Electrocardiogram-based AI algorithm identifies pulmonary hypertension before diagnosis
Click Here to Manage Email Alerts
Key takeaways:
- At diagnosis, the area under the receiver operating characteristic curve in two test sets was 0.92 and 0.88.
- This value stayed above 0.7 when using electrocardiograms taken 36 to 60 months before diagnosis.
An AI algorithm had high discriminative ability in identifying pulmonary hypertension-likely patients from control patients at and before diagnosis using electrocardiogram data, according to study results.
These findings were published in European Respiratory Journal.
“PH is commonly associated with significant delays in diagnosis,” Hilary M. DuBrock, MD, director of the Mayo Clinic pulmonary hypertension fellowship program in Minnesota, told Healio. “Our findings suggest that an AI early detection algorithm applied to standard 12-lead electrocardiograms has potential to accelerate the diagnosis and treatment of PH and ultimately improve patient outcomes.”
In an effort to see if an AI algorithm could detect PH early from standard 12-lead electrocardiogram (ECG) data, DuBrock and colleagues trained, validated and tested the PH Early Detection Algorithm (PH-EDA) using 39,823 PH-likely patients (mean age, 70.1 years; 51.9% men; 87.3% white) and 219,404 PH-unlikely patients/controls (mean age, 61.6 years; 47.7% men; 86.9% white) from Mayo Clinic with ECGs taken 30 days before or after PH diagnosis or taken on or before last screening.
Nearly half of the total cohort (48%) was used for training, which was done using ECGs taken within 1 month of a diagnosis of PH. The remaining patients either became a part of the validation set (12%) or the test set (40%).
The final test set included 16,175 PH-likely patients and 87,998 controls from Mayo Clinic, as well as 6,045 PH-likely patients (mean age, 64.2 years; 51.7% men; 78.8% white) and 24,256 controls (mean age, 58.3 years; 46.5% men; 82.8% white) from Vanderbilt University Medical Center (VUMC).
The algorithm was able to discriminate between PH-likely and control patients to “a high degree” as demonstrated by elevated area under the receiver operating characteristic curve (AUC) in both the Mayo Clinic (0.92) and VUMC (0.88) test sets.
Further, researchers observed favorable sensitivity and specificity in the Mayo Clinic set (85.5% and 83.8%) and the VUMC set (82.3% and 77%), as well as high negative predictive values (Mayo Clinic, 0.97; VUMC, 0.95).
When tested using ECGs taken 6 to 18 months before a PH diagnosis, the AUCs found above only slightly decreased in the Mayo Clinic (0.86) and VUMC (0.81) sets.
In terms of ECGs taken 18 to 36 months before a PH diagnosis, discrimination again slightly declined (Mayo Clinic, 0.82; VUMC, 0.78), and this was also the case when using ECGs taken 36 to 60 months before diagnosis, but the AUCs did not drop below 0.7 (Mayo Clinic, 0.79; VUMC, 0.73).
The subsets of patients with dyspnea in the Mayo Clinic set (n = 7,155 PH-likely; n = 16,155 controls) and the VUMC set (n = 3,719 PH-likely; n = 9,208 controls) had AUCs comparable to those above at each time point.
Researchers also analyzed physician ability to identify ECG abnormalities against the algorithm and found that less than one-third of the abnormalities in the Mayo Clinic (29.1%) and VUMC (30.2%) sets of patients with PH had been confirmed by physicians.
“Future studies will focus on multicenter validation of the algorithm in more diverse cohorts and different types of PH,” DuBrock told Healio.