以機器學習模型預測加護病房內肺炎病童死亡風險

Using Machine Learning Model to Predict Mortality for Children with Pneumonia in the Intensive Care Unit

林湘容(Siang-Rong Lin)

指導教授 ： 周佳靚

國立臺灣大學/工學院/應用力學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202201973

摘要

肺炎在台灣與全球是造成兒童住院與死亡的主要原因之一。能及時辨別肺炎病童的嚴重度與死亡率對於醫療決策、提供家屬諮詢與討論病情是至關重要的，並可能改善治療後的結果。但目前未有適用於台灣臨床環境中的兒童呼吸道感染嚴重程度的評分系統。近年來隨著電子病歷系統的普及與機器學習技術的蓬勃發展，已有諸多研究展示數據挖掘與機器學習方法應用於臨床的可能性，但在兒科研究中的應用範圍尚未得到充分的探索。因此，我們旨在開發一種機器學習模型來預測肺炎病童在加護病房中的死亡率，並辨別重要的臨床指標以支持醫療決策。我們在2010年1月至2019年12月期間從國立臺灣大學附屬醫院中納入了932名肺炎病童，共組成1,231筆加護病房入院紀錄與1,144入院紀錄。我們感興趣的結果是肺炎病童在加護病房的死亡率及其在接下來24小時內的死亡率。藉由病患之生理數據與機器學習方法提供個人化的死亡率評估，並計算特徵重要性以獲得模型解釋性。首先，本研究開發了加護病房住院早期的死亡率預測模型。該模型顯示出良好的預測性能，受試者工作特徵曲線下面積為0.85。另外，我們更進一步開發了動態死亡預測模型，該模型考慮了病童隨時間變化的生理狀況，並實現了受試者工作特徵曲線下面積為0.98 的出色預測性能（95% CI，0.96-0.99）。此動態預測模型能隨著肺炎病童在不同時間下的病程發展，提供能每日更新的個人死亡率評估，以利醫師提出即時且精準的臨床決策。該模型性能優於現有的嚴重程度評分系統，並且更符合台灣兒童的臨床特徵。透過機器學習模型，我們識別了重要的死亡風險因子，包括較低的體溫、收縮壓和較高的肌酐檢驗值等，符合醫師的臨床知識與經驗。

關鍵字

兒科患者 ； 肺炎 ； 加護病房 ； 死亡預測 ； 機器學習 ； 臨床指標

並列摘要

Identifying the severity and mortality of children with pneumonia in time is crucial to medical decision-making, counseling, and discussion of the disease situation. It has the opportunity to improve the outcomes after treatment. However, there is no scoring system for the severity of childhood respiratory infections in the clinical setting in Taiwan. In recent years, with the popularity of electronic health records and the boom in machine learning technology, numerous studies have demonstrated the potential of data mining and machine learning methods for clinical applications. Still, they have not been fully explored in pediatric research. Therefore, we aimed to develop a machine learning model to predict ICU mortality for children with pneumonia and discriminate important clinical indices for supported decision-making. We enrolled 1,231 ICU admissions from 1,144 hospitalizations and 932 unique patients with pneumonia in the cohort from National Taiwan University Hospital between January 2010 and December 2019. First, a mortality prediction model in the early stages of ICU admission was developed. The model showed good prediction performance with 0.85 of the area under the receiver operating characteristic curve (AUROC). Second, we further develop a dynamic prediction model that consider the patients’ disease situations over time and achieve an excellent AUROC of 0.98 (95% CI, 0.96-0.99). The model performance outperforms existing severity scoring systems and is better adapted to the clinical characteristics of Taiwanese children. The relative important clinical indices identified by the model, including lower body temperature, systolic blood pressure, and higher creatinine et al., are consistent with the physician’s clinical knowledge and experience.

並列關鍵字

pediatric patients ； pneumonia ； intensive care unit ； mortality prediction ； machine learning ； clinical indices

參考文獻

1. Organization, W.H., World health statistics 2019: monitoring health for the SDGs, sustainable development goals. 2019: World Health Organization.

2. Organization, W.H., World health statistics 2021: monitoring health for the SDGs, sustainable development goals. 2021: World Health Organization. 7-21.

3. Welfare, M.o.H.a. 2020 Cause of Death Statistics. 2021 January 18, 2022]; Available from: https://www.mohw.gov.tw/lp-5256-2.html.

4. Liu, L., et al., Global, regional, and national causes of under-5 mortality in 2000–15: an updated systematic analysis with implications for the Sustainable Development Goals. The Lancet, 2016. 388(10063): p. 3027-3035.

5. Eber, E. and F. Midula, ERS Handbook of paediatric respiratory medicine. 2021: European Respiratory Society. 253-254.