AI Case Study
Mayo Clinic researchers accurately diagnose hyperkalemia using a smartphone electrocardiogram device
Researchers tested the ability of a smartphone electrocardiogram device, AliveCor, as a non-invasive way to test for hyperkalemia. They found that the device was a viable tool able to detect even mild versions of the condition.
Industry
Healthcare
Healthcare Providers And Services
Project Overview
"An artificial intelligence (AI) algorithm utilizing a deep neural network was trained to detect hyperkalemia using only ECG leads with promising results... Each patient used an investigational version of a portable, AliveCor smartphone ECG device to acquire a 4-hour ECG recording during two separate dialysis sessions, with concurrent blood testing."
Reported Results
An AUC of 0.91 was achieved with 91% sensitivity and 72% specificity, which is considered good.
Technology
"An artificial intelligence (AI) algorithm utilizing a deep neural network was trained to detect hyperkalemia using only ECG leads I and II. Performance was evaluated on the testing group. Each patient used an investigational version of a portable, AliveCor smartphone ECG device to acquire a 4-hour ECG recording during two separate dialysis sessions, with concurrent blood testing."
Function
Strategy
Data Science
Background
"Hyperkalemia is commonly seen in congestive heart failure, chronic kidney disease, diabetes and is associated with significant mortality and arrhythmic risk. Because it is frequently asymptomatic, detecting hyperkalemia is challenging and image recognition AI can significantly increase the chances of detection".
Benefits
Risk reduction - Predictive diagnosis
Data
"Data from patients seen at Mayo Clinic between 1994 and 2017 with a serum potassium acquired within 12 hours of an ECG were included in this IRB-approved study. Study data included 709,000 patients, 2.1 million ECGs and 4.0 million serum potassium values. Patients were split randomly into training (2/3) and testing (1/3) groups... An additional, prospective, independent testing dataset was also obtained in 10 patients undergoing hemodialysis."