健康風險預測

Abstract

現代人對健康的重視逐漸提升，未來健康狀況的預測是現今相當重要的議題，其中，住院為重大的醫療事件，因照護成本高，不論對個人、醫院、全民健康保險都是一項較沉重的負擔，因此，未來是否會住院成為許多人所關心的議題，本研究將聚焦於次年與兩年後的住院率及長期住院率的預測。 過去的研究中，預測未來健康事件時，經常使用合併症指標(Comorbidity index)將病情嚴重性量化，現在主要有以診斷基礎合併症指標和以藥物處方為基礎合併症指標。由於合併症指標能用來表現病情的嚴重程度，因此常被許多研究者作為預測未來健康狀況的特徵值之一。因此，本研究將合併症指標做為預測特徵值，並同時使用兩種合併症指標作為預測特徵值，提升預測表現。除了使用合併症指標當作預測特徵值，也使用機器學習(Machine learning)的方式，例如：特徵值選擇(feature selection)和主題模型(Topic model)，尋找能預測患者未來健康狀況的特徵值。本研究希望透過不同的特徵值及模型，改善未來健康風險的預測。 從實驗結果來看，本研究所使用的特徵值皆具有預測未來住院的能力，而結合不同的合併症指標，比單獨使用其中一種的表現來得佳。另外，透過機器學習方法找出的特徵值，在預測未來住院的表現上，比合併症指標表現來得好，未來除了使用合併症指標預測未來健康風險，也能參考機器學習方法找出的特徵值，改善預測結果。

With increasing attention on individual health, the prediction of individual health risk is an important issue today. One of the important health risks is the hospitalization. The hospitalization is costly for individual, society and health insurance, and many people concern this issue. Therefore, this study focuses on predicting subsequent-year and subsequent-2year hospitalizations. In the past, many research use comorbidity index to quantify the diseases and present the risk of individual health. They predict the risk of individual health by evaluating one’s comorbidity score. For this reason, we use comorbidity as our feature to predict future hospitalization. We also combine different comorbidity index to improve performance. In addition, we use machine learning method, such as feature selection and topic model to find the factors which affect individual’s future health status. Our research expects to improve predicting performance of the risk of individual health by using different feature combination and model. Our results show that both comorbidity index and the feature found by machine learning methods have good performance. In addition, the performance predicted by the feature found by machine learning methods is better than comorbidity index.