Abstract: SA-PO0093
MedKD: A Knowledge Distillation-Based Lightweight Deep Learning Model for Drug Recommendation in Diabetic Kidney Disease
Session Information
- Artificial Intelligence: Kidney Imaging, Pathology, and Precision Phenotyping
October 24, 2026 | Location: Exhibit Hall A, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Author
- Tang, Wenbin, Xiangya Hospital Central South University, Changsha, Hunan, China
Background
Diabetic kidney disease (DKD) in type 2 diabetes is a leading cause of end-stage renal disease, and its pharmacological management requires coordinated use of antidiabetic, antihypertensive, lipid-lowering, and renoprotective agents. In busy clinical settings, choosing appropriate drug classes and specific agents from heterogeneous routine data is time-consuming and variable.
Methods
We developed MedKD, a lightweight student network for four-class drug recommendation in type 2 DKD that distils knowledge from a larger TabTransformer teacher using combined supervised and distillation losses. Using 22,511 hospitalization episodes (14,968 patients), we trained and evaluated MedKD and five deep learning baselines on standardized demographic, comorbidity, staging, and laboratory features with stratified five-fold cross-validation, and in a secondary analysis trained separate random forest models within the glucose-lowering, blood pressure-lowering, and lipid-lowering domains to predict specific drug names among commonly used agents.
Results
MedKD achieved an accuracy of 0.8641, with precision, recall, and F1-score of 0.8645, 0.8641, and 0.8629, respectively, while using 7.2×10^5 parameters and outperforming all baselines. Within-class random forest models showed modest top-1 but higher top-3 accuracies for drug name prediction (glucose-lowering:0.23/0.42; blood pressure-lowering: 0.12/0.32; lipid-lowering: 0.43/0.85), preferentially suggesting clinically dominant agents such as metformin, SGLT2 inhibitors, calcium channel blockers, and statins.
Conclusion
These findings indicate that a distillation-based lightweight architecture can provide accurate, resource-efficient support for class-level drug recommendation in T2DKD. Extending this framework with within-class models offers a pragmatic path towards agent-level suggestions, which could be refined using treatment combinations and longitudinal outcome data.
Funding
- Government Support – Non-U.S.