目的

本研究旨在基于大数据和AI开发重症呼吸道感染患者在治疗策略（开始和撤离有创机械通气、抗生素降阶梯治疗）实施前的实时、个体化预后风险评估模型，为临床医师早期识别高危患者和根据不同预后风险患者制定个体化治疗策略提供实用工具。

方法

本研究使用了MIMIC-III、MIMIC-IV和eICU数据库中的患者数据进行分析。本研究采用了RandomizedSearchCV进行超参数优化寻找模型最佳参数，利用受试者工作特征曲线下的面积、敏感性、特异性、精确率、准确率、校准曲线和Brier分数等来评估模型的判别能力和预测校准度。本研究采用了K折交叉验证技术评估了模型的泛化能力。使用Shapley加性解释（Shapley Additive Explanations ,SHAP）值计算了特征的重要性。

结果

1. 重症呼吸道感染患者有创机械通气（Invasive Mechanical Ventilation，IMV）治疗开始时机预后评估：研究收集IMV治疗开始前24小时变量数据构建预后模型。其中随机森林（Random Forest，RF）、自适应提升算法(Adaptive Boosting, AdaBoost)、类别变量提升(Categorical Boosting,CatBoost)、轻量级梯度提升机（Light Gradient Boosting Machine，LGBM）、人工神经网络（Artificial Neural Network，ANN）和卷积神经网络（Convolutional Neural Network，CNN）模型的曲线下面积（Area Under the Curve ，AUC）分别为：RF - 0.84，Adaboost - 0.83，Catboost - 0.84，LGBM - 0.83，ANN - 0.82，CNN - 0.82，均表现出较好的区分能力。传统评分系统AUC明显低于机器学习模型。通过对eICU合作研究数据库进行外部验证，机器学习模型的AUC也表现良好，分别是RF（0.99）、Catboost（0.94）和ANN（0.88）。RF的brier分数为0.1233，表明RF在概率预测方面较为精准。10折交叉验证中RF模型，准确率和精确度均为0.8，SHAP分析表明，去甲肾上腺素注射速率、尿素氮、乳酸水平和呼吸频率是对模型结果最关键的4个预测因子。

2. 重症呼吸道感染患者抗生素降阶梯治疗(Antibiotic De-escalation Therapy,ADE)预后评估：研究收集进行ADE治疗时间点前的24小时变量数据构建预测模型。其中RF和逻辑回归（Logistic Regression，LR）的AUC均为0.80，极限梯度提升（eXtreme Gradient Boosting，XGBoost）和多层感知器（Multilayer Perceptron，MLP）的AUC均为0.81。RF模型在灵敏度（0.82）、特异度（0.71）、准确率（0.91）和精确度（0.71）上表现优秀，尤其在准确率和精确度方面具有明显优势。LR模型的Brier得分为0.150，表现较好，XGBoost的Brier得分为0.155。在临床决策曲线分析中，RF模型在几乎所有阈值下表现出较高的净收益。10折交叉验证中，RF模型的准确率0.84，精确度0.88。SHAP分析表明，序贯器官衰竭评分、血尿氮和年龄等特征对模型输出的影响较大，是对模型结果最关键的预测因子。

3. 重症呼吸道感染患者IMV撤离时机预后评估：研究收集撤离IMV前24小时变量数据构建预测模型。其中梯度提升模型(Gradient Boosting Machine,GBM)和XGBoost的AUC分别为0.88和0.87，表现最佳。XGBoost在准确性（0.80）和特异性（0.83）方面表现优异，F1得分为0.63，显示出均衡的性能。LGBM和LR模型的Brier得分分别为0.110和0.118，显示出较好的校准性。RF和XGBoost的Brier得分分别为0.113和0.123，表现出较高的校准准确性。10折交叉验证结果表明，XGBoost在准确率（0.84）和精确度（0.74）方面表现最为出色。SHAP分析显示，收缩压、呼吸率、心率和碱性磷酸酶是对模型影响最大的4个变量。

结论

本研究基于大数据和AI分别开发了急性重症呼吸道感染患者在开始和撤离有创机械通气、抗生素降阶梯治疗前的各自实时、个体化预后风险评估模型，模型均具有良好的效能、泛化能力和可解释性。为临床医师早期识别高危患者和根据不同预后风险患者制定个体化治疗策略提供实用工具。

Objective

This study aims to develop real-time, individualized prognostic risk assessment models for critically ill patients with respiratory infections, based on big data and AI, to assist clinicians in early identification of high-risk patients and in formulating individualized treatment strategies (initiation and discontinuation of invasive mechanical ventilation, and antibiotic de-escalation therapy).

