應用深度學習於感染預警預測模型之建構

Development of Infection Warning Predicting Model Using Deep Learning

張涵(Hang Jang)

指導教授 ： 賴飛羆

共同指導教授 ： 陳宜君(Yee-Chun Chen)

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2020年)

https://doi.org/10.6342/NTU202002079

摘要

血流感染在各年齡層中都展現了高致病率以及高死亡率的特性，尤其對於免疫系統受損的病人，血流感染可能會演變為危及病人生命的侵襲性感染症，例如念珠菌血症以及敗血症，並且對於住院天數、醫療成本以及死亡率的增加都有重大影響。根據國際感染病學會（International Society of Infection Disease）的定義，血流感染意指血液培養呈現一種或多種病原體陽性反應，伴隨全身性感染跡象，例如發燒、寒顫以及／或低血壓。全球每年約有三千萬個病人發生血流感染，其中約有六百萬位病人因此死亡，以及約三百萬個新生兒和約一百二十萬個兒童轉變為敗血症。然而，現行的診斷方法有諸多限制，以當今的參考標準：血液培養為例，其檢驗時程較長，且針對特定病原菌，例如念珠菌，其敏感度並不理想。相較之下，新興的檢驗技術，例如聚合酵素連鎖反應（Polymerase Chain Reaction），雖能有效降低檢驗時程且提高敏感度，但因缺乏標準的作業流程以及容易受到污染等原因，至今仍無法完全取代血液培養作為臨床標準。較長的檢驗時程可能導致延遲治療，進而使得病人的預後惡化，但目前常見的經驗性治療經常導致藥物的過度使用。若能儘早偵測甚至鑑別病原並且準確用藥，則能有效地改善宿主病況。

今日許多醫學研究結合電子病歷以及深度學習，都得到良好的結果。本研究以 2009 年 04 月 13 日至 2017 年 12 月 31 日為研究區間，針對台大醫院住院病人血液培養陽性者，搜集其電子病歷資料，建構出一個長短期記憶網路（Long-Short Term Memory）預測模型。利用病人的基本資料、住院資料、醫療處置（包含手術記錄、用藥紀錄等）以及檢驗資料等作為輸入值，此模型將輸出一個風險預測值，若預測值高於設定的門檻值，則發出血流感染的警訊，可作為醫師是否要提早介入的參考依據。其中，不進行特徵選取（Feature Selection）所建構出之長短期記憶網路模型，搭配相同設定的風險門檻，其敏感度以及陽性預測值最高，分別來到84.87%以及85.02%，在臨床可接受範圍內，是最佳的預測模型。此研究的主要貢獻為針對細菌血症以及念珠菌血症病人，應用深度學習建構感染預警預測模型。

關鍵字

血流感染 ； 預測模型 ； 念珠菌血症 ； 細菌血症 ； 長短期記憶網路 ； 深度學習

並列摘要

Bloodstream infections (BSI) exhibit high morbidity and mortality in people of all ages, especially in patients with the compromised immune system. BSI may evolve into invasive infections that threaten the lives of patients, for instance, candidemia and sepsis, and it has a significant impact on the length of hospitalization, medical cost, and mortality as well. According to the International Society of Infection Disease, BSI is defined as “one or more positive blood cultures associated with systemic signs of infection such as fevers, chills and/or hypotension.” In worldwide, BSI affects approximately 30 million people and leads to 6 million deaths, with around 3 million newborns and 1.2 million children suffering from sepsis annually. The current diagnostic techniques have many limitations. Taking the current reference standard: blood culture, as an example, the turnaround time is slow, and the sensitivity to specific pathogens, such as Candida, is not ideal. In contrast, emerging techniques such as polymerase chain reaction (PCR) has faster turnaround time and higher sensitivity, but due to the lack of universal protocol and susceptibility to contamination and other characteristics, it cannot completely replace blood culture as a clinical standard so far. However, slow turnaround time may lead to delayed treatment, which in turn worsens the patient’s prognosis, but empirical treatments often lead to overuse of drugs. If the pathogens can be detected and even identified in the early stage, in cooperation with the customization of antibiotic therapy, it might benefit patients in several aspects, furthermore, improve patients’ outcomes.

Today, many medical research projects have combined electronic medical records and deep learning techniques, and most of which have achieved fairly good performances. In this study, we collect the electronic medical records of inpatients from April 13th, 2009 to December 31st, 2017 in National Taiwan University Hospital, of which with positive blood culture results. We construct a long-short term memory prediction model, which utilizes patients’ demographic data, hospitalization data, medical order records (including surgical records, medication records, etc.), lab data, etc. as input, and the model will output a risk score accordingly. If the risk score is higher than the set threshold, it will issue a warning sign of BSI, which could be a reference for physicians to determine whether to intervene or not. The most outperforming model is constructed without prior feature selection under the same sets of risk thresholds, which has the highest sensitivity (84.87%) and positive predictive value (85.02%) and are in the clinically acceptable range. The main contribution of this study is to use deep learning techniques to construct the infection warning prediction models for BSI patients, including bacteremia and candidemia.

並列關鍵字

Bloodstream Infection ； Prediction Model ； Candidemia ； Bacteremia ； Long-Short Term Memory ； Deep Learning

參考文獻

1. Deku, J.G., et al., The Epidemiology of Bloodstream Infections and Antimicrobial Susceptibility Patterns: A Nine-Year Retrospective Study at St. Dominic Hospital, Akwatia, Ghana. Journal of Tropical Medicine, 2019. 2019: p. 6750864.

2. Sheng, W.H., et al., Comparative impact of hospital-acquired infections on medical costs, length of hospital stay and outcome between community hospitals and medical centres. J Hosp Infect, 2005. 59(3): p. 205-14.

3. Fleischmann, C., et al., Assessment of Global Incidence and Mortality of Hospital-treated Sepsis. Current Estimates and Limitations. Am J Respir Crit Care Med, 2016. 193(3): p. 259-72.

4. Fleischmann-Struzek, C., et al., The global burden of paediatric and neonatal sepsis: a systematic review. The Lancet Respiratory Medicine, 2018. 6(3): p. 223-230.

5. Kullberg, B.J. and M.C. Arendrup, Invasive Candidiasis. New England Journal of Medicine, 2015. 373(15): p. 1445-1456.