A photoplethysmography-based algorithm for early atrial fibrillation detection in heart failure telemonitoring

Background

Atrial fibrillation (AF) and heart failure (HF) frequently coexist, share overlapping pathophysiologic mechanisms and a bidirectional causal relationship. Their concurrence is associated with worse clinical outcomes. Early detection of AF in HF patients may offer timely interventions and improve outcomes.

Purpose

To develop and validate an AF detection algorithm based on machine-learning analysis of photoplethysmography (PPG) signals, integrated into a CE-certified (MDR IIb) HF telemonitoring platform.

Methods

We included 247 HF patients (43627 individual PPG recordings, mean age 65.3 ± 11.8 years; 33.6% female, predominantly NYHA class: II 54.1% and III 26.0%, 49% non-ischemic etiology, HFpEF 32.8%, HFmrEF 16.2%, HFrEF 47.0%) from an ongoing telemonitoring study, STOP-DHF (Strategy TO Prevent Decompensated HF).

Patients were classified into two groups based on ECG and clinical data: (1) permanent AF , defined as history of permanent AF together with documented AF on a 12-lead (n = 47; 7423 PPG recordings); and (2) non-AF , defined as sinus rhythm on the index 12-lead ECG with no prior documented AF episodes (n = 200; 35,844 PPG recordings). External validation was performed on 936 PPG recordings from 167 patients, combining data from the MIMIC database (n=35; 683 PPGs) and cardiology outpatients (n=132; 253 PPGs). A second validation used a proprietary dataset (n=21; 51 PPG–ECG pairs) with frequent irregular extrasystoles. A third patient-level validation included proprietary PPG–ECG pairs from a patient with frequent sinus–AF transitions. Fifteen waveform features were extracted, including heart rate statistics, time- and frequency-domain heart rate variability measures, count-based variability indices, and autocorrelation descriptors reflecting rhythm irregularity. Three algorithms (logistic regression, random forest, and LightGBM) were trained and compared using stratified five-fold cross-validation.

Results

The LightGBM classifier demonstrated the best overall performance. In 5-fold internal cross-validation, the model achieved a mean ROC AUC = 0.98 ± 0.01, mean average precision 0.96 ± 0.04, sensitivity 0.96 ± 0.02 and specificity 0.97 ± 0.02 at the Youden threshold. In external validation, it reached AUC = 0.99 (95% CI: 0.98–1.00) and average precision of 0.98, with sensitivity 0.97 (CI: 0.94-0.98) and specificity 0.93 (CI: 0.91-0.95) at the Youden threshold. In the extrasystolic validation, specificity remained high (0.94; 95% CI: 0.84–0.98) with a low false-positive rate 0.06, despite frequent irregular ectopy. A third patient-level validation confirmed accurate discrimination during frequent sinus–AF transitions (figure 2).

Conclusions

Our PPG-based AF detection algorithm demonstrated high and consistent performance across internal, external, and patient-level validations. Its integration into HF telemonitoring platform could enable early AF recognition.Figure 1Figure 2