基於機器學習的中風患者離院時之修改過的雷氏量表及惡化預測

Abstract

腦血管疾病在台灣是主要的死亡原因之一，及時準確的結果預測在治療決策重要作用，在本篇論文中，我們建立機器學習模型並進行驗證和分析，以預測出院時的雷氏量表分數及惡化。以ROC曲線下面積（AUC）評估，隨機森林在兩個目標中均表現最佳，在對於病房中惡化（目標不平衡）的預測模型訓練中，我們進行重新採樣的實驗。整體而言，我們觀察到，過去中風的量表，包含：雷氏量表、NIHSS、巴氏量表，是預測的關鍵，同時也指出添加更多預測因子，可以略微增加模型的AUC，最後，本篇論文也示範了以SHAP及LIME解釋模型重要因子，並確認模型的可靠性。

Cerebrovascular disease is a leading cause of death in Taiwan. Timely and accurate outcome prediction plays an important role in guiding treatment decision. In this work we focus on the ML development, validation and model analysis for predicting mRS at discharge and deterioration. Random forest performs the best in both target evaluated with Area Under the ROC Curve(AUC). For deterioration during ward, which target is imbalanced, experiment with re-sampling is also included. We observe that by features obtained by assessment like mRS, NIHSS, BI are key for predicting. We conclude that not only the random forest could be the best model to use for prediction, but also point out adding more features, like blood test result, can slightly increase AUC of models. Interpretation for prediction is also described in this work.