Methods

Data from MIMIC-III, MIMIC-IV, and eICU databases were analyzed in this study. RandomizedSearchCV was employed for hyperparameter optimization to identify the best model parameters. Model performance was assessed using Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, specificity, precision, accuracy, calibration curve, and Brier score. K-fold cross-validation was used to evaluate the model’s generalization ability. Feature importance was calculated using Shapley Additive Explanations (SHAP) values.

Results

Prognostic assessment for initiation of invasive mechanical ventilation (IMV):

The study collected data from the 24 hours prior to the initiation of IMV and constructed prognostic models. The AUC for Random Forest (RF), Adaptive Boosting (AdaBoost), Categorical Boosting (CatBoost), Light Gradient Boosting Machine (LGBM), Artificial Neural Network (ANN), and Convolutional Neural Network (CNN) were as follows: RF - 0.84, AdaBoost - 0.83, CatBoost - 0.84, LGBM - 0.83, ANN - 0.82, CNN - 0.82, all showing good discriminatory ability. Traditional scoring systems performed significantly worse than machine learning models. External validation with the eICU database showed the models’ AUCs to be RF (0.99), CatBoost (0.94), and ANN (0.88). RF showed the lowest Brier score of 0.1233, indicating higher prediction accuracy in terms of probability. In

cross-validation, the RF model achieved an accuracy and precision of 0.8. SHAP analysis revealed that norepinephrine infusion rate, blood urea nitrogen, lactate levels, and respiratory rate were the most important predictive factors.

Prognostic assessment for antibiotic de-escalation therapy (ADE):

The study collected data from the 24 hours prior to the initiation of ADE and constructed predictive models. The AUCs for RF and Logistic Regression (LR) were 0.80, while the AUCs for Extreme Gradient Boosting (XGBoost) and Multilayer Perceptron (MLP) were 0.81. RF performed excellently with a sensitivity of 0.82, specificity of 0.71, accuracy of 0.91, and precision of 0.71, showing a clear advantage in accuracy and precision. The Brier score for LR was 0.150, performing well, while XGBoost had a slightly lower Brier score of 0.155. Clinical decision curve analysis showed that the RF model provided higher net benefit across almost all thresholds. In cross-validation, RF achieved an accuracy of 0.84 and precision of 0.88. SHAP analysis showed that Sequential Organ Failure Assessment (SOFA), blood urea nitrogen, and age were key predictive factors influencing the model’s output.

Prognostic assessment for the discontinuation of IMV:The study collected data from the 24 hours prior to the discontinuation of IMV and constructed predictive models. The AUCs for Gradient Boosting (GBM) and XGBoost were 0.88 and 0.87, respectively, performing the best. XGBoost showed excellent performance with an accuracy of 0.80 and specificity of 0.83, and an F1 score of 0.63, indicating balanced performance. LGBM and LR models had Brier scores of 0.1104 and 0.1179, showing good calibration. RF and XGBoost models had Brier scores of 0.1125 and 0.1229, respectively, indicating high calibration accuracy. In cross-validation, XGBoost performed best in terms of accuracy (0.84) and precision (0.74). SHAP analysis revealed that systolic blood pressure, respiratory rate, and heart rate were significant factors influencing the model's predictions.

Conclusion

This study developed real-time, individualized prognostic risk assessment models for the initiation and discontinuation of invasive mechanical ventilation and antibiotic de-escalation therapy in critically ill patients with acute respiratory infections, based on big data and AI. The models demonstrate good performance, generalization ability, and interpretability, providing a practical tool for clinicians to identify high-risk patients early and tailor individualized treatment strategies based on different prognostic risks.