Abstract: SA-PO0082
Electronic Health Record (EHR)-Based Phenotyping for Kidney Failure: A Composite Algorithm Outperforms Published Rules
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
Authors
- Shang, Ning, Columbia University, New York, New York, United States
- Khan, Atlas, Columbia University, New York, New York, United States
- Kiryluk, Krzysztof, Columbia University, New York, New York, United States
Background
Accurate identification of kidney failure (KF) from EHRs is critical for epidemiologic and outcomes research. Existing KF phenotyping rules often rely on dialysis or transplant codes and may underestimate KF prevalence. We evaluated literature-based KF phenotyping rules and developed a composite multi-domain algorithm.
Methods
We conducted a scoping review to identify EHR-based KF phenotyping algorithms. Among 105 screened studies, diagnostic (cross-sectional) and predictive (time-to-event), 45 EHR-applicable studies were selected after excluding non-EHR-based definitions.
Using literature review and clinician/informatician input, we implemented 8 literature-derived KF phenotyping rules using diagnosis, laboratory, and procedure data. Transient dialysis events were excluded using temporal filters to reduce AKI-related false positives. A composite algorithm was defined as positivity for any individual rule.
Algorithms were evaluated in an adjudicated KF cohort (2,590 KF; 5,407 non-KF) using sensitivity, specificity, PPV, and NPV.
Results
Performance varied substantially across rules (Table 1). Diagnosis-based codes showed the best single-rule performance. Procedure-based rules had high specificity but low sensitivity, lab-based rules showed intermediate tradeoffs by eGFR criteria. The composite algorithm achieved the best overall performance, with >3-fold higher sensitivity than the best published procedure-based rule.
Conclusion
Published KF rules showed markedly different performance in real-world data. A composite multi-domain strategy achieved the best sensitivity-specificity balance for KF research. External validation and NLP-based approaches are important next steps.
Table 1
| Rule | Sen (%) | Spe (%) | PPV (%) | NPV (%) |
| Diagnosis-based | ||||
| ≥1 CKD stage 5 or ESRD diagnosis code | 44.2 | 94.6 | 79.6 | 78.0 |
| Laboratory-based | ||||
| ≥1 eGFR <15 | 21.9 | 97.2 | 78.8 | 72.2 |
| Sustained low eGFR <15 with prior eGFR<90 or kidney disease diagnosis 3 months earlier) | 12.8 | 97.9 | 74.6 | 70.1 |
| ≥40% eGFR decline from baseline | 8.1 | 98.1 | 66.9 | 69.0 |
| Procedure-based | ||||
| ≥1 dialysis diagnosis or procedure code | 5.3 | 99.7 | 88.4 | 68.7 |
| ≥1 transplant diagnosis or procedure code* | 17.0 | 98.3 | 82.7 | 71.2 |
| CKD stage 5/ESRD diagnosis codes with dialysis codes within 7 days* | 3.1 | 99.8 | 85.9 | 68.2 |
| ≥3 dialysis codes spanning >30 days or ≥2 dialysis codes within 3 months* | 0.9 | 99.8 | 72.7 | 67.8 |
| Composite | ||||
| Any rule positive | 53.9 | 91.9 | 76.1 | 80.6 |
*Best-performing published rules
Funding
- Other NIH Support