Medical trajectory prediction system and method based on federated learning and differential privacy

Abstract

The application discloses a medical trajectory prediction system and method based on federated learning and differential privacy, aiming at the problems of difficult balance between privacy protection and performance, and large communication overhead in cross-institutional data collaborative training. The application embeds privacy constraints, performance requirements, data and calculation budgets into neural network design, deduces the optimal matching relationship between model hyperparameters and differential privacy parameters based on the differential privacy large model scaling law, adopts a double-branch decoupling model, configures differentiated privacy budgets and noise intensities for disease prediction and trajectory prediction tasks, and introduces differentiated coefficients and comorbidity protection coefficients for medical scene optimization. The communication overhead is reduced, and data heterogeneity is coped with through parameter increment uploading and three-dimensional weighted aggregation mechanism, and a closed-loop verification is formed in combination with a dynamic verification module. The application realizes high-precision, strong-privacy and low-overhead medical trajectory prediction, and is suitable for clinical auxiliary diagnosis and disease development trend prediction in a multi-institutional collaborative environment.

Images