Deep Learning Ultra Low-Frequency Heart Rate Variability from raw ECG

Lead Research Organisation: University of Liverpool
Department Name: Institute of Ageing and Chronic Disease


Lay Summary:
This project will use new "Machine Learning" technologies to analyse Heart Rate Variability.

If someone says, "my heart beats steady as a rock", they probably need to be told that this is a warning of the increased likelihood of an impending heart attack. In contrast to many people's intuition, a healthy heart does not beat steadily like a rock (do rocks even beat?) or a metronome, but with an irregular beat. This natural and healthy variation between heartbeats is known as "Heart Rate Variability" (HRV) and is widely measured in sports and medicine, but the causes of the variability are not well understood. In this study, we will develop novel software to facilitate analysis of this irregularity and gain a better understanding of the biology behind it.

The heart does have an inbuilt pacemaker that beats with an apparently steady rhythm throughout adult life, but on top of this regular beat, there are two well characterised subconscious mechanisms that can accelerate or decelerate the heart-beat. The behaviour of these two modulatory mechanisms has been extensively studied and causes the heartbeat to change in the second by second or minute by minute timeframe. However, the heartbeat also changes over the course of hours or days and technical limitations have made this very difficult, if not impossible to study at this level of detail in the past. Essentially, human selection and inspection of clean strips of ECG traces was necessary and this was impractical for very large datasets. In the case of rodent ECG traces, it would mean visually inspecting over a million heartbeats per day! We believe that we can make use of new computer and software developments to study the long-term changes in HRV. Specifically "deep learning" a so-called artificial network, and major type of modern artificial intelligence (AI). This is similar software to that allowing Alexa or Siri to answer verbal commands in the latest smart devices. In this project, we will develop this type of software to assist with long-range ECG analysis and use further modern computer models to infer the biological mechanisms underlying this long-term HRV.

The applications for our software would be widespread, from health monitoring in people and pets and in fitness monitoring in sports people. Since changes in the way the heart is controlled are a major risk factor in ageing, distribution of such software will benefit the healthy ageing agenda.

Technical Summary

Heart rate variability (HRV) is a well-recognised phenomenon in people and other animals and the relatively high-frequency components, in the range 0.015 to 4Hz (termed very low-frequency vLF, low-frequency LF, and high frequency, HF) are used frequently and there is evidence that they derive from the activity of the autonomic nervous system. In reality, these are all rather high-frequency components in a physiological context, with time periods of even so-called "low frequency" components stretching to only a minute or so. Lower frequency components, ultra-low frequency, (uLF or those frequencies one might call sub-ultra-low frequency or frequencies in the microhertz (muHz)) range are have rarely been studies and the origins of this variability are completely unknown. Part of the reason for this knowledge gap is the fact that very good quality ECG signals are necessary for HRV and traces are typically inspected and selected by eye before conducting the analysis. This is infeasible for long-range signals and introduces significant experimental bias. Furthermore, long-term recording requires the subject to be freely moving and incurring "noise" dislodging electrodes etc. Commercial apparatus for such analysis disguise missing data and the user never knows, but the HRV signal recorded would become scientifically worthless. The past solution to the long-term analysis of HRV is to pick and analyse short segments taken over the course of a longer-term experiment, or session. The bias inherent in such workflows is obvious.
Our proposal is to use deep learning to automatically process ECG signals to create frequency periodograms with robust sub-ultra low-frequency powers, and then to simulate the biology of the muHz frequency powers with a Bayesian model. We hypothesis the circadian rhythm contributes to the microhertz frequency variability of HRV and we will test this using telemetric recording of HRV from rodents.

Planned Impact

Further details INCLUDING timelines included in the Pathways to Impact attachment.

When we have robust prototype deep learning models available we will approach our current collaborators Medtronic with regards to adapting their analysis pipeline to include deep learning; we have already spoken to them about data exchange, but we cannot offer validated prototypes until this project is completed. We also have good relations with Millar Telemetry and CED and have discussed collaboration; CED stated they would be happy to incorporate our models once they are distribution ready.

