Campi Riccardo, De Santis Antonio, Colombo Paolo, Scarpazza Paolo, Masseroli Marco
Politecnico di Milano, Dipartimento di Elettronica, Informazione e Bioingegneria, Piazza Leonardo Da Vinci 32, Milano, MI, 20133, Italy.
Azienda Socio Sanitaria Territoriale della Brianza, Via Santi Cosma e Damiano 10, Vimercate, MB, 20871, Italy.
Comput Methods Programs Biomed. 2025 Mar;260:108574. doi: 10.1016/j.cmpb.2024.108574. Epub 2024 Dec 30.
Helmet-Continuous Positive Airway Pressure (H-CPAP) is a non-invasive respiratory support that is used for the treatment of Acute Respiratory Distress Syndrome (ARDS), a severe medical condition diagnosed when symptoms like profound hypoxemia, pulmonary opacities on radiography, or unexplained respiratory failure are present. It can be classified as mild, moderate or severe. H-CPAP therapy is recommended as the initial treatment approach for mild ARDS. Even though the efficacy of H-CPAP in managing patients with moderate-to-severe hypoxemia remains unclear, its use has increased for these cases in response to the emergence of the COVID-19 Pandemic. Using the electronic medical records (EMR) from the Pulmonology Department of Vimercate Hospital, in this study we develop and evaluate a Machine Learning (ML) system able to predict the failure of H-CPAP therapy on ARDS patients.
The Vimercate Hospital EMR provides demographic information, blood tests, and vital parameters of all hospitalizations of patients who are treated with H-CPAP and diagnosed with ARDS. This data is used to create a dataset of 622 records and 38 features, with 70%-30% split between training and test sets. Different ML models such as SVM, XGBoost, Neural Network, Random Forest, and Logistic Regression are iteratively trained in a cross-validation fashion. We also apply a feature selection algorithm to improve predictions quality and reduce the number of features.
The SVM and Neural Network models proved to be the most effective, achieving final accuracies of 95.19% and 94.65%, respectively. In terms of F1-score, the models scored 88.61% and 87.18%, respectively. Additionally, the SVM and XGBoost models performed well with a reduced number of features (23 and 13, respectively). The PaO2/FiO2 Ratio, C-Reactive Protein, and O2 Saturation resulted as the most important features, followed by Heartbeats, White Blood Cells, and D-Dimer, in accordance with the clinical scientific literature.
头盔式持续气道正压通气（H-CPAP）是一种无创呼吸支持方式，用于治疗急性呼吸窘迫综合征（ARDS）。ARDS是一种严重的病症，当出现严重低氧血症、胸部X光显示肺部有阴影或不明原因的呼吸衰竭等症状时可确诊。ARDS可分为轻度、中度或重度。H-CPAP疗法被推荐为轻度ARDS的初始治疗方法。尽管H-CPAP治疗中重度低氧血症患者的疗效尚不清楚，但随着新冠疫情的出现，其在这些病例中的使用有所增加。在本研究中，我们利用维梅尔卡特医院肺病科的电子病历（EMR）开发并评估了一个能够预测H-CPAP疗法对ARDS患者治疗失败情况的机器学习（ML）系统。
维梅尔卡特医院的EMR提供了接受H-CPAP治疗并被诊断为ARDS患者的所有住院病例的人口统计学信息、血液检查结果和生命体征参数。这些数据被用于创建一个包含622条记录和38个特征的数据集，训练集和测试集的比例为70%-30%。不同的ML模型，如支持向量机（SVM）、极端梯度提升（XGBoost）、神经网络、随机森林和逻辑回归，以交叉验证的方式进行迭代训练。我们还应用了一种特征选择算法来提高预测质量并减少特征数量。
支持向量机和神经网络模型被证明是最有效的，最终准确率分别达到95.19%和94.65%。在F1分数方面，这两个模型分别得分88.61%和87.18%。此外，支持向量机和极端梯度提升模型在减少特征数量（分别为23个和13个）的情况下仍表现良好。根据临床科学文献，动脉血氧分压/吸入氧分数值（PaO2/FiO2）、C反应蛋白和血氧饱和度是最重要的特征，其次是心率、白细胞和D-二聚体。
