Federated Learning for Diabetes Prediction: Investigating Client-Level Statistical Heterogeneity
编号:217
访问权限:仅限参会人
更新:2026-11-19 02:26:41
浏览:46次
In-person
摘要
Machine learning can flag diabetes risk early, but there are privacy issues related to centralized training, which depends on pooling patient records by hospitals. Federated learning trains the global model without transferring records to a central server, although how class imbalance, difference in client datasets, and the number of clients influence this technique are not known. This paper compares centralized and federated logistic regression on 96,146 records of a publicly available diabetes dataset. The problem of class imbalance is handled by applying SMOTENC locally for each client, training with FedProx, and aggregation based on the sample size. One threshold of 0.8752, which maximizes the F1-score for out-of- fold predictions, is used for all models. The clients are split into IID or non-IID (Dirichlet, alpha of 0.30) partitions for 3, 5, and 10 clients across ten seeds. All federated IID models perform equally with the centralized baseline with F1-score close to 72.45 percent in all experiments. In case of non-IID splits, ROC-AUC is 0.955-0.959 against 0.960, while the mean F1-score is 68.70 percent with 3 clients and recall is 35 to 39 percent in two partitions. Calibration analysis attributes this to clients having insufficient diabetic records to sample from, which leads to under-prediction by the global model; across 30 non-IID runs, the effective positive share correlates with recall at 0.95 but not with ROC-AUC. FedProx coefficients of 0.05 to 0.1 decreased the variability relative to FedAvg.
关键词
Federated Learning,Diabetes Prediction,logistic regression,class imbalance,SMOTENC,FedProx,Client heterogeneity
稿件作者
Aswana Lal
Rajalakshmi Engineering College
Prathyush R
Rajalakshmi Engineering College
Sanjana R
Rajalakshmi Engineering College
Srevarsha N
Rajalakshmi Engneering College
Naveed Ahmed Basha
Rajalakshmi Engineering College
发表评论