Point-of-care screening for heart failure with preserved ejection fraction using simple pulse oximetry

Background

Heart failure with preserved ejection fraction (HFpEF) prevalence is rising and is expected to become the predominant heart failure (HF) subtype, yet it remains under-recognized and under-treated. We developed and cross-validated a machine learning (ML) system that detects HFpEF by analyzing photoplethysmographic (PPG) signals recorded with a simple pulse oximeter.

Purpose

To develop and validate a widely accessible, reliable, and cost-effective method for HFpEF screening.

Methods

Patients attending routine ambulatory check-ups were included in the study. HF diagnosis was based on the latest ESC guidelines. Left ventricular ejection fraction was determined by standard transthoracic echocardiography. The PPG signal was recorded from 60 to 120 s using a simple finger pulse oximeter while calmly standing. The PPG recording was uploaded through a dedicated application to a secured cloud. Signal quality was evaluated prior to processing, including noise removal, filtering, artifact detection, and pulse wave identification.

A total of 82 features were extracted from the PPG signal, capturing temporal, morphological, and variability-related characteristics. Based on a combination of expert-guided selection and data-driven variable importance analysis, a subset of 13 PPG features was selected and utilized as input for ML models to facilitate HFpEF classification. Three different ML processing pipelines were developed, evaluated using 5-fold cross-validation and compared. We additionally evaluated models combining PPG-derived features with basic clinical variables, including age, sex, body mass index, atrial fibrillation and diabetes status, to assess whether integrating routinely available clinical information provides incremental discriminatory value beyond PPG-derived features alone.

Results

In total, 287 patients were included in the study, including 133 HFpEF patients (56.4% female, average age 70.70 ± 11.62) and 154 non-HF patients (48.1% female, average age 55.61 ± 16.77). The logistic regression classifier trained on PPG-derived features alone achieved an average cross-validated ROC AUC of 0.87 ± 0.05, corresponding to a sensitivity of 0.86 ± 0.09 and a specificity of 0.83 ± 0.09 at the Youden-optimal threshold. Additionally, when the same model was trained using a combination of PPG-derived features and the clinical variables aforementioned, performance further improved to a ROC AUC of 0.93 ± 0.03, with sensitivity of 0.87 ± 0.05 and specificity of 0.91 ± 0.07.

Conclusions

Our results suggest that HFpEF can be effectively screened using a simple pulse oximeter recording, enhanced with ML signal analysis. This approach could be adapted into a non-invasive, reliable, and cost-effective fast-track diagnostic tool for use in primary care. Early detection of HFpEF and timely referral for prompt evaluation and treatment initiation could positively impact patient outcomes.ROC and precision-recall curvesFor image description, please refer to the figure legend and surrounding text.