Director of the DPhil Programme (Social Data Science)
Research Fellow, Wolfson College, Oxford
Lecturer, Lady Margaret Hall, Oxford


Oxford Internet Institute
University of Oxford
1 St Giles
Oxford OX13JS, UK

Research interest

Data Science: Machine Learning, Statistics, Modelling
Computational Medicine: Digital Health, Healthcare Policy, Social Data Science




Machine Learning in Healthcare

From around February 2020 I lead Oxford COVID-19 (OxCOVID19) Project aimed at increasing our understanding of the COVID-19 pandemic and elaborate possible strategies to reduce the impact on the society through the combined power of Statistical, Mathematical Modelling and Machine Learning techniques. [Read more].
ICU is the most heterogeneous population in the hospital, with the highest rates of acute and chronic multimorbidity. In this project we apply machine learning clustering algorithms to identify subgroups of ICU patients having similar characteristics. We show how to use global sensitivity analysis to test the stability of the resulting clustering algorithms.

In this project I lead the effort (with H Khandahari) to apply Active Learning techniques to ECG labelling for arrhythmia classification in the ICU. We aim to re-label all the MIMIC waveforms to produce one of the largest dataset of its kind.

Sepsis is a life-threatening condition occurring in an estimated 30 million people worldwide and with 6 million people die from sepsis each year. Here (with J Hawrych and J Morrill) we develop a machine learning algorithm that predicts the occurrence of sepsis in the ICU using routinely collected data: vital signs, laboratory values and demographics. This is a follow-up work on the 2019 Physionet Challenge (with MAF Pimentel, A Mahdi, O Redfern, MD Santos, L Tarassenko), Uncertainty-Aware Model for Reliable Prediction of Sepsis in the ICU. Computing in Cardiology (2019), 1-4. [Pdf]
We benchmark a feature-based and a deep learning approach in classifying short ECG segments as proposed by the Physionet/Computing in Cardiology Challenge 2017.

F Andreotti, O Carr, MAF Pimentel, A Mahdi, M De Vos
Comparing feature-based classifiers and Convolutional Neural Networks to detect arrhythmia from short segments of ECG. Computing in Cardiology 44 (2017), 1-4. [Pdf]
We found that a large number of hospitalised patients in the UK were found to have mean blood pressures that exceeded the diagnostic threshold for hypertension, without any notation of treatment or diagnosis of hypertension on their medical records.

A Mahdi, LC Armitage, L Tarassenko, P Watkinson
Estimated prevalence of hypertension and undiagnosed hypertension in a large inpatient population: A cross-sectional observational study.
We found that hospitalised patients’ circadian patterns of BP largely mirror those found in the community. High-quality hospital data may allow for the identification of patients at significant cardiovascular risk through either opportunistic screening or systematic screening.

A Mahdi, P Watkinson, RJ McManus, L Tarassenko
Circadian blood pressure variations computed from 1.7 million measurements in an acute hospital setting.
American Journal of Hypertension 32 (2019), 1154-1161 [Pdf]
We developed a new physiology-based, data-driven ABP model capable of generating many known characteristics of human blood pressure waveforms.

A Mahdi, GD Clifford, SJ Payne
A model for generating synthetic blood pressure waveform