A Dual Genetic Algorithm for Hyperparameter Tuning and Model Aggregation in Federated Learning for Depression Prediction in Multimorbid Chronic Disease Populations
摘要
In healthcare, especially when predicting depression in patients with multimorbid chronic diseases, protecting patient data privacy and ensuring model effectiveness are essential and challenging. Centralized machine learning methods are risky and expose sensitive information; therefore, solutions are needed to address this problem. In this work, we propose a Federated Learning (FL)-based architecture for healthcare that promotes efficient model training while guaranteeing the privacy of patients’ medical information. The FL process allows multiple FL nodes (e.g., hospitals) to collaboratively build and train a model without sharing raw patient data. On the FL node side, a Light Gradient Boosting Machine (LightGBM) fine-tuning process with a Genetic Algorithm (GA) metaheuristic is introduced to boost prediction accuracy. A new evaluation metric, Weighted FN/FP trade-off Metric for Healthcare (WFTM), which combines the False Negative (FN) rate with a logarithmic transformation of the False Positive (FP) rate, is defined to assess the proposed model. To enhance the performance of the aggregated or global model, we also employ a GA to combine local models by evaluating their performance independently. This approach aims to balance a capable and accurate model while maintaining data privacy in a health-focused context. All FL nodes merged with similar parameters (