急性呼吸窘迫症候群是病人進入加護病房的常見原因之一,並且有很高的死亡率,雖然現今已經有很多研究運用臨床資料和機器學習方法探討及時診斷與提前預測的模型,但幾乎沒有研究同時考慮了數值資料及影像資料。本研究使用了公開資料集(MIMIC-IV 以及 MIMIC-CXR)以獲取病人的臨床資料及胸部X光片影像資料,應用機器學習方法建立決策樹(Decision Tree)、隨機森林(Random Forest)、極限梯度提升(XGBoost)、神經網路(Neural Network)等多種模型,並應用了多模態機器學習分析,比較單模態與多模態模型的表現。使用晚期融合的多模態模型在診斷及12小時、24小時及48小時前的預測,接受者操作特徵曲線下面積(AUROC)約為0.7951至0.8502,與單模態模型相比約可以提高6.0%至9.3%的模型表現,這個研究將可以協助急性呼吸窘迫症候群的診斷及早期預測。
Acute respiratory distress syndrome (ARDS) is one of the most common causes of admission to the intensive care unit and has a high mortality rate. Although there were several studies applying machine learning techniques to the issue of ARDS prediction, few studies combined numerical and image data. This study collected clinical data and chest radiograph images from publicly available databases (MIMIC-IV and MIMIC-CXR) and applied machine learning methods to establish models such as Decision Tree, Random Forest, XGBoost, and Neural Networks. Moreover, multimodal machine learning were applied and the performance of single- and multi-modality models were compared. The multi-modality models with late-level fusion demonstrated the AUROC of 0.7951~0.8502 for onset identification, 12-, 24-, and 48-hr prediction, which improved about 6.0%~9.3% compared with the single-modality models. This study can assist improved prediction and early recognition of ARDS.