| 摘要: | 背景
先天性心臟病在台灣新生兒的盛行率為千分之十三，而顯著先天性心臟病的盛行率為千分之一點五。若無法及時診斷先天性心臟病，病人在嬰幼兒時期可能會出現心臟衰竭的症狀，造成生長發育遲緩，甚至是造成死亡。標準的心臟超音波檢查價格昂貴，操作者稀少，無法大規模推廣。目前的常規篩檢方式是血氧濃度測試，但其僅能篩檢出部分嚴重的先天性心臟病，若為左至右分流的先天性心臟病則無法檢出。心電圖為低成本的檢查，目前已有利用心電圖預測先天性心臟病的人工智慧模型，但其對象多為學齡期以上的兒童及成人。針對嬰幼兒的心電圖預測模型仍待發展。此外，顯著先天性心臟病童的心電圖與正常嬰幼兒的心電圖差距較大，顯著先天性心臟病也才需要早期介入治療。本研究建立0至5歲嬰幼兒顯著右心疾病(如心房中膈缺損、法洛氏四重症)及顯著左心疾病(如心室中膈缺損、開放性動脈導管)的預測模型，並依上述模型建立利用心電圖預測顯著先天性心臟病(包含左心及右心疾病)的人工智慧模型。
研究材料及方法
本研究收集2013年1月至2020年1月於林口長庚醫院內接受心導管檢查或治療的先天性心臟病住院病人及2020年12月至2021年3月於林口長庚醫院接受門診心臟超音波檢查的先天性心臟病人的心電圖資料作為實驗組，並收集2020年12月至2021年3月於林口長庚醫院接受門診心臟超音波檢查的正常心臟結構病人的心電圖資料作為對照組。將先天性心臟病分為右心與左心疾病兩分類，再依是否出現血行動力學顯著影響分為輕微及顯著兩次分組。將心電圖訊號進行連續小波轉換獲得頻譜圖後，藉Resnet-18、Inceptionresnet-V2及Nasnetmobile等三個預訓練模型以遷移式學習探討輕微先天性心臟病是否能由模型鑑別，並建立預測0至5歲嬰幼兒是否有顯著先天性心臟病的模型。
結果
本研究共收集1035位0至5歲病人的心電圖，其中234位正常心臟結構、100位輕微右心疾病、291位顯著右心疾病、141位輕微左心疾病、269位顯著左心疾病。針對正常心電圖及右心疾病的組合，預測顯著右心疾病最好的是基於Resnet-18預訓練模型，具有正確率(Accuracy rate) 0.789 ± 0.009 (95%信賴區間0.778—0.799)，敏感度(Sensitivity) 0.735 ± 0.032 (95%信賴區間0.695—0.775)，特異性(Specificity) 0.839 ± 0.022 (95%信賴區間0.812—0.865)，F1分數(F1 score) 0.770 ± 0.014 (95%信賴區間0.753—0.787)，馬修斯相關係數(Matthews correlation coefficient) 0.579 ± 0.016 (95%信賴區間0.558—0.599)，ROC曲線下方面積(area under the receiver operating characteristic, AUROC) 0.852 ± 0.011 (95%信賴區間0.838—0.866)。針對正常心電圖及左心疾病的組合，預測顯著左心疾病最好的是基於Inceptionresnet-v2預訓練模型，具有正確率0.744 ± 0.033 (95%信賴區間0.704—0.785)，敏感度(Sensitivity) 0.696 ± 0.038 (95%信賴區間0.512—0.638)，特異性(Specificity) 0.779 ± 0.043 (95%信賴區間0.725—0.832)，F1分數0.695 ± 0.038 (95%信賴區間0.652—0.738 )，馬修斯相關係數 0.476 ± 0.065 (95%信賴區間0.396—0.556)，ROC曲線下方面積0.816 ± 0.035 (95%信賴區間0.773—0.859)。針對臨床可用設計，以檢出顯著先天性心臟病為目標的最佳模型為基於Resnet-18，左心及右心疾病訓練模型異常結果的聯集，有著正確率0.736 ± 0.031 (95%信賴區間0.697—0.775)，敏感度(Sensitivity) 0.709 ± 0.070 (95%信賴區間0.622—0.795)，特異性(Specificity) 0.768 ± 0.032 (95%信賴區間0.728—0.808)，F1分數0.742 ± 0.044 (95%信賴區間0.687—0.796)，馬修斯相關係數 0.535 ± 0.118 (95%信賴區間0.388—0.682)。
結論
針對0-5歲的嬰幼兒，利用十二導程心電圖可以預測顯著先天性心臟病。以人工智慧預測臨床場域中顯著先天性心臟病的模型能夠增進篩檢效能，使病人能於疾病的早期就接受介入治療。本模型針對顯著先天性心臟病進行篩檢的特性也可以和常規的血氧濃度篩檢測試搭配，在缺乏兒童心臟專家的偏遠地區可以減少轉診需求，提升整體醫療品質及滿意度。
Background:
The prevalence rate of congenital heart disease in newborns in Taiwan is 1.3 per thousand, and the prevalence rate of significant congenital heart disease is 0.15 per thousand. If congenital heart disease is not diagnosed in a timely manner, patients may develop symptoms of heart failure during infancy, leading to growth and developmental delays, and even death. Standard echocardiography for the heart is expensive and requires scarce operators, making it difficult to be widely implemented. The current routine screening method is the measurement of blood oxygen saturation, but it can only detect some severe cases of congenital heart disease and cannot detect left-to-right shunt defects. Electrocardiography (ECG) is a low-cost examination, and artificial intelligence models using ECG to predict congenital heart disease have been developed, but they mainly focus on school-age children and adults. There is still a need for the development of ECG prediction models for infants and young children. Furthermore, there is a significant difference between the ECGs of children with significant congenital heart disease and those of normal infants. Significant congenital heart disease requires early intervention. In this study, we aimed to establish predictive models for significant right heart disease (such as atrial septal defects and Tetralogy of Fallot) and significant left heart disease (such as ventricular septal defects and patent ductus arteriosus) in infants aged 0-5 years. We also aimed to develop an artificial intelligence model using ECG to predict significant congenital heart disease (including left and right heart diseases) based on the models mentioned before.
