A novel machine learning procedure to detect and remove artefacts in heart rate data obtained from photoplethysmography wearables: A prospective cohort study

Key Result

A novel machine learning procedure operating on device-generated HR data effectively removed artefacts, achieving high agreement with reference activity scores (Pearson's r = 0.89-0.96).

Key Points

• The aim is to validate a novel machine learning method for detecting artefacts in heart rate data from PPG devices.
• 149 participants wore PPG devices and reference chest straps for 12 weeks.
• Participants were classified as 'PPG-compatible' based on prior testing.
• Artefact and activity detection models were trained on 75% of the data, while 25% was used for testing.
• Multiple probability thresholds were used to balance artefact removal and activity preservation.
• The Antwerp Activity Index was calculated for evaluating the effectiveness of the thresholds.
• 53.8% of participants were classified as PPG-compatible, with over 5 million data points collected.
• Artefact detection sensitivity was 58.5% in daily life and 76.6% during exercise.
• Specificity ranged from 74.9% to 91.3%, indicating effective artefact removal with minimal incorrect exclusions.
• The best threshold combination achieved a high agreement with reference AAI scores (Pearson’s r = 0.89–0.96).