Artificial intelligence using ECG criteria could assist diagnosis of HFpEF
Click Here to Manage Email Alerts
Key takeaways:
- Five-year all-cause death was higher in patients with undiagnosed HFpEF vs. those who received a clinician-assigned diagnosis.
- AI may be able to identify undiagnosed patients who meet criteria for HFpEF.
Individuals with undiagnosed HF with preserved ejection fraction represent an at-risk group whose diagnosis could be improved by artificial intelligence trained with electronic health record data, according to researchers in the U.K.
“Given that HFpEF prediction scores and diagnostic criteria are relatively recent developments, it is probable that there are many patients with a prior diagnosis of HF where the formal diagnosis of HFpEF has not been made,” Jack Wu, research platform engineer at the School of Cardiovascular and Metabolic Medicine & Sciences, British Heart Foundation Centre of Research Excellence, King’s College London, and colleagues wrote. “We therefore aimed to use artificial intelligence methods — specifically natural language processing (NLP) — to identify all patients with a clinician-assigned diagnosis of HFpEF at a large London hospital and compare this cohort to a population of HF patients at the same center who meet the European Society of Cardiology criteria but who remain without a formal diagnosis of HFpEF.”
Image: Adobe Stock
For this retrospective, single-center study, Wu and colleagues established a single anonymized database of adults with a clinical diagnosis of HF according to EHRs at King’s College Hospital National Health Service (NHS) Foundation Trust from 2010 to 2022.
Patients with a clinical diagnosis of HF and left ventricular EF of 50% or more were categorized into one of four groups:
- patients with a clinician-assigned diagnosis of HFpEF;
- patients with HF, LVEF of 50% or more and clinical evidence of diastolic dysfunction meeting the ESC criteria but without a HFpEF diagnosis;
- patients with a clinical diagnosis of HF and LVEF of 50% or more who did not meet ESC criteria and did not receive a HFpEF diagnosis; and
- patients with severe valvular heart disease, hypertrophic cardiomyopathy, restrictive cardiomyopathy, constrictive pericarditis and cardiac amyloidosis.
Demographic and clinical data were retrieved using open-source NLP informatics (CogStack ecosystem; MedCAT; and MedCATTrainer) at King’s College Hospital NHS Foundation Trust.
The study was published in the European Journal of Heart Failure.
The researchers identified 8,606 patients with HF.
Among 3,727 consecutive patients with HF and LVEF of 50% or more, 8.3% had a clinician-assigned diagnosis of HFpEF, despite AI analysis demonstrating that as many as 75.4% met ESC diagnostic criteria for HFpEF but did not receive a formal diagnosis.
Manual validation of 100 randomly selected patients who met ECG criteria per AI analysis confirmed true positive in 100%.
Moreover, undiagnosed patients with HFpEF meeting ESC criteria had a higher 5-year mortality compared with those with a clinician-assigned diagnosis of HFpEF (P = .005) or those with HFpEF who did not meet ESC criteria (P = .002), according to the study. However, HF hospitalization was more common at 5 years in those with a clinician-assigned diagnosis of HFpEF compared with undiagnosed patients with HFpEF (P = .0255) or those with HFpEF who did not meet ESC criteria (P < .001).
These findings were validated in an external cohort of 1,765 patients with HF and LVEF greater than 50% at Royal Brompton Hospital.
“It is likely that the increased mortality in the Undiagnosed HFpEF (ESC Criteria) group is related to a combination of factors. Despite being younger and less comorbid, this group were less likely to be seen by a cardiologist and there was lower utilization of HF medications, both of which are important in the prognosis of HF,” the researchers wrote. “It is unclear why patients with clinical features meeting the ESC Criteria for HFpEF remain without a formal diagnosis. ... User-friendly machine learning techniques therefore have potential to complement existing diagnostic algorithms via automated identification of suspected/likely HFpEF who can then undergo formal diagnostic assessment by an expert clinician.”
Collapse
Click Here to Manage Email Alerts