利用機器學習預測住院病人出院之目的地

Discharge Destination Prediction for Hospitalized Patients Using Machine Learning

葉沛陽(Pei-Yang Ye)

指導教授 ： 賴飛羆

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2018年)

https://doi.org/10.6342/NTU201902183

摘要

出院準備服務是醫院端提供安排後續照護的橋樑，如果銜接的夠早夠好，可以減少因安排後續歸宿而拖延住院日數，也因安排得當可減少出院後再入院的比率，同時增加病人與家屬對醫療團隊的滿意度，也可以增進病人出院後返家/入住機構的照顧品質。在決定出院目的地時，常常因需要醫護人員、病人與家屬三方的多次溝通而拖延到出院時間。
在這項研究中，我們希望能建立一個機器學習的分類模型，能夠藉由輸入病人的資料以快速決定病人合適的出院目的地。我們採取的病人資料包括病人的基本資料、入院護理評估、出院照護摘要、入出院時部門、跌倒危險性分數、住院期間各項治療次數(物理、職能、心理、語言)以及診斷碼。在預測出院目的地中，我們定義了三個類別(轉院、轉機構、回家)。由於我們面臨到很嚴重的資料不平衡問題，我們採取了集成方法(Ensemble method)來解決它。而考慮到現實情況，有些病人即使需要轉院或轉機構卻執意想要回家。因此，我們利用分群方法(K-means)來移除回家類別中的異常值(Outlier)。最後，我們利用梯度提升決策樹(Gradient Boosted Decision Tree)並使用混淆矩陣以及自定義一個weighted macro f5 score來評估模型的預測效果。
在預測病人出院目的地中，轉院類別的準確率達到65%，轉機構類別的準確率達到89%，回家類別的準確率達到83%。整體上來看，我們自訂的weighted macro f5 score達到0.823。還找出了一些重要的特徵，例如巴氏量表分數、住院期間、年齡、出院後主要照顧者。

關鍵字

出院目的地 ； 機器學習 ； 多分類問題 ； 資料不平衡 ； 異常值 ； 重要特徵

並列摘要

Discharge preparation service is a bridge from the hospital to provide follow-up care. If the service is executed early enough, it can reduce the extra length of hospital stay caused by arranging follow-up, then reduce the re-hospitalization rate, increase the patient satisfaction in the hospital and medical staff, and even improve the quality of follow-up care after discharge.
In this study, we hope to establish a classification model for machine learning that can quickly determine the appropriate discharge destination for patients. The patient information we took included the patient's basic information, admission note, discharge note, department, falls-risk score, the number of therapies during the hospital stay (physical, occupational, psychological, speech) and diagnosis code. In predicting discharge destinations, we define three classes (“Hospital,” “Facility,” and “Home”). Since we faced a very serious data imbalance problem, we had adopted the Ensemble method to solve it. Considering the reality, some patients insist on going home even if they need to go to hospitals or facilities. Therefore, we use K-means to remove outliers from the class “Home.” Finally, we adopted the Gradient Boosted Decision Tree algorithm. And, we used the confusion matrix and a weighted macro f5 score (〖F(5)-Score〗_am) to evaluate the model's predictions.
In the result, the accuracy rate of the class “Hospital” reached 65%, the accuracy rate of the class “Facility” reached 89%, and the accuracy rate of the class “Home” reached 83%. Overall, 〖F(5)-Score〗_am reached 0.823.

並列關鍵字

discharge destination ； machine learning ； multi-class classification ； data imbalance problem ； outlier removal ； feature importance

參考文獻

[1] P. W. New, N. Andrianopoulos, P. Cameron, J. Olver, and J. U. Stoelwinder, "Reducing the length of stay for acute hospital patients needing admission into inpatient rehabilitation: a multicentre study of process barriers," Intern. Med. J., vol. 43, no. 9, pp. 1005-1011, 2013.

[2] M. Rosman, O. Rachminov, O. Segal, and G. Segal, "Prolonged patients’ In-Hospital Waiting Period after discharge eligibility is associated with increased risk of infection, morbidity and mortality: a retrospective cohort analysis," BMC Health Serv. Res., vol. 15, no. 1, p. 246, 2015.

[3] A. Rojas‐García, S. Turner, E. Pizzo, E. Hudson, J. Thomas, and R. Raine, "Impact and experiences of delayed discharge: A mixed‐studies systematic review," Health Expectations, vol. 21, no. 1, pp. 41-56, 2018.

[4] J. Gaughan, H. Gravelle, and L. Siciliani, "Testing the bed‐blocking hypothesis: does nursing and care home supply reduce delayed hospital discharges?," Health Econ., vol. 24, pp. 32-44, 2015.

[5] R. R. Nasrabad, "Introducing a new nursing care model for patients with chronic conditions," Electronic physician, vol. 9, no. 2, p. 3794, 2017.