Attend the Virtual Global Study Fair | Register Now Attend the Virtual Global Study Fair | Register Now

Uncertainty quantification for machine learning-based analysis of photoplethysmography signals using synthetic data

   Faculty of Life Sciences & Medicine

This project is no longer listed on and may not be available.

Click here to search for PhD studentship opportunities
  Dr Jordi Alastruey, Dr Manasi Nandi  No more applications being accepted  Funded PhD Project (UK Students Only)

About the Project

Applications are invited for a 3 year full-time PhD studentship starting on 1st October 2023.

Background and Project Aim:

Photoplethysmography (PPG) signals are generated by smartwatches and contain a wealth of valuable physiological information for monitoring and diagnosing various health conditions, but this information is not routinely exploited [1]. PPG signals can be used to estimate important clinical measurements, such as systolic and diastolic blood pressure, pulse pressure, and pulse wave velocity. They can also be used to detect various diseases or conditions. These tasks involve regression and classification problems using both machine learning (feature-based) and deep learning (image-based) models. Confidence in the predictions of both types of models is crucial, since false negatives can be harmful to patients and potentially fatal, and false positives can lead to unnecessary treatment and patient anxiety.

This project is part of the EURAMET European Partnership on Metrology project 22HLT01 “Uncertainty quantification for machine learning models applied to photoplethysmography signals” that seeks to quantify the uncertainty of existing machine learning (ML) and deep learning (DL) models using PPG data. The PhD project’s goal is to create new datasets of synthetic (simulated) PPG signals that can be used to compare the accuracy and uncertainty of existing ML and DL. By doing so, we aim to identify models which have high accuracy and low uncertainty for a wide range of tasks. Synthetic signal datasets enable pre-clinical testing of pulse wave analysis algorithms across a wide range of cardiovascular conditions, in a relatively quick and inexpensive manner [2].

The project objectives are to:

  • Generate a dataset of synthetic PPG signals in about 25,000 virtual subjects of both sexes aged 25 to 75 years, under both physiological and pathophysiological conditions;
  • Validate the synthetic dataset by comparing it to in vivo data;
  • Quantify the uncertainty of at least one classification and one regression existing ML and/or DL model.

Project Description:

The following tasks will be carried out to fulfil these objectives:

Task 1: Cardiac contraction/filling simulation (6 months) – Our existing in-house code for simulating blood flow in the larger systemic arteries will be further developed to include the contraction and filling of the heart. To achieve this, the one-dimensional (1-D) formulation solved by the code will be coupled to an elastance zero-dimensional (0-D) model of the heart [3]. The existing model enables the simulation of PPG signals, and this task will investigate the sensitivity of the simulated PPG signals [4] to changes in cardiac and central vascular properties.

Task 2: Generation and validation of datasets of virtual subjects (12 months) – Datasets will be generated using the blood flow model developed in Task 1, following a methodology developed at KCL [4] which involves (i) identification of the most influential physical parameters that change with age, sex and disease from clinical data; (ii) simulation of virtual subjects using all possible combinations of the age-, sex-, disease-specific mean value for all parameters in step (i); and (iii) verification by comparing the simulated hemodynamic measures and derived indices with corresponding in vivo data, at the cohort level. Data from the clinical literature will be used to validate synthetic PPG signals under physiological and pathophysiological conditions affecting the heart (e.g., HFpEF and aortic valve disease) and vasculature (e.g., vascular ageing and peripheral arterial disease). The resulting benchmark datasets for classification and regression problems will be make publicly available, along with a brief description of the benchmark problems.

Task 3: Uncertainty quantification (UQ) (18 months) – The PhD student will collaborate with other members of the EURAMET consortium to investigate which UQ methods are suitable for different ML and DL models based on the use of features and image transformations of the signal, including the SPAR attractor [5]. The UQ methods employed will consists of both model-dependent and independent approaches encompassing both aleatoric (data) and epistemic (model) uncertainties. This study will consider PPG signals of varying quality, signals from different sites on the body, and different demographics, as well as both classification and regression problems.

The ultimate goal is to determine which models have high accuracy and low uncertainty for a wide range of tasks, allowing for the reliable use of PPG signals in a medical context. This will be of great benefit to digital health companies that are developing software for use in both medical devices and consumer wearable devices, some of which are part of the EURAMET consortium. Research papers will be submitted for publication in high impact peer-reviewed journals and the PhD work will be presented at relevant international conferences. At the end of the project all synthetic PPG datasets will be made widely available to the medical device and digital health communities.


  1.  Charlton et al. Wearable photoplethysmography for cardiovascular monitoring. Proceedings of the IEEE 110(3) 355–381, 2022
  2.  Existing open access datasets of in silico pulse waves at KCL (
  3. Alastruey et al. Arterial pulse wave modelling and analysis for vascular age studies: a review from VascAgeNet. American Journal of Physiology – Heart and Circulatory Physiology (online ahead of print), 2023
  4. Charlton et al. Modelling arterial pulse waves in healthy ageing: a database for in silico evaluation of haemodynamics and pulse wave indices. American Journal of Physiology – Heart and Circulatory Physiology 317(5) H1062–H1085, 2019
  5. Aston et al. Beyond HRV: attractor reconstruction using the entire cardiovascular waveform data for novel feature extraction. Physiological Measurements 1;39(2):024001, 2018

Informal email enquiries from interested students to the supervisors are encouraged (contact details below).

Dr Jordi Alastruey – email [Email Address Removed]

Dr Manasi Nandi – email [Email Address Removed]


Application consists of completing an online application. Please visit the studentship webpage before applying for further details on the programme and eligibility.

PhD saved successfully
View saved PhDs