Background
More than 64 million patients with heart failure (HF) are burdened with poor quality of life, frequent hospitalizations, and high readmission rates. While HF with reduced ejection fraction (HFrEF) accounts for approximately half of these patients, there is currently no effective tool for performing cost-effective and reliable early-stage screening. For this purpose, we derived and cross-validated a machine learning (ML) based system capable of identifying HFrEF based on analysis of the photoplethysmographic (PPG) signal recorded by a simple pulse-oximeter.
Methods
Patients attending a routine ambulatory check-up at an outpatient clinic were included in the study. Left ventricular ejection fraction was determined by standard transthoracic echocardiography. Diagnosis of HF was based on the latest ESC recommendations for HF. PPG was measured using a standard oximeter modified to record raw PPG signal. Before processing, the quality of the measured PPG signal was assessed by signal quality indices to determine whether the quality of signal was sufficient for further analysis. In brief, the signal processing involved noise removal, filtering for specific frequencies, automatic artifact detection and identification of individual heartbeats, on which specific parts of the signal and signal characteristics, necessary for further analysis, were identified. Four different ML processing pipelines with different approaches to feature selection and class balancing were deployed, validated using the K-fold cross-validation and compared.
Results
A total of 371 patients were included in the study: 54 patients with HFrEF (mean age 66.87 ± 14.35, 74.07% male) and 317 patients with HFmrEF (4), HFpEF (42) or no history of HF (271) (mean age 54.12 ± 17.79, 51.73% male). From the 4 ML pipelines, LightGBM with SMOTE + Edited Nearest Neighbors balancing had the highest cross-validated predictive performance for detection of HFrEF (c-statistics 0.8, CI 0.69 – 0.89).
Conclusions
Our results suggest that HFrEF can be effectively screened using a simple pulse oximeter, thanks to ML signal analysis. This solution could be adapted to a simple, non-invasive, readily available, time-saving, and inexpensive point-of-care diagnostic tool for primary care. Early disease detection and referral to a cardiologist (or HF specialist) would have a positive impact on patient morbidity and mortality.Receiver Operating Characteristics CurveModel Performance Metrics