Table1_Machine learning models to predict 30-day mortality for critical patients with myocardial infarction: a retrospective analysis from MIMIC-IV database.pdf

BackgroundThe identification of efficient predictors for short-term mortality among patients with myocardial infarction (MI) in coronary care units (CCU) remains a challenge. This study seeks to investigate the potential of machine learning (ML) to improve risk prediction and develop a predictive model specifically tailored for 30-day mortality in critical MI patients.
MethodThis study focused on MI patients extracted from the Medical Information Mart for Intensive Care-IV database. The patient cohort was randomly stratified into derivation (n = 1,389, 70%) and validation (n = 595, 30%) groups. Independent risk factors were identified through eXtreme Gradient Boosting (XGBoost) and random decision forest (RDF) methodologies. Subsequently, multivariate logistic regression analysis was employed to construct predictive models. The discrimination, calibration and clinical utility were assessed utilizing metrics such as receiver operating characteristic (ROC) curve, calibration plot and decision curve analysis (DCA).
ResultA total of 1,984 patients were identified (mean [SD] age, 69.4 [13.0] years; 659 [33.2%] female). The predictive performance of the XGBoost and RDF-based models demonstrated similar efficacy. Subsequently, a 30-day mortality prediction algorithm was developed using the same selected variables, and a regression model was visually represented through a nomogram. In the validation group, the nomogram (Area Under the Curve [AUC]: 0.835, 95% Confidence Interval [CI]: [0.774–0.897]) exhibited superior discriminative capability for 30-day mortality compared to the Sequential Organ Failure Assessment (SOFA) score [AUC: 0.735, 95% CI: (0.662–0.809)]. The nomogram (Accuracy: 0.914) and the SOFA score (Accuracy: 0.913) demonstrated satisfactory calibration. DCA indicated that the nomogram outperformed the SOFA score, providing a net benefit in predicting mortality.
ConclusionThe ML-based predictive model demonstrated significant efficacy in forecasting 30-day mortality among MI patients admitted to the CCU. The prognostic factors identified were age, blood urea nitrogen, heart rate, pulse oximetry-derived oxygen saturation, bicarbonate, and metoprolol use. This model serves as a valuable decision-making tool for clinicians.

【背景】在冠心病监护病房（coronary care units, CCU）的心肌梗死（myocardial infarction, MI）患者中，高效识别短期死亡预测因子仍是一项挑战。本研究旨在探究机器学习（machine learning, ML）改善风险预测的潜力，并开发一款专门针对重症MI患者30天死亡率的预测模型。
【方法】本研究的研究对象为从重症医学信息库第四版（Medical Information Mart for Intensive Care-IV, MIMIC-IV）中提取的MI患者。将患者队列随机分层分为推导集（n=1389，占比70%）与验证集（n=595，占比30%）。通过极端梯度提升（eXtreme Gradient Boosting, XGBoost）与随机决策森林（random decision forest, RDF）方法识别独立危险因素。随后采用多因素logistic回归分析构建预测模型。利用受试者工作特征（receiver operating characteristic, ROC）曲线、校准曲线及决策曲线分析（decision curve analysis, DCA）等指标评估模型的区分度、校准度与临床实用性。
【结果】本研究共纳入1984例患者（平均年龄[标准差]：69.4[13.0]岁；女性659例，占比33.2%）。基于XGBoost与RDF构建的模型预测效能相近。随后利用筛选出的相同变量开发了30天死亡率预测算法，并通过列线图（nomogram）可视化呈现回归模型。在验证集中，该列线图的曲线下面积（Area Under the Curve, AUC）为0.835，95%置信区间（Confidence Interval, CI）为[0.774–0.897]，其对30天死亡率的区分能力优于序贯器官衰竭评分（Sequential Organ Failure Assessment, SOFA）（AUC=0.735，95%CI=(0.662–0.809)）。列线图的准确率为0.914，SOFA评分的准确率为0.913，二者均表现出良好的校准度。决策曲线分析显示，列线图的表现优于SOFA评分，在死亡率预测中可提供净临床获益。
【结论】基于机器学习的预测模型在预测CCU收治的MI患者30天死亡率方面具有显著效能。本研究识别出的预后因素包括年龄、血尿素氮（blood urea nitrogen）、心率、脉搏血氧饱和度（pulse oximetry-derived oxygen saturation）、碳酸氢盐（bicarbonate）水平及美托洛尔（metoprolol）使用情况。该模型可作为临床医师的实用决策辅助工具。