Materials and Methods:
In this study, we collected ECG data of patients with congenital heart disease who underwent cardiac catheterization or treatment at Linkou Chang Gung Memorial Hospital from January 2013 to January 2020, as well as ECG data of patients with congenital heart disease who underwent outpatient echocardiography at Linkou Chang Gung Memorial Hospital from December 2020 to March 2021. ECG data of patients with structurally normal hearts who underwent outpatient echocardiography at the same hospital from December 2020 to March 2021 were collected as the control group. Congenital heart disease was classified into right heart and left heart diseases, and further divided into mild and significant groups based on the presence of hemodynamic significance. After performing continuous wavelet transform on the ECG signals to obtain spectrograms, three pre-trained models, Resnet-18, Inceptionresnet-V2, and Nasnetmobile, were used for transfer learning to evaluate if mild congenital heart disease could be discriminated. Furthermore, model were established to predict significant congenital heart disease in infants aged 0-5 years.
Results:
A total of 1035 ECGs from patients aged 0-5 years were collected in this study, including 234 with structurally normal hearts, 100 with mild right heart disease, 291 with significant right heart disease, 141 with mild left heart disease, and 269 with significant left heart disease. For the combination of normal ECGs and right heart disease, the best prediction of significant right heart disease was achieved using the pre-trained Resnet-18 model, with an accuracy rate of 0.789 ± 0.009 (95% confidence interval [CI]: 0.778-0.799), sensitivity of 0.735 ± 0.032 (95% CI: 0.695-0.775), specificity of 0.839 ± 0.022 (95% CI: 0.812-0.865), F1 score of 0.770 ± 0.014 (95% CI: 0.753-0.787), Matthews correlation coefficient of 0.579 ± 0.016 (95% CI: 0.558-0.599), and area under the receiver operating characteristic (AUROC) of 0.852 ± 0.011 (95% CI: 0.838-0.866). For the combination of normal ECGs and left heart disease, the best prediction of significant left heart disease was achieved using the pre-trained Inceptionresnet-v2 model, with an accuracy rate of 0.744 ± 0.033 (95% CI: 0.704-0.785), sensitivity of 0.696 ± 0.038 (95% CI: 0.512-0.638), specificity of 0.779 ± 0.043 (95% CI: 0.725-0.832), F1 score of 0.695 ± 0.038 (95% CI: 0.652-0.738), Matthews correlation coefficient of 0.476 ± 0.065 (95% CI: 0.396-0.556), and AUROC of 0.816 ± 0.035 (95% CI: 0.773-0.859). For clinically applicable model to predict significant congenital heart disease, the best model was the union of positive results of left heart and right heart diseases using the pre-trained Resnet-18 model, with an accuracy rate of 0.736 ± 0.031 (95% CI: 0.697-0.775), sensitivity of 0.709 ± 0.070 (95% CI: 0.622-0.795), specificity of 0.768 ± 0.032 (95% CI: 0.728-0.808), F1 score of 0.742 ± 0.044 (95% CI: 0.687-0.796), Matthews correlation coefficient of 0.535 ± 0.118 (95% CI: 0.388-0.682).
Conclusion:
For infants aged 0-5 years, significant congenital heart disease can be predicted using twelve-lead ECGs. The use of artificial intelligence models to predict significant congenital heart disease in a clinical setting can enhance screening efficiency and enable early intervention for patients. The characteristics of this model for screening significant congenital heart disease can be combined with conventional blood oxygen saturation screening tests, reducing the need for referrals in remote areas with a shortage of pediatric cardiac experts and improving overall healthcare quality and satisfaction. |
