Abstract: TH-PO012
Noninvasive Artificial Intelligence (AI)-Enhanced Electrocardiographic Detection of Hyperkalemia in the Emergency Department (ED) and ICU
Session Information
- AI, Digital Health, Data Science - I
November 02, 2023 | Location: Exhibit Hall, Pennsylvania 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
- Dillon, John J., Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Liu, Kan, Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Dugan, Jennifer, Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Jentzer, Jacob, Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Attia, Zachi I., Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Friedman, Paul, Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
- Harmon, David M., Mayo Foundation for Medical Education and Research, Rochester, Minnesota, United States
Group or Team Name
- Nephrology/Cardiology.
Background
We have previously trained a convolutional neural network using an all-patient cohort to produce an AI algorithm that can detect hyperkalemia from the surface ECG. In this validation study, we assessed the network's performance among ED and ICU patients.
Methods
We included adult patients presenting to the ED at all Mayo Clinic sites between February and August 2021 (ED cohort) and patients admitted to the ICU at Mayo Clinic St. Mary's Hospital, Rochester, MN between August 2017 and February 2018 (ICU cohort) if they provided research authorization, had a standard 12-lead supine ECG and had a blood K value within 4 hours of the ECG. The network analyzed leads I and II of the 12-lead ECG to calculate the probability of hyperkalemia, defined as K>6 mEq/L. The ED and ICU cohorts were analyzed separately. Exploratory subgroup analyses were performed for patients with eGFR<45 ml/min and eGFR<30 ml/min.
Results
40,128 ED patients and 2636 ICU patients were included. The prevalence of hyperkalemia was 0.9% in the ED cohort and 3.3% in the ICU cohort. The AI-ECG had AUCs of 0.88 in both cohorts with sensitivities and specificities ≥ 80%. Negative predictive values (NPVs) were >99% in both cohorts. Although positive AI-ECGs quadrupled the probability of hyperkalemia, positive predictive values (PPVs) were relatively low: 3.5% in the ED and 14% in the ICU in part due to low hyperkalemia prevalences. Low eGFR subgroups had higher hyperkalemia prevalences and higher PPVs as shown in the Table.
Conclusion
The AI-ECG demonstrated excellent discrimination with AUCs of 0.88 in both cohorts. It was highly effective at ruling out hyperkalemia with NPVs>99% in both cohorts, but with much lower PPVs, suggesting that it is most useful as a screening test to exclude hyperkalemia. One method by which PPVs can be increased is by limiting testing to high-risk populations, such as those with reduced eGFR.
| ED | ICU | |||||
| All | eGFR<45 | eGFR<30 | All | eGFR<45 | eGFR<30 | |
| N | 40128 | 5298 | 2298 | 2636 | 745 | 429 |
| AUC | 0.88 | 0.85 | 0.83 | 0.88 | 0.85 | 0.85 |
| Sensitivity | 80% | 86% | 86% | 82% | 84% | 88% |
| Specificity | 80% | 65% | 60% | 82% | 72% | 67% |
| PPV | 3.5% | 10% | 15% | 14% | 23% | 29% |
| NPV | 99.8% | 99.0% | 98.2% | 99.2% | 97.9% | 97.4% |
| K>6 Prevalence | 0.87% | 4.6% | 7.7% | 3.3% | 9.3% | 13.3% |
Funding
- Clinical Revenue Support