Abstract: SA-PO013
Identifying Peripheral Artery Disease in Persons with and without CKD from Electronic Health Records
Session Information
- Augmented Intelligence, Large Language Models, and Digital Health
October 26, 2024 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Parsons, Georgia, Stanford University School of Medicine, Stanford, California, United States
- Parvathinathan, Gomathy, Stanford University School of Medicine, Stanford, California, United States
- Stedman, Margaret R., Stanford University School of Medicine, Stanford, California, United States
- Chang, Tara I., Stanford University School of Medicine, Stanford, California, United States
Background
Peripheral artery disease (PAD) may be more difficult to identify in CKD due to altered pathophysiology. We aimed to develop a model to identify PAD in persons with and without CKD from electronic health records (EHR) to facilitate future studies.
Methods
We used the Stanford Medicine Research Repository (STARR) EHR mapped to the OMOP Common Data Model to identify adults with ≥1 PAD encounter 1/2020 to 9/2022. We randomly selected 993 patients for chart review for ground truth PAD status. Potential variables included PAD billing codes, demographics, comorbidities, diagnostic testing, and specialist visits. We built logistic regression models using age, diabetes and the top 10 percentile of variables selected by random forest. We assessed sensitivity, specificity, positive predicted value (PPV), negative predicted value (NPV), and accuracy overall and by CKD status (defined using ≥2 codes/labs). We chose the threshold probability to maximize Youden’s index with specificity>sensitivity. We calculated 95% confidence intervals [CI] by bootstrapping 1,000 samples.
Results
In our cohort (Table), 222 (22%) had PAD (130 with CKD, 92 without). Our model had high sensitivity, specificity, NPV and accuracy but low PPV (Figure). Results were similar by CKD status.
Conclusion
Our study is among the first to focus on PAD ascertainment in CKD. Using discrete EHR data elements, our model generally performed well regardless of CKD status, but due to the lower prevalence of PAD, the PPV was relatively low. Future approaches will use natural language processing to incorporate clinical notes and diagnostic reports to improve model performance.
Baseline Characteristics
Model Performance
Funding
- Other NIH Support