運用人工智慧管理心肌梗塞患者：利用初期緊急例行實驗室血液測試

Abstract

背景： 為了診斷患有非ST區段上升急性心肌梗塞的患者，心臟肌鈣蛋白的測量具有重要價值，然而約三分之一的患者在檢測第二次心臟肌鈣蛋白值之後始得確診。在本研究中，我們引入了一些機器學習算法，結合初始高敏感度心肌鈣蛋白T的測量以及人口特徵和常規實驗室檢查，這些檢查通常在一小時內產生結果。我們的目標是評估常規實驗室檢查是否通過機器學習，從而對於預測心肌梗塞(AMI)有額外的診斷價值。這種方法旨在促進臨床實踐中快速的診斷決策。 方法： 這項回顧性研究利用了來自台灣大學醫學院附設醫院急診科的44,000名患者的初始高敏感度心肌鈣蛋白T數據來訓練非ST區段上升急性心肌梗塞機器學習模型。10,636名來自台大新竹分院的患者則被視為獨立的測試數據集。排除標準包括患有ST區段上升急性心肌梗塞的患者、缺乏至少兩次高敏感度心肌鈣蛋白T以及缺乏完整血球計數數據的患者。機器學習模型整合了23項常規實驗室檢查，藉以計算非ST區段上升急性心肌梗塞的可能性。比較了包括極端梯度提升、隨機森林和羅吉斯回歸在內的3種不同算法的性能，並評估包括受試者操作特徵曲線下面積(area under the receiver-operating-characteristic curve , AUC)和精確召回曲線的平均精度(average precision , AP)在內的診斷性能指標。為了提高模型的解釋性，分析了SHAP (Shapley Additive Explanations) 值，藉以解釋特定常規實驗室特徵對模型的影響。通過由風險分數的敏感性、陰性預測值、特異性或陽性預測值確定的閾值，將患者分為低、中、高風險組。 結果： 在2016年5月1日至2021年12月31日期間，共納入了15,096名患者，其中690名患者(4.6%)被診斷為非ST區段上升急性心肌梗塞。在獨立的引導抽樣測試數據集中，極端梯度提升模型的AUC值為0.862 (0.848-0.877)，而單獨使用初始高敏感度心肌鈣蛋白T時的AUC值為0.711 (0.691-0.730)。風險分數低於2.0者，在被歸類為低風險的患者中，有51.0%被有效排除具非ST區段上升急性心肌梗塞，此時陰性預測率為98.7。相反地，風險分數高於37.5者，在被歸類為高風險的患者中，有2.9%被確訂為非ST區段上升急性心肌梗塞，此時陽性預測率為73.3。 結論： 我們建立機器學習算法，結合了常規實驗室檢查，成功預測了疑似非ST區段上升急性心肌梗塞患者的風險。這種方法提供早期、快速且精確的心肌梗塞分診一種臨床可行的解決方案。

Background To diagnose patients with non-ST-segment elevation acute myocardial infarction (NSTEMI), cardiac troponin measurements play an important role, while about one-third of the patients remain undetermined until the second cardiac troponin values are assessed. In this study, we introduce some machine learning (ML) algorithms incorporating the initial measurement of high-sensitivity cardiac troponin T (0 h hs-cTnT) along with demographic features and routine laboratory tests that usually release data within an hour. Our aim is to evaluate whether routine laboratory tests provide additional diagnosis value to 0 h hs-cTnT by machine learning for AMI prediction. This approach could facilitate rapid rule-in and rule-out decisions in clinical practice. Method The retrospective study recruited 44,000 patients with 0 h hs-cTnT from the National Taiwan University Hospital (NTUH) emergency department for training NSTEMI machine learning models. Additional 10,636 patients from the NTUH Hsin-Chu Branch were recruited as an independent cohort. Exclusion criteria involved patients with ST-segment elevation myocardial infarction (STEMI) and those lacking serial hs-cTnT and complete blood count data. The efficacy of 3 algorithms, including eXtreme Gradient Boosting (XGBoost, XGB), Random Forest (RF), and Logistic Regression (LR), were compared to each other. The diagnostic performance metrics were examined, based on the area under the receiver-operating-characteristic curve (AUC) and the average precisions (AP) of the precision-recall curve. To evaluate the feasibility of this model, Shapley Additive Explanations (SHAP) values were analyzed to find out the impacts of specific routine laboratory feature. Low-, middle- and high-risk groups are defined based on the low and high thresholds of risk scores which were determined by sensitivity, negative predictive value (NPV), specificity, or positive predictive value (PPV). Findings The cohort of 15,096 patients was recruited from May 1, 2016, to December 31, 2021. Among them, 690 patients (4.6%) were diagnosed with NSTEMI. In an independent testing dataset with bootstrapping, the AUC value for XGBoost was 0.862 (0.848-0.877), compared to the AUC of 0.711 (0.691-0.730) when using 0 h hs-cTnT alone. When individuals with a risk score below 2.0 were classified as low-risk, 51.0% of patients were effectively excluded, given an NPV of 98.7. In contrast, when a risk score above 37.5 was defined as high-risk, 2.9% of patients were identified as high-risk, with a PPV of 73.3. Interpretation When incorporating with routine laboratory tests, our established machine learning algorithm could accurately predict the risk for patients with suspected NSTEMI. This approach offers a clinically valuable strategy for early, rapid, and precise triage of myocardial infarction.