Completing a BHF Immediate Postdoctoral Research Fellowship

In 2020 I was awarded a 5-year Immediate Postdoctoral Basic Science Research Fellowship by the British Heart Foundation (BHF). During the fellowship I investigated how clinical and consumer devices (such as wearables) could be used in screening for atrial fibrillation (AF, a common arrhythmia). In this piece I summarise the key achievements of the fellowship.

Scientific Findings

Summary: This fellowship advanced methods for detecting atrial fibrillation (AF) using wearable and mobile devices. We showed that wearables are broadly acceptable to older adults, developed signal processing tools to improve accuracy, and investigated strategies to make AF screening more reliable and efficient. These findings support the design of future AF screening programmes and wearable technologies.

SAFER Wearables Study:

Wearables hold promise for identifying undiagnosed AF in daily life. However, it is not yet clear whether wearables are acceptable to older adults (who are at greatest risk of AF), or how well they perform in practice. The SAFER Wearables Study investigated the acceptability and performance of wearables for AF screening in older adults. Community-dwelling adults aged 70 and over were asked to wear three devices simultaneously for one week: a chest-patch that recorded a continuous ECG signal, a wristband-style heart rate monitor, and a smartwatch-style device that continuously monitored rhythm and prompted four 30-second ECGs each day. In the published interim analysis, we found:

• Most participants were happy to wear any of the three devices for one week
• Some participants experienced skin irritation, particularly from the chest-patch. In a minority of cases, this led to early removal of devices.

Final results will be published soon.

Cardiovascular signal processing:

A key step in establishing criteria for detecting possible AF from everyday wearables is to develop a robust pipeline for detecting AF from the cardiovascular signals that wearables measure. We contributed at each stage of the pipeline:

• Detecting heartbeats in cardiovascular signals: We benchmarked open-source algorithms for beat detection in ECG and PPG signals, identifying the most reliable algorithms and the conditions in which they failed (e.g. during movement or low signal quality). We also developed a state-of-the-art PPG beat detection algorithm that combines high accuracy with high efficiency. This work provided recommendations for device design to ensure heartbeats are detected as accurately as possible. The algorithms were made openly available in the ppg-beats toolbox, allowing others to use them.
• Automatically determining when interbeat-intervals are accurate: We investigated methods for automatically determining when interbeat-intervals are likely to be accurate enough for AF detection. We found that many periods of inaccurate PPG-derived interbeat-intervals can be identified by analysing simultaneous accelerometry signals to determine whether the subject is sufficiently still for accurate measurements. We also found that accuracy could be improved further by running multiple beat detection algorithms, and only analysing those periods in which the algorithms agreed on heartbeat timings.
• Identifying AF from interbeat-intervals: We assessed the performance of several statistical metrics for discriminating between AF and non-AF heart rhythms, identifying those with best performance.

Predicting AF from sinus rhythm ECGs:

At the start of the project, we assumed patients would receive in-person training on screening devices, and that training ECGs might be useful to predict who would show AF during subsequent monitoring, potentially avoiding the need for some patients to take a device home.

Therefore, we investigated the feasibility of predicting who would show paroxysmal AF during screening based on a single, sinus-rhythm ECG. When using a deep-learning algorithm to analyse single-lead ECGs collected in two large studies (SAFER and STROKESTOP), we found that it was possible to predict paroxysmal AF (AUC of 0.80 in SAFER, and 0.62 in STROKESTOP), and that performance was improved with a wider age distribution (SAFER had an age range of 65-90+, whereas STROKESTOP was 75-76 year olds).

Investigating the feasibility of AF screening using mobile ECG devices:

Many approaches to AF screening involve using a mobile ECG device (such as a smartwatch equipped with ECG recording functionality, or a handheld ECG recorder). Such devices typically allow the user to capture a 30-second ECG which can be used for diagnosis. During this fellowship we analysed data collected in the SAFER Programme to investigate AF screening strategies using handheld devices. Our key findings were:

• The group found that it was feasible to screen “for all ages over 64 years, including people aged 85 and over”. I contributed ECG-level analyses to this study.
• Furthermore, the group demonstrated the feasibility of “population based screening using hand-held devices without face-to-face contact” (here). Similarly, I contributed ECG-level analyses to this study.
• I led a study of diagnostic agreement between cardiologists. Our main finding was that for every 100 participants on whom two cardiologists agreed on a diagnosis of AF, there would be a further 70 where they disagreed. We identified ECG signal quality as a key factor associated with the reliability of diagnoses.
• Following this, we investigated approaches for improving ECG quality, and found that amongst those participants who did not record adequate quality ECGs, the quality of their ECGs naturally improved over time without the need for further training by study staff.
• We also investigated strategies for reducing the clinical workload associated with screening for AF, whether by excluding ECGs from clinical review according to initial automated analyses, or prioritising for review those showing the highest likelihood of AF based on automated analyses.

