Abstract: FR-PO0005
Artificial-Intelligence-Based Trajectory Clustering Analysis Identifies High-Risk and Previously Overlooked Patients with CKD
Session Information
- Artificial Intelligence and Digital Health at the Bedside
November 07, 2025 | Location: Exhibit Hall, 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
- Kanda, Eiichiro, Kawasaki Ika Daigaku, Kurashiki, Okayama Prefecture, Japan
- Epureanu, Bogdan I., University of Michigan, Ann Arbor, Michigan, United States
- Adachi, Taiji, Kyoto Daigaku, Kyoto, Kyoto Prefecture, Japan
- Sasaki, Tamaki, Kawasaki Ika Daigaku, Kurashiki, Okayama Prefecture, Japan
- Pennathur, Subramaniam, University of Michigan, Ann Arbor, Michigan, United States
- Okada, Hirokazu, Saitama Ika Daigaku, Iruma, Saitama Prefecture, Japan
- Nangaku, Masaomi, Tokyo Daigaku, Bunkyo, Tokyo, Japan
- Kashihara, Naoki, Kawasaki Ika Daigaku, Kurashiki, Okayama Prefecture, Japan
Background
Many chronic kidney disease (CKD) patients have not been properly diagnosed, and their characteristics remain unclear. It has been difficult to identify their clinical courses by conventional analysis. Therefore, our study aimed to identify overlooked CKD patient groups at high risk of dialysis or death, addressing unmet clinical needs.
Methods
We employed data mining techniques using a newly developed artificial intelligence (AI) algorithm—time-series K-means analysis—to classify patients on the basis of 34 time-series variables, including background, laboratory test results, and medication from a CKD cohort study (n=3,129). This method utilizes dynamic time warping to assess the similarity between clinical data trajectories. Subsequently, we conducted a barycenter transition analysis of all time-series data to visually examine the trajectories.
Results
Their mean age and eGFR were 62.0 years and 50.6 mL/min, respectively. They were distinctly classified into seven groups on the basis of human-understandable backgrounds (Figure 1). Notably, one group exhibited low urinary protein levels, no hypertension, and no identifiable cause of CKD. They were assumed to have a favorable prognosis and were not specifically treated. However, their outcomes were worse than those of appropriately treated hypertensive patients, who had the best prognosis (hazard ratio, 2.82; 95% confidence interval, 1.58–5.09; p<0.0001). Although each group initially displayed different clinical courses, their progression patterns gradually aligned over time, ultimately converging into the final common pathway (Figure 2).
Conclusion
The AI-based trajectory clustering analysis successfully uncovered the distinctive clinical course of previously overlooked CKD patients with unknown etiologies. The AI algorithm is valuable in identifying unmet clinical needs within large-scale patient datasets.