肺炎住院病人14天再入院之分析與應用：人工神經網絡

A Predictive Model for Detecting Patients with Pneumonia of 14-Day Hospital Readmission using Artificial Neural Networks (ANN)

在醫療機構中，非計劃再入院情形十分頻繁且耗醫療資源。大多數減少再入院的解決方案，都是在病人出院前的加強醫療照顧及出院後的住院服務或關心。對於臨床醫生在住院期間早期識別高危非計劃再入院的可能病人線索十分有限。本研究目的，是為肺炎住院病人建立14天非計劃再入院的預測模式。由2016年至2018年在台灣南部三所醫院中下載ICD-10：J12－J18的肺炎住院病人資料。總計21,892件[UPRA＝1208（6％）]沒有遺漏值的資料。在MS Excel中，建立人工神經網絡（ANN）預測模式，使用Weka軟體粹取14天非計劃再入院的特徵變數，並比較其他預測模式與ANN模式之差異，指標包括敏感度，特異度和ROC曲線下的面積。然後開發一具預測非計劃再入院的應用模組，將模式的估計參數納入系統做為預測肺炎病人非計劃再入院之分類。結果顯示（1）本研究粹取出17個特徵變數；（2）在ANN模式中ROC曲線下面積比其他預測模式皆高，達0.73（95％ CI 0.72－0.74）；（3）開發預測非計劃再入院的應用程式，可套用至醫院的資訊系統，該模式具有53個估計參數，可提高非計劃再入院的準確度。該應用模式將幫助臨床醫生提早預測住院的肺炎患者，並讓臨床醫生得以在病人出院前及後制定照護準備及其計劃。

Unplanned hospital readmission (UPRA) is frequent and costly in healthcare settings. Most readmission reduction solutions are on complementing inpatient care with enhanced care transition and post-discharge interventions. Few clues were provided for clinicians to identify high-risk patients early during hospitalization. The study was aimed to build a predictive model for early detection of 14-day UPRA for patients with pneumonia. We downloaded data of pneumonia inpatients with ICD-10: J12*-J18* in three hospitals in Southern Taiwan during the years from 2016 to 2018. A total of 21,892 cases (1208 (6%) for UPRA) were eligible without missing data. Weka software (v.3.8.4) was used to extract feature variables and classify UPRA and Non-UPRA in 14 days after discharge from hospitalization in comparison with the predictive model with artificial neural network (ANN) developed in MS Excel using several model indicators including sensitivity, specificity, and area under the receiver operating characteristic curve, AUC). An app predicting UPRA was then developed involving the model's estimated parameters as a website prediction and classification. The results show that (1) 17 feature variables extracted from this study in the ANN model yielded a higher AUC of 0.73 (95% CI 0.72-0.74) than other counterparts, (2) an ready and available app for predicting UPRA was developed and used in health-care settings. The 17-item ANN model with the 53 parameters estimated by the ANN for improving the accuracy of UPRA has been developed. The app would help clinicians to predict inpatients with pneumonia at an early stage and allow clinicians to make preparedness plans before and after patient discharge from hospitalization accordingly.