Investigating the determinants of PPG signal quality

The PPG signal, which is often used in smartwatches for heart rhythm monitoring, is highly susceptible to noise. We investigated factors associated with the quality of PPG signals, and found that lying down and holding the hand at heart height were associated with higher quality signals. Furthermore, we found that “automatic adjustment of the light intensity of the sensor may help maintain high signal quality across different skin tones” (here).

Additional Important Achievements

Prizes

During the fellowship I was awarded the following prizes:

• The AJP-Heart and Circulatory Physiology 2024 Best Review Article Award
• The Martin Black Prize for the best paper published in Physiological Measurement in 2022
• An Academic Early Career Award from the Institute of Physics and Engineering in Medicine, in 2021

Additional funding

I was also part of teams who secured additional funding for:

1. A collaborative European project investigating uncertainty quantification for PPG analyses, including work on detecting AF from the PPG (€2.1million).
2. A network of researchers focused on ‘multimodal AI’, initially funded by the Alan Turing Institute, and subsequently funded by the EPSRC (£1.75 million).

In addition, i secured funding for a PhD studentship on the topic of ECG analysis for use in AF screening.

Leadership

• I led two chapters for a leading textbook on photoplethysmogaphy.
• I contributed to work on establishing the potential role of wearables in cardiovascular monitoring, by leading a group of 50 researchers to develop The 2023 Wearable Photoplethysmography Roadmap, and contribute to an International Society of Hypertension position paper on the use of wearables for BP monitoring.
• I held a leadership role in a network of European researchers investigating vascular ageing, including an award-winning review of using PPG devices to assess vascular ageing.
• I made datasets openly available for further research into cardiovascular signal processing and AF detection.
• I contributed towards key open-source toolboxes for cardiovascular signal processing (namely pyPPG, and Neurokit).
• I contributed to a review of the impact of skin tone on pulse oximeter accuracy, which was an important contribution at the time of the COVID-19 pandemic, when pulse oximeters were widely used, but found to be less accurate in patients with darker skin.

Development

Developing others

The fellowship provided many opportunities to contribute to others' development.

For instance, I supervised a total of 17 students, 11 of whom published their work in journals or at international conferences (see here for further details). This included:

• Obtaining funding for and supervising a PhD student as primary supervisor, whose research focuses on using mobile ECG devices for AF screening.
• Seeing two PhD students through to completion as an additional supervisor, both of whom were studying cardiovascular monitoring technologies.
• Supervising several research projects, including: four technical master’s projects, one medical academic foundation programme placement, eight medical students undertaking short projects (mainly 6-week Student-Selected Component projects), and one bachelor’s student undertaking a short project.

Ran educational sessions on wearable data analysis, reaching over 200 students and professionals from a range of disciplines, including engineers and clinicians.

Personal Development

During the course of my BHF fellowship, I developed significantly as a researcher, gaining expertise in supervision, clinical research, academic writing, collaboration, leadership, and technical skills.

I grew in confidence in supervision, having mentored multiple undergraduate projects and taken on the role of principal PhD supervisor for the first time. I supervised students from both medicine and engineering, guiding them to first-author journal publications in both medical and engineering journals.

Through leading the SAFER Wearables Study as Principal Investigator, I gained hands-on experience in running a clinical study. This involved taking responsibility for all aspects, from study design and ethical approval to NHS recruitment, participant engagement, data analysis, and dissemination. Additionally, being part of a team running a large clinical trial provided valuable insights into clinical research at scale.

I refined my academic writing skills, developing a structured and efficient approach to preparing research manuscripts. I now feel much more confident in structuring papers, presenting findings effectively, and streamlining the writing process.

I also gained confidence and experience in collaborations, initiating new academic partnerships and working closely with industrial partners. This broadened my understanding of the commercial aspects of developing wearable technologies. Additionally, through my involvement in a large trial of ECG-based screening for atrial fibrillation (AF), I gained a deeper understanding of public health, screening strategies, and the clinical applications of ECGs.

Finally, I developed leadership skills by coordinating various teams, including leading a group of 50 researchers to roadmap the future of wearable technology, chairing the ‘Physiological Measurement’ special interest group within the Institute of Physics and Engineering in Medicine, and leading a multi-disciplinary European team on physiological sensing (funded by the European Cooperation in Science and Technology). I also organised international conference workshops, special sessions, and national events on physiological measurement in clinical practice. Recently, my colleagues and I were successful in securing funding for a network on ‘multimodal AI’, funded by the EPSRC, providing further opportunity to lead teams and serve the scientific community.

From a technical perspective, I expanded my programming skills by working with Python, which has supported my research in physiological measurement.

At the end of the fellowship, I moved to Nokia Bell Labs to continue research into wearables in healthcare. This role provides opportunity to deepen my machine learning skills, and to learn more about commercialising research.