supervision

2024

1. PhD

   Towards the development of a continuous model for the assessment of bacteremia, bloodstream infection and sepsis

   Oscar Escudero-Arnanz, and Bernard Hernandez

   PhD in Artificial Intelligence, Visiting researcher 2024
2. PhD

   Dynamic graph machine learning for early detection and characteristation of antimicrobial resistance outbreaks

   Oskar Fraser-Krauss, Bernard Hernandez, Alison H Holmes, and 1 more author

   PhD in Artificial Intelligence, AI4Health CDT Programme 2024

2023

1. PhD

   Clinical Decision Support Systems to support antimicrobial cessation

   William Bolton, Bernard Hernandez, Timothy M Rawson, and 2 more authors

   PhD in Artificial Intelligence, AI4Health CDT Programme 2023

2022

1. M.Sc.

   Signal Analysis and Classification of Photoplethysmography (PPG) Waveforms for Predicting Clinical Outcomes

   Wang Jingkai, Pantelis Georgiou, and Bernard Hernandez

   M.Sc. in Applied Machine Learning 2022

   Abs PDF Poster

   The photoplethysmogram (PPG) is a signal that measures pulsatile volume of blood in tissue and is collected by portable, non-invasive pulse oximeters. PPG signals can be used to predict the severity of dengue patients during dengue epidemics. However, raw PPG waveforms are susceptible to artifacts and noise. The classification of PPG signal quality is therefore crucial. In this work, PPG waveforms are classified into three categories according to their quality, and seven signal quality indices (SQIs) are applied to extract the features of each segment of PPG. At the same time, the spectrogram of each PPG sample is calculated as another feature. With the help of k-means clustering on the SQI features, each sample is labelled with a signal quality as ground truth. Apply an autoencoder to reduce the dimension of SQI features to observe the data samples in 2-D, and compare the PCA method. Using the above PPG samples and labels as dataset, apply decision tree, random forest, and AdaBoost classifiers on raw PPG data or SQI features. Build a convolutional neural networks (CNN) model for 2-D spectrograms to learn the training set data, and predict the test set data. Experimental results show that: SQIs are effective PPG features that facilitate quality labelling and dimensionality reduction. The spectrogram is also an effective PPG feature, and the CNN has good stability, good flexibility, high accuracy and fast speed to be an ideal PPG signal quality classification method.

2. M.Sc.

   Susceptibility Data Analysis and Visualization for Surveillance in Antimicrobial Resistance

   Jin Yujia, Pantelis Georgiou, and Bernard Hernandez

   M.Sc. in Applied Machine Learning 2022

   Abs PDF Poster

   In the early 20th century, the discovery of various antimicrobials solved the medical problem of death caused by infection with pathogenic microorganisms. With the large-scale use of antimicrobials, hospitals have found that during the use of antimicrobials, some microbes gradually develop resistance, making the antimicrobials ineffective. Each year, the number of resistant pathogenic microorganisms is increasing rapidly, but the number of new antimicrobials is not enough. If unchecked, address and deal with the Antimicrobial Resistance (AMR) problem, the annual number of deaths due to AMR will increase to 10 million by 2050. One of the most important causes of the AMR problem is the inap- propriate and excessive use of antimicrobials by doctors and pharmacists. Therefore, this project builds an AMR surveillance web page system that can interact with users. The system provides doctors with clear and accurate AMR information and recommendations based on the collection, processing, analysis, and visualization of recorded susceptibility test data.

2021

1. M.Eng.

   Analysis and Implementation of Data Imputation Techniques for Laboratory Data

   Agrim Manchanda, Pantelis Georgiou, and Bernard Hernandez

   M.Eng. in Electronic and Information Engineering 2021

   Abs PDF Slides

   Missing data is a ubiquitous problem in the clinical domain which impacts the fidelity of predictive modelling as many of the models rely on the completeness of data. Simple methods such as mean or median imputation exist but they do not preserve the relations between variables. As such, they do not capture and exploit feature correlations which exists between variables in laboratory data. This project presents the design and implementation of an imputation framework which provides a robust methodology to investigate two data imputation techniques, Machine Learning and Bayesian Networks, that inherently exploit feature relations to infer missing data values from those that are observed. The project carries out an empirical study on a real-life laboratory data set to compare the performance of the aforementioned techniques against simple median imputation. The project concludes with an evaluation of the studied techniques and recommendations on how the methods can be integrated as part of an existing clinical decision support system.

2. M.Eng.

   Patient similarity retrieval based on laboratory data cluster analysis and visualisation

   Stiff Oliver, Pantelis Georgiou, and Bernard Hernandez

   M.Eng. in Electronic and Information Engineering 2021

   Abs PDF

   Large scale medical datasets and electronic health records have become increasingly common, prompting the creation of clinical decision support systems which utilise data to enhance health care. Tools have been designed to predict the likelihood of infection, automate drug dosing and prescriptions, and evaluate a patient’s risk of death. Adoption, however, remains low due to inherent issues, most notably, concern from clinicians regarding the quality of recommendations and lack of transparency about how results are obtained. This report covers the design and implementation of a system that uses laboratory test data to retrieve and present information relevant to a current patient, allowing clinicians to make quicker diagnoses and create tailored treatment plans based on past data. Thus, using the increasing amount of healthcare data without impinging on clinicians’ diagnosis and decision-making process and avoiding the typical objections to automated systems. The proposed solution incorporates the results of dimensionality reduction and clustering algorithms into an interactive, web-based similarity retrieval application, demonstrating the potential of unsupervised-learning based tools in clinical settings.

