	2023-12-15 16:09:09 (UTC+8)
摘要:	研究背景: 當今使用的大多數用藥安全警示系統都是基於規則式的邏輯架構建立;然而，這些警示系統的準確性往往不盡理想，進而導致警示疲勞。為了解決這個問題，先前研究人員構建了基於台灣健保資料庫的人工智慧機器學習用藥錯誤警示系統。不過，這種基於機器學習的用藥錯誤警示系統是否具備跨國可應用性，目前尚未有定論。研究目的: 本研究檢驗了人工智慧機器學習用藥錯誤警示系統的跨國可應用性，以及聯盟式學習方法是否可以進一步提高該機器模型的準確性。研究材料與方法: 研究材料包括來自美國兩個大型醫療中心的 667,572 份的門診處方籤。我們的機器學習模型用於構建初始模型、當地模型和混合模型。初始模型是基於台灣健保資料庫的 13 億筆處方籤開發完成。我們另外使用聯盟式學習策略訓練了混合模型。兩名的醫師評估者將測試處方集分類為可解釋或未能解釋，進而用於評估模型性能。結果: 在將處方分類為可解釋或未能解釋方面，兩位醫師評估者之間的一致性達到顯著水準。閾值範圍為 0.5 到 1.5，初始模型的警示準確度範圍為 75% 到 78%，當地模型為 76% 到 78%，混合模型為 79% 到 86%。結論: 該人工智慧機器學習模型用藥錯誤警示系統在美國數據集中站展現了良好的跨國可應用性。透過聯盟式學習結合當地醫院的數據集可以進一步提高該模型的準確性。 Research Background: The majority of the medication safety systems in use today are rule-based; however, these systems often have poor accuracy, which can lead to alert fatigue. To address this problem, researchers have constructed a model through machine learning (ML) using the Taiwan national health insurance database (TNHID) in the past. However, it is not clear whether or not this model can be transferred to healthcare systems in other countries. Research purposes: The current research assesses the cross-national transferability of a ML-based prescription error detection system. We also evaluated whether federated learning could improve the system's transferability. Methods: The study material includes 667,572 prescriptions from two hospitals in the United States. We constructed the Taiwan (TW) model, the United States (US) model, and the Hybrid (HB) model following the appropriateness of the prescription model method. The TW model was developed using 1.3 billion prescriptions from the TNHID. A validation set consisting of 9% of the prescriptions was randomly sampled, and the remaining 91% served as the local training set for the US model. Using a federated learning method, the HB model integrated the association values with the highest co-occurrence frequency from both the TW and US models. Two independent physician reviewers classified the test set as substantiated or unsubstantiated, which was then used to evaluate model performance. Result: As for classifying whether the prescription is substantiated or unsubstantiated, the inter-rater agreement was significant. With varying thresholds, the alert accuracy for the original model ranged from 75% to 78%, for the local model from 76% to 78%, and for the hybrid model from 79% to 86%. Conclusion: Our prescription error alert system based on TW clinical data presents a reasonable cross-national transferability among US clinical data. Combining local US data through federated learning could further improve the system’s performance.
描述:	博士指導教授:李友專委員:吳俊穎委員:郭博昭委員:雪必兒委員:張資? 委員:李友專
資料類型:	thesis
顯示於類別:	[醫學資訊研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	90	檢視/開啟

在TMUIR中所有的資料項目都受到原著作權保護.

TAIR相關文章

資料載入中.....