DeepAsthmaNet: A Time-Aware Federated Prognostic Framework for Personalised Paediatric Asthma Risk Stratification in Primary Care

Abstract

Asthma in children is still a common chronic illness with a wide range of clinical manifestations and progressions. Timely action and individualised care depend on early and precise risk assessment. This study presents a federated architecture used to train DeepAsthmaNet integrates deep neural architectures and attention mechanisms to enable early, accurate prediction of childhood asthma by combining clinical, genetic and environmental data. After rigorous preprocessing, feature engineering and redundancy elimination, the pipeline leverages inputs from genetic biomarkers, wearables, environmental sensors, and electronic health records. At an optimal learning rate of 0.80, it achieves superior performance with an MSE of 0.00015, an MAE of 0.0051, and an AUC of 0.9773, outperforming baseline models (ANN, DNN, RNN, LSTM). To handle heterogeneous distributed data, three federated aggregation strategies, FedAvg, FedProx and Krum, were tested, with FedAvg delivering the best balance of accuracy, convergence stability and communication efficiency.