3. M.Sc.

   Model Development for Understanding the Relationship Between Photoplethysmography (PPG) and Physiological Parameters of Dengue Patients

   Philippos Arkis Hadjimarkou, Pantelis Georgiou, and Bernard Hernandez

   M.Sc. in Applied Machine Learning 2021

   Abs PDF Poster

   Dengue is a mosquito-born virus that can cause a range of complications to those affected; ranging from skin rashes to even death. With no effective treatment currently available and vaccines that can have serious side effects, the virus is a major risk to human lives. Detecting and predicting a severe case of Dengue can be the turning point in the Dengue epidemic management, increasing survival rates of patients experiencing Dengue shock and improving healthcare facility processes. Using wearable photoplethysmography devices, which are cost effective and user-friendly, patients’ vital parameters can be continuously monitored and analysed, allowing for the recognition of patterns that may lead to the development of Dengue detection and prediction algorithms. This study will focus on the utilisation of photoplethysmography signals for the exploration, design and implementation of supervised machine learning models for the classification of Dengue patients’ records, based on clinically recorded physiological parameters. Signal and data processing techniques will be utilised, leading up to the definition of evaluative experiments and model development. From design and implementation to testing and evaluation, we will be presenting our findings in the following exploratory report, examining a recently formed PPG dataset.

4. M.Sc.

   Application of machine learning to inform clinical management of infectious diseases in Vietnam

   Anh Nguyen, Pantelis Georgiou, and Bernard Hernandez

   M.Sc. in Applied Machine Learning 2021

   Abs PDF Poster

   Throughout this project and thesis, a Severe Dengue Predictor was developed and deployed on a web-based application. The model uses LSTM as the basis for the neural network architecture because of its ability to handle time-series data, especially clinical data, which was proven several times through different researches. From the data set provided by OUCRU, a small number of features were selected based on correlation test and WHO Severe Dengue guidelines: vomiting, body temperature, respiratory rate, haemoglobin, haematocrit percent, platelet count, bleeding vaginal, bleeding mucosal, and abdominal pain. The neural network architecture has had it hyper-parameter tuned before beginning the training process. The training result was validating through K-Fold cross-validation process, which ensures that biased result is minimal. With models’ performance ranging from adequate to excellent, a web-based application has been developed to deploy those train models. A random simulation for the application is also shown in this thesis. Through those simulation, the performance of this machine learning application is verified to be adequate in real-life scenarios.

5. M.Sc.

   Natural language processing for characterization of computerized tomography and x-ray text reports

   Zhao George, Yuting Xing, Pantelis Georgiou, and 1 more author

   M.Sc. in Applied Machine Learning 2021

   Abs PDF Poster

   Named entity recognition (NER) systems are commonly build using supervised methods which requires corpora manually annotated with named entities. However, manual annotation is very expensive and laborious. In this paper, a novel method is proposed for training clinical NER systems that does not require any manual annotation. It only needs text corpus, a semantic type database like UMLS that can give entities semantic types, and a pre-trained BERT representation on biomedical data. Using these resources, with a few analyses, annotations are automatically obtained to train the BERT token classifier for NER tasks.

2020

1. Group

   COVID-19: Tracking

   Oliver Stiff, and Chen Su

   M.Eng. in Electronic and Information Engineering 2020

   Abs Supp Code Poster Slides

   The aim of the project is to develop a map-based visualisation platform which sources its data directly from the Lacewing diagnostic device. Featuring some of the most comprehensive and advanced features in the market today, it allows users to understand and analyse the extent of the virus’s proliferation.

2. Group

   COVID-19: Decision Support

   Yannis Panagis, Selena Tabbara, Yichen Liu, and 2 more authors

   M.Eng. in Electronic and Information Engineering 2020

   Abs PDF Code Poster

   The aim is to implement a clinical decision support system module for COVID-19. EPIC IMPOC (Enhanced, Personalised and Integrated Care for Infection Management at the Point of Care) is a clinical decision support system used in the NHS to improve the management of infectious diseases by facilitating data collection, infection diagnostics and antimicrobial therapy advice at the point of care. This will allow clinicians and medical care workers not only to retrieve similar patients from a database, but also to use pathology tests to identify whether or not COVID-19 as well as perform a risk assessment. This module will be integrated within EPIC IMPOC in the near future and tested in Clinical Trials in early July.

2015

1. B.Eng.

   Applying machine learning to the development of an intelligent decision support system for antimicrobial prescribing

   Jamie Kan Yan Ting, Bernard Hernandez, Pantelis Georgiou, and 1 more author

   B.Eng. in Electrical and Electronic Engineering 2015

   Abs PDF Poster

   An Intelligent Decision Support System (IDSS) was developed to model the decision making process of an expert so as to suggest suitable antimicrobial therapy options for a new case. A case is described by its attributes such as age and lactate levels. Using the technique of Case Based Reasoning (CBR), similar cases were retrieved using the K-Nearest Neighbour (KNN) algorithm. A novel approach was used to define the distance metric of the KNN algorithm by taking a weighted average of each attribute’s distance. A grid-search algorithm was used to find the optimal set of weights that best represents the relative importance of the attributes. Using the IDSS, the proposed solution had an average specificity accuracy of 94%, showing that the concept of CBR and KNN can be used for antimicrobial prescribing.