1. Machine Learning Prediction Modelling in Computational Healthcare and Mental Health

(1.a) Predicting Risk of Dementia with Machine Learning using Routine Primary Care Records – CPRD

Participants: Daniel Stamate, Fionn Murtagh, Mihai Ermaliuc, John Langham, Charlotte Wu, in collaboration with Prof David Reeves and team at the Centre for Primary Care, University of Manchester.

We develop novel machine learning and statistical approaches to predict dementia risk using routine primary care records (CPRD), contributing to prevention and improved diagnosis rates. This work builds on our project on predicting the risk of dementia using primary care data , which has received media coverage including the BBC.

(1.b) Predicting Alzheimer’s and Dementia with Machine Learning and Statistical Approaches on ADNI, EMIF‑AD and ELSA

Participants: Daniel Stamate, Daniel Stahl, David Reeves, Henry Musto, Rostislav Vorobev, Ruslan Tsygankov, Olesya Ajnakina, in collaboration with King’s College London, UCL, Oxford University, EMIF‑AD partners, and University of Manchester.

We apply neural networks, deep learning, gradient boosting, Gaussian processes, SVMs, survival models, and other methods to ADNI, EMIF‑AD, and ELSA cohorts to predict Alzheimer’s disease and dementia.

• ADNI: deep learning, gradient boosting, Gaussian processes, SVM, survival ML
• EMIF‑AD: gradient boosting machines, random forests, deep learning
• ELSA: survival random forests, survival elastic net, Cox models, gradient boosting classification

(1.c) AI for Predicting Psychosis

Participants: Daniel Stamate, Daniel Stahl, Wajdi Alghamdi, Andrea Katrinecz, in collaboration with King’s College London, Maastricht University Medical Centre, and Yale University School of Medicine.

We investigate prediction modelling and pattern detection for first‑episode psychosis associated with cannabis use, and prediction of psychosis from experience sampling data using machine learning.

• Prediction Modelling and Pattern Detection Approaches for the First‑Episode Psychosis Associated to Cannabis Use
• Predicting Psychosis from Experience Sampling Data using Machine Learning

2. Identifying and Measuring Playful Parenting Using Machine Learning

Participants: Daniel Stamate, Caspar Addyman, Mark Tomlinson, Irene Uwerikowe, Jeremiah Ayock Ishaya, in collaboration with Stellenbosch University and the Global Parenting Initiative led by Oxford.

We use video data and machine learning to automatically estimate the quality of parent–child interactions via nonverbal synchrony. Models such as BiLSTM and BiGRU trained on movement patterns distinguish high and low synchrony, identifying dyads that may benefit from further support.

3. Predicting Spectral Reflectance Curves and Applications in Coatings Industry

Participants: Daniel Stamate, Asei Akanuma, Alexandra Stepanenko, in collaboration with Sherwin‑Williams.

Through Innovate UK Knowledge Transfer Partnership and Accelerated Knowledge Transfer projects co‑funded by Innovate UK and Sherwin‑Williams, we develop state‑of‑the‑art neural network and deep learning approaches to colour reflectance curve prediction to optimise the design of new coatings.

4. Machine Learning and NLP Sentiment Analysis in Finance

Participants: Daniel Stamate, Raph Olaniyan, and Frederic Marechal.

We analyse the relationship between sentiment‑filled online information and stock markets, focusing on whether sentiment resolves uncertainty or induces volatility. Our frameworks explicitly model non‑linearity, heteroscedasticity, and asymmetric influences of positive and negative sentiment, and extend towards financial fraud detection with NLP and ML.

5. Soft Computing, Evolutionary Algorithms and Applications

Participants: Doina Logofatu, Mihaela Breaban, Daniel Stamate, Ida Pu, in collaboration with Frankfurt University of Applied Sciences and University of Iasi.

We design efficient algorithms for optimisation and reasoning with imperfect information using evolutionary computing (genetic algorithms, particle swarm optimisation) and soft computing based on multi‑valued logics. Applications include query optimisation, integrating and querying imperfect information, and improving time‑efficiency in networking and concurrency problems.

6. Generative based one-class classifier algorithms

Participants: Mihai Ermaliuc, Daniel Stamate, George Magoulas, Dragos Gavrilut, in collaboration with Birkbeck, University of London, University of Iasi, and Bitdefender

This direction investigates novel one-class classification methods based on generative AI models achieving state of the art classification performances. One approach we investigate is to use ensembles of Generative Adversarial Network (GAN) discriminators to distinguish a target class from everything else, leading to a GAN based one class classification algorithm we introduced - GANOCC.

Instead of modelling the target class directly, GANOCC focuses on generating synthetic out-of-class samples near the target class boundary and training multiple discriminators to separate them from real data. Applications range from novelty, to anomaly and malware detection.