運用LSTM建立血液腫瘤科病人初次化學治療後口腔黏膜炎發生之預測性模型

目的：罹患癌症的人數逐年上升，其主要治療方式為化學藥物治療，口腔黏膜炎為治療後常見的副作用，其可能原因為化學治療而抑制免疫系統及骨髓的功能，當骨髓幹細胞遭受破壞，無法產生紅血球、白血球、血小板，導致嗜中性白血球低下及血小板不足，間接破壞口腔黏膜容易受細菌等微生物侵犯而感染，因此產生口腔黏膜炎。當發生口腔黏膜炎時，大部分病人會有疼痛、甚至影響進食，可能導致營養不良，若免疫力低下且併發有感染的情況時，有可能需中斷療程，甚至有生命的危害，住院天數延長、醫療費用增加等。因此本研究的目的在於利用深度學習之LSTM學習血球檢驗值變化與口腔黏膜炎等級變化的時間序列關係，產生口腔黏膜炎變化的預測模型。材料與方法：從中部某醫學中心的臨床資料庫收集資料，申請條件為血液腫瘤科住院病人初次有施打化學藥物且有口腔黏膜評估等級，排除接受放射線治療，後續每次住院於血液腫瘤科別的病人。資料數據排除門診檢驗值、一入院就有口腔黏膜炎、住院期間只有一筆抽血值、兩次抽血值之間小於七天的區段，共460人次(無發生254人次、有發生206人次)，資料10,119筆數(無發生4,982筆、有發生5,317筆)。接著，我們先利用視覺化方式，將血液腫瘤病患之血液檢驗數據：白血球(WBC)、絕對嗜中性白血球數目(ANC)、血紅素(HGB)、血小板(PLT)的檢驗值，由檢驗值變化的視覺化圖形與口腔黏膜炎變化的發生彼此交互比對，使得專家能夠直觀地觀察，找出彼此之間的規則。接著使用深度學習之LSTM(長短期記憶模型)學習檢驗值變化與口腔黏膜炎等級變化的時間序列關係，產生口腔黏膜炎變化的預測模型。結果與討論：本論文以收案結果80%的資料做為訓練組，20%的資料為測試組，訓練完畢後以測試組做為驗證，發現LSTM的預測正確率為92%。研究結果發現，ANC的變化幅度較為劇烈時，口腔黏膜炎等級的變化也跟著發生，所以ANC為主要的預測因子，其次為WBC是次要的預測因子。由於ANC是WBC推算的結果，具高度共線性，因此預測能力一致，再其次則是PLT(血小板)，HGB(血紅素)的預測能力不顯著。結論：本論文以Matlab所建立的預測模型雖然尚未將正確率提升到臨床可接受的程度，然而本論文只有單一醫院的數據，建議將來可以擴大收案來源到其他醫院，如此應該可將預測正確率提高，更符合臨床的需求。

關鍵字

口腔黏膜炎；預測模型；長短期記憶；時間序列

並列摘要

Purpose: The number of people suffering from cancer is increasing year by year. The main treatment is chemotherapy. Oral mucositis is a common side effect after treatment. The possible cause is that chemotherapy inhibits the function of the immune system and bone marrow. When the bone marrow stem cells are destroyed, it cannot be It produces red blood cells, white blood cells, and platelets, leading to low neutrophils and insufficient platelets, indirectly destroying the oral mucosa and being easily infected by bacteria and other microorganisms, resulting in oral mucositis. When oral mucositis occurs, most patients will have pain and even affect eating, which may lead to malnutrition. If the immune system is low and infections are complicated, the treatment may need to be interrupted, even life-threatening, and the length of hospital stay will be extended. , Medical expenses increase, etc. Therefore, the purpose of this study is to use the deep learning LSTM to learn the time series relationship between changes in blood cell test values and changes in oral mucositis grades to generate a prediction model for changes in oral mucositis. Materials and methods: Collect data from the clinical database of a medical center in the central part of China. The application conditions are that the inpatients in the department of hematology and oncology have been given chemical drugs for the first time and have an oral mucosal assessment grade, excluded from receiving radiotherapy, and each subsequent hospitalization in the department of hematology and oncology Patient. The data excludes outpatient test values, oral mucositis upon admission, only one blood draw value during hospitalization, and the interval between two blood draw values less than seven days, a total of 460 people (no occurrence of 254 person-times, 206 person-times) , The number of data is 10,119 (4,982 without occurrence, 5,317 with occurrence). Next, we first use the visualization method to visualize the blood test data of hematological tumor patients: white blood cell (WBC), absolute neutrophil number (ANC), hemoglobin (HGB), platelet (PLT) test value, and the test value The changing visual patterns and the occurrence of changes in oral mucositis are interactively compared with each other, allowing experts to observe intuitively and find out the rules between each other. Then use the deep learning LSTM (Long Short-Term Memory Model) to learn the time series relationship between the changes in the test value of the oral mucositis and the changes in the grade of oral mucositis to generate a prediction model of the changes in oral mucositis. Results and discussion: In this paper, 80% of the received results are used as the training group and 20% as the test group. After the training is completed, the test group is used as verification. It is found that the LSTM prediction accuracy rate is 92%. The results of the study found that when the amplitude of ANC changes more drastically, changes in the grade of oral mucositis also occur, so ANC is the main predictor, followed by WBC as the secondary predictor. As ANC is the result of WBC calculation, it has a high degree of collinearity, so the predictive ability is consistent, followed by PLT (platelet), HGB (hemoglobin), the predictive ability is not significant. Conclusion: Although the prediction model established by Matlab in this paper has not improved the accuracy rate to a clinically acceptable level, this paper only has data from a single hospital. It is recommended that the source of admissions can be expanded to other hospitals in the future, so that the prediction should be correct. The rate is increased to meet the clinical requirements.