Abstract: FR-PO0002
Identification of a Strategic "Intervention Window" in Rapid Progression of Diabetic Kidney Disease: A Temporal Cascade Analysis Providing up to 12 Months of Lead Time
Session Information
- Artificial Intelligence: Clinical Prediction, Decision Support, and Digital Kidney Care
October 23, 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
Authors
- Imakiire, Yukihiro, Chiba Daigaku, Chiba, Chiba Prefecture, Japan
- Hananoe, Ayaka, Rikagaku Kenkyujo, Wako, Saitama Prefecture, Japan
- Zhao, Xuyang, Osaka Daigaku Daigakuin Igakukei Kenkyuka Igakubu, Suita, Osaka Prefecture, Japan
- Kawakami, Eiryo, Chiba Daigaku, Chiba, Chiba Prefecture, Japan
Background
Diabetic Kidney Disease (DKD) requires strategic intervention, yet clinical markers often fail to capture the nonlinear "tipping point" of rapid eGFR decline. While conventional models excel in static risk assessment, they lack sufficient lead-time for proactive care. We leveraged a Transformer-based temporal model to identify a 12-month "Intervention Window" using a 20-year longitudinal dataset, aiming to intervene prior to irreversible damage.
Methods
Using Electronic Medical Record (EMR) data from 11,551 patients (20-yr follow-up), a Transformer-based model was trained to predict 'Rapid Decliners' (eGFR slope < -5 mL/min/1.73m2/yr). After preprocessing for noise/missing values, SHapley Additive exPlanations (SHAP) was employed to analyze feature importance across Early (-12m), Mid (-6m), and Late (-2 to 0m) phases to map the temporal cascade of clinical drivers.
Results
Our model achieved Area Under the Curve (AUC) of 0.993 at 12-month lead-time. Performance remained robust across all windows. Temporal Cascade analysis showed phased transitions: Early Phase (-12m) driven by baseline eGFR and metabolic markers; Mid Phase (-6m) by eGFR slope; and Late Phase (-2m) by acute blood pressure (BP) and proteinuria fluctuations before the rapid decline event.
Conclusion
Transformer-based analysis provides up to 12 months of lead-time for DKD intervention, shifting the paradigm from reactive management to proactive stabilization during the identified strategic window.
Figure 1: Trajectories of eGFR and predictive scores identifying the strategic lead-time.
Figure 2: Temporal SHAP analysis revealing the phased transition of predictive factors.