Real World Analytics & AI, IQVIA, Cambridge, United Kingdom.
Strategic Analytics & Insights, IQVIA, Saint-Prex, Switzerland.
PLoS One. 2022 Jul 28;17(7):e0271227. doi: 10.1371/journal.pone.0271227. eCollection 2022.
Identifying COVID-19 patients that are most likely to progress to a severe infection is crucial for optimizing care management and increasing the likelihood of survival. This study presents a machine learning model that predicts severe cases of COVID-19, defined as the presence of Acute Respiratory Distress Syndrome (ARDS) and highlights the different risk factors that play a significant role in disease progression.
A cohort composed of 289,351 patients diagnosed with COVID-19 in April 2020 was created using US administrative claims data from Oct 2015 to Jul 2020. For each patient, information about 817 diagnoses, were collected from the medical history ahead of COVID-19 infection. The primary outcome of the study was the presence of ARDS in the 4 months following COVID-19 infection. The study cohort was randomly split into training set used for model development, test set for model evaluation and validation set for real-world performance estimation.
We analyzed three machine learning classifiers to predict the presence of ARDS. Among the algorithms considered, a Gradient Boosting Decision Tree had the highest performance with an AUC of 0.695 (95% CI, 0.679-0.709) and an AUPRC of 0.0730 (95% CI, 0.0676 - 0.0823), showing a 40% performance increase in performance against a baseline classifier. A panel of five clinicians was also used to compare the predictive ability of the model to that of clinical experts. The comparison indicated that our model is on par or outperforms predictions made by the clinicians, both in terms of precision and recall.
This study presents a machine learning model that uses patient claims history to predict ARDS. The risk factors used by the model to perform its predictions have been extensively linked to the severity of the COVID-19 in the specialized literature. The most contributing diagnosis can be easily retrieved in the patient clinical history and can be used for an early screening of infected patients. Overall, the proposed model could be a promising tool to deploy in a healthcare setting to facilitate and optimize the care of COVID-19 patients.
识别最有可能发展为重症感染的 COVID-19 患者对于优化护理管理和提高生存率至关重要。本研究提出了一种机器学习模型，用于预测 COVID-19 重症病例，定义为急性呼吸窘迫综合征（ARDS）的存在，并强调了在疾病进展中起重要作用的不同危险因素。
使用美国 2015 年 10 月至 2020 年 7 月的行政索赔数据，创建了一个由 289351 例 2020 年 4 月确诊的 COVID-19 患者组成的队列。对于每位患者，从 COVID-19 感染前的病史中收集了 817 项诊断信息。本研究的主要结局是 COVID-19 感染后 4 个月内出现 ARDS。研究队列被随机分为训练集用于模型开发、测试集用于模型评估和验证集用于真实世界性能评估。
我们分析了三种机器学习分类器来预测 ARDS 的存在。在所考虑的算法中，梯度提升决策树的性能最高，AUC 为 0.695（95%CI，0.679-0.709），AUPRC 为 0.0730（95%CI，0.0676-0.0823），与基线分类器相比，性能提高了 40%。还使用了五名临床医生小组来比较模型的预测能力与临床专家的预测能力。比较表明，我们的模型在精度和召回率方面与临床医生的预测相当或优于临床医生的预测。
本研究提出了一种使用患者索赔历史预测 ARDS 的机器学习模型。模型用于进行预测的危险因素在专业文献中与 COVID-19 的严重程度广泛相关。最主要的诊断可以在患者的临床病史中轻松检索，并可用于对感染患者进行早期筛查。总体而言，该模型可以作为一种有前途的工具，在医疗保健环境中部署，以促进和优化 COVID-19 患者的护理。
