Federated Learning with Interpretable Deep Models for Diabetes Prediction in Non-IID Settings Using the Flower Framework

In the era of privacy-preserving Machine Learning (ML), Federated Learning (FL) presents a transformative example for collaborative model training across distributed data sources without exposing sensitive information. This paper investigates the application of FL in healthcare using the Pima Indians Diabetes dataset, with a strong emphasis on non-Independent and Identically Distributed (non-IID) data partitioning, local client updates, and model interpretability. Three fully connected layers in a neural network and ReLU activations, implemented in PyTorch, are trained across five simulated clients using the Flower (FLWR) framework. The dataset is standardized, and clients receive shards of label-sorted data to replicate real-world heterogeneity across healthcare providers. Each Client trains its model using the Adam optimizer and cross-entropy loss, with local training loss monitored over multiple epochs. Post-training, interpretability techniques-LIME (Local Interpretable Model-agnostic Explanations)- were employed to explain distinct predictions and global feature influence. Experimental results demonstrate that while federated learning can achieve reasonable performance in non-IID settings, interpretability insights vary significantly across clients due to data distribution disparities. The findings highlight the need for client-aware personalization and future enhancements in federated optimization strategies, communication efficiency, and explainable AI in sensitive domains like healthcare.