以機器學習方法改善新生兒異位性皮膚炎風險之預測

Abstract

研究背景：異位性皮膚炎是最常見的皮膚疾病，許多患者在幼兒時期即已罹病，2009年一篇研究使用邏輯斯迴歸，建立新生兒6個月大時罹患異位性皮膚炎的風險預測模型，近年來機器學習演算法快速發展，廣泛應用於臨床醫學，有潛力改善2009年研究建立的異位性皮膚炎預測模型。 研究目的：使用機器學習演算法，以先天及後天的風險因子建立6個月大新生兒異位性皮膚炎的風險預測模型，並與邏輯斯回歸模型比較。 研究方法：研究使用與2009年研究相同的資料集，即「台灣世代研究」資料庫，此資料庫抽樣收集台灣88個鄉鎮於2005年出生的新生兒資料，於新生兒6個月大時進行首次調查。本研究先移除遺漏值，並以性別將資料分開，再以80%：20%的比例將原資料集切成訓練集與測試集。在機器學習模型部分，預測變數使用19個特徵，首先依照臨床上合理的切點將特徵離散化，並新定義倆性別的風險組，分為極低、低、高、極高四組，接著對訓練集進行100次「隨機特徵集選取、風險重標籤」以創造出新訓練集，使用新訓練集訓練XGBoost模型，並使用測試集以「5組驗證」的方式驗證模型，透過窮舉搜索的方式調整參數，找出預測各風險組的最佳模型，再定義「二模型混合預測」與「三模型混合預測」規則，採用三模型混合預測作為機器學習模型預測結果；在邏輯斯迴歸模型部分，使用與2009年相同的8個特徵訓練邏輯斯迴歸模型，並使用測試集以「5組驗證」的方式驗證模型。兩模型最終以混淆矩陣呈現，以對角線和、均方根誤差、加權誤差等作為模型表現的指標。 研究結果：本研究最終使用的資料集包含20235名新生兒（9607名女性，占47%），女性異位性皮膚炎比例約6%，男性異位性皮膚炎比例約8%，女性機器學習三模型混合預測準確率為：低風險組0.953、高風險組0.753、極高風險組0.706，混淆矩陣對角線和2.412，均方根誤差0.533，加權誤差0.302，女性邏輯斯迴歸模型預測準確率為：低風險組0.958、高風險組0.734、極高風險組0.644，混淆矩陣對角線和2.337，均方根誤差0.580，加權誤差0.370。男性機器學習三模型混合預測準確率為：低風險組0.963、高風險組0.811、極高風險組0.816，混淆矩陣對角線和2.590，均方根誤差0.394，加權誤差0.175，男性邏輯斯迴歸模型預測準確率為：低風險組0.936、高風險組0.772、極高風險組0.821，混淆矩陣對角線和2.529，均方根誤差0.412，加權誤差0.227。 結論：本研究將機器學習方法應用於一個具全國代表性的出生世代資料集，為6個月大的新生兒建立異位性皮膚炎風險預測模型，研究顯示機器學習模型比過去的邏輯斯迴歸模型表現更佳，有效提高預測準確率，可以協助臨床醫師預測新生兒罹患異位性皮膚炎的風險並採取預防措施。

Background: Atopic dermatitis (AD) is the most common skin disorder and many patients develop symptoms early. A risk prediction model of AD in 6-month-old newborns was established in 2009 using logistic regression (LR). Recently, machine learning (ML) methods keep gaining popularity and have been applied in various clinical settings. Whether ML can outperform LR remains inconclusive. Objective: To apply ML methods to set up AD risk prediction model among 6-month-old newborns based on hereditary and environmental risk factors, and to compare performance between ML model and LR model. Methods and Participants: Taiwan Birth Cohort Study (TBCS) was used in this study, same as the study in 2009. Babies born in 2005 in 88 townships in Taiwan were sampled and the first follow-up interview took place when the babies were 6 months old. Data with missing values were removed. The data were stratified based on gender and were split to a train set and a test set in 80-20 ratio. Nineteen features (risk factor) were included in the ML model. Feature discretization, 100 rounds of random feature set selection and AD risk level relabeling were performed sequentially to create a new train set. The ML model was trained on the new train set and was validated by 5-run validation on the test set. Through exhaustive grid search of parameters, the best model of each risk level was identified. We assigned prediction rules of 2-model and 3-model mixed prediction. The 3-model mixed prediction was the final ML model. The LR model was set up using the same 8 features as the study in 2009 and was validated by 5-run validation on the test set. Standardized confusion matrix was used to summarize the final prediction results of two models. Sum of diagonals, RMSE and weighted error were calculated to compare performance between ML and LR. Results: A total of 20235 newborns (9607 female [47%]) were analyzed. The AD percentage was about 6% in female and about 8% in male. The prediction accuracy of ML model of female was 0.953, 0.753, and 0.706 in low, high and very high risk group, respectively and the sum of diagonals, RMSE and weighted error were 2.412, 0.533 and 0.302, respectively. The prediction accuracy of LR model of female was 0.958, 0.734 and 0.644 in low, high and very high risk group, respectively and the sum of diagonals, RMSE and weighted error were 2.337, 0.580 and 0.370, respectively. The prediction accuracy of ML model of male was 0.963, 0.811, and 0.816 in low, high and very high risk group, respectively and the sum of diagonals, RMSE and weighted error were 2.590, 0.394 and 0.175, respectively. The prediction accuracy of LR model of male was 0.936, 0.772 and 0.821 in low, high and very high risk group, respectively and the sum of diagonals, RMSE and weighted error were 2.529, 0.412 and 0.227, respectively. Overall, compared to the LR model, the ML model of female had 3.2% higher sum of diagonals, 8.1% lower RMSE and 18.4% lower weighted error. Compared to the LR model, the ML model of male had 2.4% higher sum of diagonals, 4.4% lower RMSE and 23% lower weighted error. Conclusions: In this study, a novel ML approach combining with XGBoost was applied on a national representative birth cohort to set up AD risk prediction models in 6-month-old newborns. For both genders, the ML model had better overall performance than the LR model. Our ML model can help clinicians stratify newborns into different risk levels with high accuracy and help clinicians design preventive strategies based on the risk.