Training is a big part of this; our growing group will train many undergraduate students (RBJ had 5 undergrad project student this year) in far more mathematical and statistical projects than is typical in biology. The value of this project to the undergraduate skills base should not be overlooked since both physiology and mathematical biology are listed as endangered biological skills by the BBSRC. Our PDRA will be further trained in both computational analyses with deep learning and Bayesian inference to compliment her existing skills in cardiovascular biology.

This project will develop technology to significantly improve the amount of data that can be extracted from animal experiments using telemetric recording and so if the project is funded, we will be able to approach the NC3Rs with a proposal to deliver HRV analysis to animal units throughout the country. Telemetric recording will reduce severity (refinement) and reduce the numbers of animals needed (reduction), we also hope that our HRV model may allow replacement of animals in some mechanistic studies.

We will disseminate through the RBJ conceived and created Meet-The-Scientists event in Liverpool World Museum. We will demonstrate heart rate variability to museum visitors, which largely, but not exclusively consist of families. We will invite members of the public to come along and joint in some science! We will equip participants with HR monitors and explain about HRV and its pitfalls and (with adults) the limitations of the consumer devices they are most probably wearing. We will also demonstrate our machine learning detection methods to adults and older children. For biological scientists, we will host a workshop at the Physiological Society summer annual meeting and for data-scientists and hobbyists, we will run a Kaggle machine-learning competition.
Description DEVS Research Support
Amount £1,980 (GBP)
Organisation University of Liverpool 
Department Department of Eye and Vision Science
Sector Academic/University
Country United Kingdom
Start 01/2023 
End 06/2023
Description Japan Partnering Award: The paraventricular nucleus of the hypothalamus; networks and mathematical models.
Amount £50,755 (GBP)
Funding ID BB/S020772/1 
Organisation Biotechnology and Biological Sciences Research Council (BBSRC) 
Sector Public
Country United Kingdom
Start 05/2019 
End 06/2024
Description Signalling In Space And Time: Intracellular Cyclic AMP Dynamics In Human Vascular Smooth Muscle
Amount £446,542 (GBP)
Funding ID BB/V002767/1 
Organisation Biotechnology and Biological Sciences Research Council (BBSRC) 
Sector Public
Country United Kingdom
Start 04/2021 
End 11/2024
Title Systemic application of the TRPV4 antagonist GSK2193874 induces tail vasodilation in a mouse model of thermoregulation 
Description In humans, skin is a primary thermoregulatory organ, with vasodilation leading to rapid body cooling, whereas in Rodentia the tail performs an analogous function. Many thermodetection mechanisms are likely to be involved including transient receptor potential vanilloid-type 4 (TRPV4), an ion channel with thermosensitive properties. Previous studies have shown that TRPV4 is a vasodilator by local action in blood vessels, so here we investigated whether constitutive TRPV4 activity effects Mus muscularis tail vascular tone and thermoregulation. We measured tail blood flow by pressure plethysmography in lightly sedated Mus muscularis (CD1 strain) at a range of ambient temperatures, with and without intraperitoneal administration of the blood brain barrier crossing TRPV4 antagonist GSK2193874. We also measured heart rate and blood pressure. As expected for a thermoregulatory organ, we found that tail blood flow increased with temperature. However, unexpectedly we found that GSK2193874 increased tail blood flow at all temperatures, and we observed changes in heart rate variability. Since local TRPV4 activation causes vasodilation that would increase tail blood-flow, these data suggest that increases in tail blood flow resulting from the TRPV4 antagonist may arise from a site other than the blood vessels themselves, perhaps in central cardiovascular control centres. 
Type Of Material Database/Collection of data 
Year Produced 2021 
Provided To Others? Yes  
Description School Career Talk 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach Local
Primary Audience Schools
Results and Impact Ridgeway high school career talk
Year(s) Of Engagement Activity 2023