利用巨量資料及深度學習預防用藥錯誤

he incidence rate of medication error is about 0.5% to 5%. It is estimated that there are 0.75 to 7.5 million inappropriate prescriptions occurred annually in Taiwan and 2.25 to 22.5 million ones in the US. Annual expense spent on medication errors is around 30 billion US dollars and causes 17500 to 35000 deaths. For the rule-based clinical decision support system we use right now, there is still room for improvement in terms of preventing medication errors. Like the high overridden rate for alerts, not cost effective, lack of clinical validation, and the need of self-learning ability.

Therefore, the main goal of this project is to connect the probabilistic and deep learning models, based on the clinical trial experience in Shuang Ho Hospital, with the rich resources in medical information system provided by Tatung Medical and Healthcare Technologies, the medication safety system developed by the Principal Investigator is actually introduced into the hospitals. The model also include other variables, such as physician department, patient gender, age, etc., to enhance the positive predictive rate of medication error detection. At the same time, the physician's response to the suggestion is continuously collected, and the data is used to feedbackthe model using Reinforcement Learning.

The system has been introduced in 4 medical institutions (Wan Fang Hospital, Shuang Ho Hospital, Taipei Medical University Hospital, and AHMC), and collected more than 100,000 prescriptions. About 170 doctors from 22 different departments participated in the trial. The alert rate is about 3%. Through Tatung Medical and Healthcare Technologies 's professional in the medical and healthcare market, the alert interface of the system was updated to version 2.0, and the system response time was reduced from the average of 4 seconds to about 0.05 seconds to ensure that the process of doctor’s visit was not affected. Twelve experts were invited to re-evaluate the prescriptions for different departments of the system. For 938 suggestions, the ratio of 2 out of 3 physicians agreed on the reminders was 70% or more. Therefore, after the implementation of this system, it can effectively detect inappropriate prescription and reduce medication errors, thereby improving patient safety and improving medical quality.