應用深度學習於血液腫瘤病患臨終警示系統

End-of-Life Warning System on Hematological Malignancy Patients Using Deep Learning

吳翊群(Yi-Chun Wu)

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(2019年)

https://doi.org/10.6342/NTU201903295

摘要

近年來，台灣致力於發展安寧緩和醫療並得到相當的成效。然而在臨床
上我們發現血液腫瘤病患簽署DNR 的時間過晚，超過 50%的病人在死亡前一週以內才簽署DNR ，其主要原因是目前臨床上沒有準確的血液腫瘤預後資訊。然而，過晚簽署DNR 可能導致受到無效醫療的機會增加，醫療花費也會跟著增加，且血液腫瘤病情起伏較大，導致臨床上血液腫瘤病患常面臨到轉介安寧的時間已經過晚的問題。如果能夠有一個準確的臨終預測系統，能夠在病人接近臨終階段時發出警示，及早轉介安寧，便可改善以上這些狀況。
現今，越來越多醫學研究運用深度學習結合電子病歷，且得到相當好的結果。本研究提出一個基於長短期記憶網路 (LSTM) 模型，將血液腫瘤病患之基本資料、生命徵象、檢驗值、身體評估、嗎啡類藥物紀錄，組合在一起輸入模型，輸出為一個風險值，若風險值高於某門檻值，則發出警訊。我們的模型經過評估能有不錯的表現。此外，我們也支援 FHIR 輸入的服務，讓模型更容易被應用及具彈性。

關鍵字

血液腫瘤 ； 長短期記憶網路 ； 安寧緩和照護 ； 臨終預測 ； 深度學習

並列摘要

In recent years, Taiwan has been committed to the development of palliative care and has achieved considerable results. However, in clinical practice, we found that patients with hematological malignancies signed DNR late. More than 50% of patients signed DNR within one week before death. The main reason is that there is no accurate prognosis of hematological malignancies in the clinic. Signing DNR late may lead to increased chances of receiving medical futility and also increased medical costs. What’s more, the illness trajectory of hematological malignancies is fluctuating, leading to late referral to palliative care. If we have an accurate end-of-life prediction system that can alert us when the patient is nearing the end of life, and refer to palliative care as early as possible, these conditions can be improved.
Today, more and more medical research uses deep learning combined with electronic medical records, and has achieved fairly good results. In our research, we propose a long short term memory network (LSTM) model that input the combination of patients’ demographics, vital signs, laboratory tests, physical examination, illness categories, and opioids analgesics usage of patients with hematological malignancies and output a risk score. If the risk score is higher than the threshold, a warning will be sent. After our evaluations, the model can perform well. Moreover, our model support FHIR input services to make the model more flexible and applicable.

並列關鍵字

Hematological malignancies ； End-of-life prediction ； Palliative care ； Deep learning ； LSTM

參考文獻

[1] Harris I., Murray S.A. Can palliative care reduce futile treatment? A systematic review.

[2] Temel JS, Greer JA, Muzikansky A, et al. Early palliative care for patients with metastatic non-small-cell lung cancer.

[3] Garrido MM, Balboni TA, Maciejewski PK, Bao Y, Prigerson HG. Quality of Life and Cost of Care at the End of Life: The Role of Advance Directives.

[4] Maetens A, Beernaert K, De Schreye R, et al Impact of palliative home care support on the quality and costs of care at the end of life: a population-level matched cohort study.

[5] 溫純芳, 林文德(2014) 癌末死亡病人簽署不施行心肺復甦時點與醫療利用之相關性研究