Translational Artificial Intelligence for Healthcare

Carousel imageCarousel imageCarousel imageCarousel imageCarousel imageCarousel imageCarousel imageCarousel imageCarousel imageCarousel imageCarousel imageCarousel imageCarousel image

We are an interdicplinary team interested in the development and translation of computational methods to address challenges involved in making sense of complex biomedical and health data.


We collaborate with clinicians and neurobiologists on a wide spectrum of problems arising in the context of mental health and neurodegeneraition. We develop tools to augment clinical decision making and effectively leverage information available from routinely collected electronic health records. We develop novel methods for dealing with sequential data and ordered events.


We apply methods of deep learning and computer vision to high-content fluorescent imaging and cell painting to understand the morphological changes in brain cells to advance drug discovery for Alzhemer's disease.


Much of my research is dedicated to a number of deeply interconnected themes:

A Path Signature Approach for Sequential Data

Sequential data are ubiquitous. Chronological medical records, time-series financial data, strings of texts, videos and many more other examples share the same characteristic that the data elements are ordered. An ordered collection of data points can be regarded as a path, often in high dimensions. The signature is an object associated with a path which captures many of the path’s important analytic and geometric properties. The paths signature represent a non-parametric representation of a path and servers as a feature set for downstream analytical tasks. The signature embeddings can also be learnt as part of the neural network training procedure through backpropagation.

Natural Language Processing for Medical Records

Digitalisation of medicine offers a unique opportunity to access and analyse voluminous biomedical data at scale. Electronic medical record systems collect and collate data over time, augmented with state-of-the art computational methods, hold promise to transform population and at individual patient level health. Unstructured free-text clinical notes contain rich information about the patient's health, exposures and outcomes, but require complex processing to extract clinicaly meaningful information. Natural Language Processing (NLP) and Information Extraction methods can parse strings of text into structured format, suitable for donwstream analytical tasks. The UK-CRIS is a federated network of 12 Mental Health NHS Trusts that manages access to the largest database of secondary care electronic medical records of over 3 million patients in England. This provides an unpresedented opportunity for large scale studies with observational data. The availability of geographical and populational diversity of free-text data from UK-CRIS NHS Trusts, allows researchers to delve into nuanced and cultural psychiatry studies in secondary mental healthcare.

Augmented Clinical Decision Support

Statistical and machine learning models are capable of identifying patterns in large datasets and provide physicians, clinical staff, patients and other stakeholders with meaningful insights and person-specific information to inform clinical practice and care. The algorithms can early identify deterioriating patietns or estimate a risk score of future events and alert clinical staff. In collaboration with Translational Gastroenterology Unit and using the TrueColours Inflammatory Bowel Disease (TC-IBD) online symptom monitoring system, we developed a novel prognostic tools to estimate the disease activty risk. This tools allows to rank patients based on the severity of their symptoms, visualise patients' symptoms and optimise workload for clinical staff.

Deep Learning for Cell Painting and Phenotyping

Dementia and neurological disorders are on the rise in the UK and globally. It is estimated that over 1 million people in the UK will be affected by dementia by 2025, which costs over £26.3 billion a year. Drug discovery for dementa is challenging due to the lack of drugable targets and accurate estimation of cellular phenotypes. Advances in computer vision and deep learning offer new opportunities for high-throughput image analysis, patter recognition of celluar perturbations and the discovery of biochemical interaction through which a candidate drug interacts with molecular targets. Using the Cell Painting method, which as an image-based morphological assay, deep learning algorithms and topological data analysis, we develop novel methods to quantify drug-induced morphological changes in cells for drug discovery.


Funded by: