A Federated Learning Approach for Disease Prediction and Remedies Recommendation
A Sehag, Varshini Jayasankar, S. Poojitha, Shreyas Sai Raman, V Sarasvathi · 2024
Deep learning breakthroughs have transformed disease prediction and treatment recommendation systems, yet the confidentiality of sensitive medical data remains a major worry. In this paper, we propose a federated learning(FL) methodology that harnesses the capabilities of deep learning models such as Recurrent Neural Networks (RNN), Long ShortTerm Memory Networks (LSTM), Gated Recurrent Units (GRU), and Bidirectional Encoder Representations from Transformers (BERT) to address this challenge. Our approach provides a framework for multiple medical institutions to collaboratively train a deep learning model without sharing patients’ private medical records. Leveraging the Flower module, we establish a federated learning architecture comprising decentralized trainers and a global server responsible for aggregating model weights from all trainers. This setup ensures data privacy while enabling the model to learn from the diverse medical records across institutions. The trained model demonstrates significant accuracy in predicting approximately 40 common diseases from rich natural language medical reports. By integrating natural language processing capabilities, our model not only diagnoses ailments but also recommends tailored remedies. Since most of the diseases in our dataset are common and non-severe, users can access these remedies, bypassing the necessity for immediate medical visits, thus saving time and expenses associated with consultations. The proposed methodology offers a viable solution to establish a framework for collaborative learning in healthcare while safeguarding patients’ private data- a crucial aspect in today’s data driven medical landscape. Additionally, it offers a way to set up a highly accurate disease prediction and remedies recommendation system, empowering users to take proactive measures for their health. BERT+LSTM model which we developed showed an accuracy of 96.5% with centralized training whereas training the model on the FL architecture provided an accuracy of 98.75%.