This page summarizes some of the key research areas that members of the group work on. Most of these projects build on national and international collaborations with academic and industrial partners, and are by nature inter-disciplinary.
Large externally-funded projects
This project aims to optimize the use of innovative ambient and passive data collection methods towards assessing sleep and circadian rhythms, developing personalized models and facilitating early intervention.
This project builds on a large national collaboration, led by colleagues from the University of Cambridge, UCL, Edinburgh and KCL. The project focuses on understanding visceral pain, specifically ways to understand and improve how to treat people with visceral diseases. Thanasis is a Co-I and our group contributes in the analysis of the multimodal data collected (clinical tests, self-reports, data from wearable sensors).
This project focuses on endometriosis in collaboration with Prof. Andrew Horne and Prof. Philippa Saunders. Thanasis is the Lead for Digital Health within the EXPPECT team and our group contributes in the analysis of the longitudinal multimodal data collected (self-reports, data from wearable sensors).
This project aims to provide new insights into health-related problems particularly for people who are in their 50's. This is work in collaboration with Prof. Wendy Loretto and team at the Edinburgh Business School and colleagues from Design Informatics at the University of Edinburgh. Thanasis is a Co-I and our group contributes in the analysis of the multimodal data collected (self-reports, data from wearable sensors).
Data analytics methods with wide impact and applicability
Time series analysis & signal processing processing
We are developing generic signal processing tools with wide impact and applicability.
speech signal processing algorithms based on nonlinear concepts, e.g. see JRSI2011. See also the GitHub repository for the MATLAB source code on speech signal processing algorithms.
signal processing algorithms to process multimodal data, e.g. in wearable sensor data, see JMU2020; Sensors2022; Sensors2023. See also further information in the GitHub repository for MATLAB source code.
signal processing algorithms towards mining geolocation data, e.g. see TBME2017
time-series analysis algorithms towards extending the applicability of wavelets combined with exponential weighted local smoothing, e.g. see FHN2015. See also the GitHub repository for the MATLAB source code.
time-series analysis algorithms towards exploring patterns in longitudinal self-reports, e.g. see JAD2016
Statistical machine learning
We are developing generic statistical machine learning tools with wide impact and applicability.
Indicative areas of particular interest include:
Feature selection algorithms, e.g. see Patterns2022. See also the GitHub repository for the MATLAB and Python source code for feature selection.
Information fusion algorithms, e.g. see JASA2014. See also the GitHub repository for the MATLAB source code for information fusion.
Feature selection/projection and cluster analysis, e.g. see IJMI2023
Application areas of ongoing research interest
Biomedical speech signal processing
We are so used to speaking from a young age, that is easy to underestimate the complexity of this convoluted process. Speech requires the delicate co-ordination of a large number of muscles and organs, starting from the lungs, the vocal folds, and the vocal tract. Viewed under this prism, it is then hardly surprising that speech signals may convey considerable information which can be capitalized for biomedical applications.
We have been developing novel signal processing algorithms to characterize speech. For example, we have demonstrated that we can differentiate people with Parkinson's disease from healthy controls with almost 99% accuracy, we can track Parkinson's disease symptom severity more accurately than the difference between two trained specialists (inter-rater variability), and we can automate the process for self-monitoring voice rehabilitation with about 90% accuracy.
More recently, we have demonstrated that the developed methodology can be applied to assist in forensic phonetic analysis, where we can probabilistically very accurately determine whether a suspect is the culprit: remarkably, we have demonstrated this is possible merely using voice fillers.
Longitudinal monitoring of neurodegenerative disorders
We had previously demonstrated that we can accurately differentiate healthy controls from people with Parkinson's disease, and also replicate UPDRS, the standard Parkinson's disease symptom severity metric, with accuracy which is less than the difference between expert neurologists (inter-rater variability) using speech signals. Although extremely promising, these results 'only' manage to demonstrate that we can automatically replicate the clinicians' assessments, and do not capture the true underlying fluctuation of symptoms on a daily basis which really matter to patients.
We are currently working towards integrating additional wearable sensors and smartphones to elicit daily responses from people with Parkinson's disease and motor neurone disease, in an attempt to understand better the daily fluctuations of motor and non-motor symptoms in Parkinson's disease.
We envisage collecting a high-resolution dataset which may facilitate key new insights into Parkinson's disease symptom trajectory, whilst providing a platform to associate self-perception and clinical assessment of symptoms.
Physical and mental health assessment using wearable sensors & smartphones
Mental health problems are on the rise globally. Clinical diagnosis and assessment is challenging, subjective, and relies on highly specialized experts, whereas treatment effects are very difficult to be quantified and objectively monitored.
We have developed a novel approach combining Patient Reported Outcome Measures (PROMs), where patients themselves self-monitor their symptom daily, and the use of wearable sensors and smartphones. The self-reported PROMs are recorded on an Android smartphone app we developed in house, time-stamped, and processed offline to assess trends and symptom variability: this enabled us to gain key new insights which are otherwise missed in interim patient visits to the clinic. Furthermore, we reported new insights into diurnal variability in people diagnosed with bipolar disorder or borderline personality disorder using electrocardiography signals, and also demonstrated we can assess depression by analyzing geolocation patterns collected from smartphone data. Collectively, these tools provide a better day-to-day monitoring of patients.
We are currently extending this prior work, also expanding on additional disorders.
Asthma characterization and monitoring
Asthma is a chronic disease characterized by inflammation of the airways to the lungs, estimated to affect 235 million people worldwide. Despite both the availability of, and high spending on, asthma treatments, the burden of asthma mortality and morbidity is high. We are working in close collaboration with AUKCAR and the team of Aziz Sheikh to study practical aspects of asthma.
Specifically, we mine electronic health records and investigate medication non-adherence. Failing to assess non-adherence as a causal factor for poor asthma control may lead to inappropriate dose escalation and underestimated incidence of adverse effects. We are working towards characterizing adherence taking into account patient specific characteristics, including history of asthma attacks, and drug prescriptions.
Furthermore, asthma is increasingly recognized as an umbrella term used to describe a group of distinct diseases leading to similar symptoms. Numerous groupings of asthma have been suggested (e.g. early-onset atopic, severe) but at present none fully account for the heterogeneity observed across asthma patients. We are working on developing a new robust clustering approach which will identify clinically interpretable subtypes.
Cardiovascular disease assessment
In early days we developed a comprehensive closed loop model of the cardiovascular system to understand the interactions of lumped parts, and study the effect of fitness, smoking, drugs, and cardiac arrest.
More recently, we have investigated the use of high sensitivity troponin, a biomarker used to assess heart attack (formally known as myocardial infarction), to explore whether we can improve clinical outcome assessment.
We are currently working with some of the world's top experts in this field within Nick Mill's team at the University of Edinburgh to extend these findings aiming to provide better patient subtyping through clustering algorithms, and automating the process of myocardial infarction assessment using a range of routinely collected clinical tests and state-of-the-art statistical machine learning algorithms.