Patient-Centric Digital Twin Framework with Hybrid Knowledge Distillation for Federated Class-Incremental Learning in...

Area of Science:

• Computational biology and bioinformatics
• Artificial intelligence in healthcare
• Digital twin technology

Background:

• Precision medicine necessitates adaptive, patient-centric digital twins for collaborative healthcare across institutions.
• Federated learning in medicine faces challenges from heterogeneous patient data (spatial divergence) and evolving disease categories (temporal dynamics).

Purpose of the Study:

• To propose a novel digital twin-enabled precision medicine framework using hybrid knowledge distillation.
• To address dual heterogeneity challenges in federated learning for medical applications.

Main Methods:

• Developed a hybrid knowledge distillation framework integrating adaptive patient classification loss, clinical semantic distillation loss, and biomarker attention distillation loss.
• Implemented an adaptive weighting mechanism using gradient magnitudes to dynamically adjust category preservation strength during training.
• Simultaneously preserved soft-label disease relationships, intermediate convolutional features, and gradient-based adaptive weighting.

Main Results:

• The framework demonstrated consistent superiority over baseline methods across various benchmarks.
• Achieved 71.05% accuracy on CIFAR100 (1.87% improvement), and accuracy gains of 2.18%, 1.92%, and 1.95% on OrganAMNIST, OrganCMNIST, and OrganSMNIST medical imaging datasets, respectively.
• Validated with clinical laboratory data, showing a 2.54% accuracy improvement.

Conclusions:

• The proposed hybrid knowledge distillation framework is effective for digital twin-enabled precision medicine.
• It successfully addresses heterogeneity challenges, enabling continuous diagnostic knowledge expansion without compromising patient privacy.
• Establishes practical viability for real-world precision medicine deployment.