到院前分類重大創傷之人工智慧模型

Abstract

在交通系統中，意外事故帶來的傷亡是備受關注且亟需解決的議題。在這些需緊急送醫的傷患中，難以判斷其創傷輕重程度，將這些患者送至較遠的創傷中心或較近的地區醫院，為急難救助的一大難題。美國外科醫學會對現場檢傷提出期望，希望使檢傷不足的比例小於百分之五，過度檢傷的比例小於百分之三十五，提供傷患合適的救治，降低死傷比例。在過去的研究中，許多檢傷工具是以死亡率進行傷患的分診。現有的檢傷工具，準確度無法滿足美國外科醫學會的期望。事實上創傷的判定繁複且包含人體許多部位的屬性資料。 本研究希望運用能處理屬性資料的機器學習方式，以 ISS 分數為判斷重大創傷的依據，並以此為預測目標，建立到院前判斷是否為重大創傷的檢傷模型，且希望使模型達成美國外科醫學會對現場檢傷的期望。 此研究以機器學習模型 CatBoost，並以2016年至2020年的Pan-Asian Trauma Outcomes Study (PATOS)資料集為研究範本，進行模型的訓練。初始結果顯示在韓國與台灣地區都能有十分亮眼的表現，經由參數的調整更可達到美國外科醫學會檢傷不足比例小於等於5%，過度檢傷比例小於等於35%的期望。依欄位對模型預測的重要性進行篩選後，在大幅減少所需欄位的情況下，也可符合美國外科醫學會的期望。希望未來能將此系統運用於到院前的現場分診，幫助急難救助人員選擇合宜的醫療院所，使傷患獲得最適當治療、降低其死傷比例。

Accidents and trauma attributed to the transportation system operations are a crucial issue worldwide. There are global efforts to reduce mortality and morbidity from major trauma, and the prehospital triage is one of the core initiatives. Over and under triage generate various problems and both should be avoided. Thus, the American College of Surgeons set up goals of less than 5% of under triage and less than 35% of over triage. In previous studies, many triage tools were analyzed by mortality. In addition, several prevailing triage techniques that are based on vital signs cannot meet ACS expectations. The triage procedures are very complicated and involve a lot of categorical data. We want to use ISS scores as judgment criteria, and use machine learning techniques that can process categorical data to build a prehospital triage model that meets ACS goals. This research uses CatBoost, a machine learning model to train and analyze the PATOS data. The preliminary results show promising outcomes in Korea and Taiwan. Both less than 5% under triage and less than 35% over triage can be achieved via parameter adjustment. We further refined the model with minimal number of parameters, while maintaining triage sensitivity and specificity. We expect to apply our research to prehospital triage practice, facilitating better hospital selection that leads to optimal treatment and outcome.