■高患病率,发病率和死亡率,慢性阻塞性肺疾病(COPD)的疾病异质性导致来自不同医疗单位患者就诊的分散数据。整合分散的数据进行分析建模的巨大成本,以及对患者隐私保护的法律需求导致数据孤岛的出现。
■在保护患者隐私的前提下,整合不同医疗单位患者的零散数据进行高质量建模,有利于促进数字健康的发展。基于此,我们使用FedAvg开发了一种分布式COPD疾病诊断系统,称为COPD平均联合学习(COPD_AVG_FL).
■首先,为了构建COPD_AVG_FL,从现实世界中收集COPD患者的临床数据,并进行数据预处理以清理不正确的数据,离群值样本和缺失值。然后,经典的联邦学习体系结构被设计为COPD_AVG_FL。最后,为了评估已建立的COPD_AVG_FL系统,我们开发了集中式机器学习(CML)。
■我们的结果表明,在COPD_AVG_FL的协助下,绝对改善率为13.4%(准确度),13.3%(精度),12.8%(召回),测试数据的13.1%(F1-Score)和12.9%(AUC),分别。模型训练和原始训练数据之间的解耦保护了患者的隐私,并有助于安全地集成来自不同医疗单位的更多COPD数据,以生成更全面的模型COPD_AVG_FL。这种方法促进了COPD医学的明智信息技术在现实临床世界中的落地。我们模型的代码将在https://github.com/Cczhh/COPD_AVG_FL/tree/master上提供。
UNASSIGNED: The high prevalence, morbidity and mortality, and disease heterogeneity of chronic obstructive pulmonary disease (COPD) result in the scattered data derived from patient visits in different medical units. The huge cost of integrating the scattered data for analysis and modeling, as well as the legal demand for patient privacy protection lead to the emergence of data island.
UNASSIGNED: On the premise of protecting patient privacy, integrating scattered data of patients from different medical units for high-quality modeling is beneficial to promoting the development of digital health. Based on this, we develop a distributed COPD disease diagnosis system termed COPD average federated learning (COPD_AVG_FL) using FedAvg.
UNASSIGNED: First, to build the COPD_AVG_FL, the clinical data of COPD patients from the real world is collected and the data pre-processing is performed to clean the incorrect data, outlier samples and missing values. Then, a classical federated learning architecture is designed as COPD_AVG_FL. Finally, to evaluate the established COPD_AVG_FL system, we develop Centralized Machine Learning (CML).
UNASSIGNED: Our results suggest that, with the assistance of COPD_AVG_FL, the absolute improvement rates are 13.4% (accuracy), 13.3% (precision), 12.8% (recall), 13.1% (F1-Score) and 12.9% (AUC) on the test data, respectively. The decoupling between model training and raw training data protects the patients\' privacy, and helps to securely integrate more COPD data from different medical units to generate a more comprehensive model COPD_AVG_FL. This approach promotes the landing of wise information technology of medicine for COPD in the real clinical world. Code for our model will be made available at https://github.com/Cczhh/COPD_AVG_FL/tree/master.