Abstract: FR-PO0032
Harnessing Artificial Intelligence (AI) and Clinical Guidelines to Identify At-Risk Patients with CKD and Optimize Management of 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
- Desai, Pooja N., AstraZeneca, Wilmington, Delaware, United States
- Veldman, Anouk, AstraZeneca, The Hague, Netherlands
- Malvolti, Elmas, AstraZeneca, Cambridge, United Kingdom
- Evans, Sean, AstraZeneca, Wilmington, Delaware, United States
- Tanwar, Ashwani, Pangaea Data, London, United Kingdom
- Tao, Guanyu, Pangaea Data, London, United Kingdom
- Chowdhury, Marzana, Pangaea Data, London, United Kingdom
- Everman, David, Pangaea Data, South San Francisco, California, United States
- Zhang, Jingqing, Pangaea Data, London, United Kingdom
- Gupta, Vibhor, Pangaea Data, South San Francisco, California, United States
Background
Currently, an estimated 843.6 million people worldwide live with chronic kidney disease (CKD), yet 90% remain unidentified and under-treated, missing opportunities for early intervention (bit.ly/4kfbxbF, bit.ly/3H3aWv0). This leads to poorer patient outcomes and increased strain on healthcare systems. Moreover, manual review and documentation occupy around 37% of a clinician’s workday and are both time-intensive and error-prone, preventing clinicians from focusing on patient care (bit.ly/4dv6u4v). By leveraging guideline-based AI on electronic health records, healthcare professionals can efficiently find, review and manage those at risk for CKD and the under-treated CKD patients.
Methods
This analysis evaluated whether an AI platform, configured according to KDIGO guidelines, could accurately find confirmed CKD patients and individuals at-risk of CKD. The AI leveraged techniques including retrieval (e.g. Natural Language Processing) and reasoning (e.g. symbolic reasoning) to find CKD and at-risk CKD patients. Once configured, the AI was applied to a retrospective US dataset consisting of ICU patient records which included structured data (e.g. serum creatinine, UACR) and unstructured data (e.g. clinical notes).
To evaluate the accuracy of AI, the patient records of 341 patients were analyzed by the AI and manually reviewed by clinicians. Two clinicians independently assessed which of the three cohorts (i.e. confirmed diagnosis of CKD, at-risk of CKD, and no signs of CKD) each patient belongs to, following KDIGO. Any conflicts were resolved by a third clinician.
Results
It was demonstrated that AI achieved a 97% sensitivity and 85% precision in finding confirmed CKD patients, and AI achieved 87% sensitivity and 87% precision in finding patients at risk of CKD.
Conclusion
This analysis demonstrates the application potential of AI that is configured on clinical guidelines to accurately find CKD patients and people at-risk for CKD. By proactively and reliably identifying these patients at the point of care, patients can receive earlier medical assessment and prompt intervention, thereby reducing the clinical and wider health economic burden of CKD.
Funding
- Commercial Support – AstraZeneca