應用多管道潛藏狄利克雷分配偵測藥物清單遺失項目

Abstract

Medication Error (用藥疏失) 是指因為不當使用藥物而對病患安全產生危害的事件，醫師開立之處方箋與病患實際所服用藥物有差異是造成其發生的主要原因之一。在眾多藥物資訊差異型態中，尤以藥物資訊遺失為大宗。若能修正藥物清單的錯誤，還原遺失的藥物資訊，便能使病患受到更完善的醫療照顧，故本研究希望提出一個有效的方法偵測藥物清單上的遺失項目。 本研究以類比主題建模 (topic modeling) 的方式，找出診斷資訊與藥物資訊共同的潛藏因子，建立兩者之間的關聯，透過可觀察的資訊預測藥物清單上可能遺失的項目。本研究以Latent Dirichlet Allocation (潛藏狄利克雷分配，簡稱LDA) 主題模型為基礎，針對研究問題與資料將其延伸發展為Multiple Channel Latent Dirichlet Allocation (多管道潛藏狄利克雷分配，簡稱MCLDA) 模型，期望能有效地進行預測。 本研究採用全民健康保險研究資料庫的健保資料進行實驗，實驗結果證明MCLDA模型表現優於先前研究所提出的協同過濾模型，能大幅地提升預測的準確程度。MCLDA模型建立了診斷資訊與藥物資訊之間的關聯性，故亦可將偵測藥物清單遺失項目的概念延伸應用於偵測藥物清單異常項目；相同地，MCLDA模型亦可透過藥物清單的幫助進而偵測診斷清單中的異常項目。因此，本研究建構的全民健保診斷與藥物之聯合機率模型可作為往後全民健保方面相關研究的基礎。

A medication error is any preventable event that leads to patient harm because of inappropriate drug use. One major reason is because discrepancies may exist between drugs prescribed and drugs taken by patients. Drug omission is a major cause of discrepancies. The goal of this study is to develop a drug omission detection method to help resolve these discrepancies and improve health care quality. This study developed a method for automatic detecting omissions in medication lists. Our key insight is that the diagnosis items and drug items from one medical visit have the same latent health condition factors. The problem is, to some degree, analogous to the topic modeling framework which is increasingly used to inference latent topics in documents. We modified the Latent Dirichlet Allocation model according to the characteristics of the health insurance claim data and developed the Multiple Channel Latent Dirichlet Allocation (MCLDA). MCLDA discovers the relationship between diagnoses and drugs and compute the conditional probability that a given drug is omitted given the observed drugs and diagnosis. We evaluate the effectiveness of MCLDA model using the medication data from National Health Insurance (NHI) Research Database. Compared with previous Collaborative Filtering methods, the results show that MCLDA has improved accuracy for drug omission predicting.