Michael Petrus (Namibia & Green Templeton 2025) is a medical doctor pursuing a DPhil in Medical Sciences (Cardiovascular Science.) His research integrates multi-omics, radiotranscriptomics, and deep learning applied to large-scale CT imaging to develop AI-driven precision diagnostics for heart disease.
The signal we are missing in heart disease: can CT scans tell us more than we’re currently seeing?
Wednesday 08 April, 2026
Michael during a clinical placement in Namibia
My six years of medical training in Namibia were intense. At the time, it all felt like long days, too many patients, and unfortunately, often not having enough to work with. Looking back now, I realise what that experience gave me. It showed me how our healthcare system actually functions, and that the challenges we encountered were not unique to Namibia. Limited resources, staff shortages, long waiting times, these are shared across much of Africa. In our settings, medical training often comes with early responsibility. We were part of clinical decisions, often with very little to work with. That level of exposure shaped how I understood and approached medicine.
By the time most patients came to us with cardiovascular disease, we were already too late. Patients arrived with chest pain that had been ongoing for months, sometimes years. Others came after a stroke, or with heart failure that had progressed quietly without obvious warning signs. There was often a lingering sense that things could have been different, that we might have picked it up earlier, or that patients could have presented sooner.
One case has stayed with me. A man in his fifties presented multiple times to local clinics with recurrent chest pain. Each time, his ECG, which records the electrical activity of the heart and is one of the first tests done when a heart problem is suspected, came back normal or inconclusive. When he was eventually seen at a district hospital, blood tests were done to check for cardiac biomarkers, but they were not raised. After five months, he was referred for a coronary CT scan, which uses X-ray imaging to produce detailed pictures of the heart's blood vessels. It showed no obstruction and no clear evidence of an acute event. He was treated, reassured, and discharged.
The chest pain kept coming. He moved between clinics and the hospital, returning again and again with the same complaint. The pattern repeated itself, no definitive findings, no clear escalation, only the absence of something obviously wrong. Months passed like this until he came back one final time. This time, he did not leave. He died of a myocardial infarction (a heart attack). This was not an isolated case. It left me asking: what would it take to find patients like him earlier?
Cardiovascular disease is the leading cause of death globally, yet in many settings our approach is still largely reactive. We wait for symptoms and respond once disease has already taken hold. Prevention is often based on broad risk factors such as blood pressure, cholesterol, and family history, measures that do not always reflect what is actually happening inside a patient's body.
Michael in a lab in Oxford
At Oxford, I am a first-year DPhil student in the Antoniades Group, working on a different way of approaching this problem. My research uses artificial intelligence, particularly machine learning, to extract more information from CT scans that patients are already having, whether for chest pain, suspected lung disease, or other reasons. These scans contain far more than what is used in standard clinical reporting. Beyond structural findings, they capture subtle patterns in fat distribution and tissue characteristics, signals invisible to the human eye but detectable through computational analysis.
My research here keeps bringing me back to that patient in Namibia, he already had a CT scan. By standard reporting, it showed nothing obviously wrong. But what if that same scan contained information that was never being read? What if the fat around his heart, or the way tissue appeared on imaging, was already carrying a signal that conventional analysis was missing? If we could detect that signal earlier, we could identify higher-risk patients before severe symptoms develop and direct them toward treatments and closer monitoring that already exist, while there is still time.
The focus of my work is fat, not body weight as we usually think about it, but the biology of where fat accumulates and what it looks like on imaging. Body Mass Index (BMI), the measure most of us are familiar with, is a blunt tool. Two people with the same BMI can have very different metabolic profiles. One might carry most of their fat under the skin, where it is relatively inactive. Another might have fat around the heart, within muscle tissue, or in the liver, depots that are biologically active and closely linked to inflammation, insulin resistance, and cardiovascular disease. A scale cannot tell the difference. A CT scan, interpreted properly, might.
My work builds on research from our group showing that the texture and density of fat around the heart, features invisible to the naked eye but detectable through machine learning, can identify individuals at higher cardiometabolic risk and predict future cardiovascular events. I am now extending this by bringing together signals from multiple fat depots across the chest, fat around the heart, beneath the skin, within muscle, and in the liver, to build a more complete imaging biomarker of cardiometabolic health. The aim is to extract this from scans that patients have already had, so that a single scan, read more carefully, becomes an early warning.
Michael in his department in Oxford
Here at Oxford, I am working in research environments with access to technologies that hospitals in Namibia, and across much of Africa, do not yet have. But imaging is becoming more widely available across the continent, and if we can show that meaningful risk is already encoded in the scans hospitals are collecting and build tools to read it automatically, those tools can travel well beyond Oxford.
The patient I described never had access to what we are working on, and neither do many others. That is the gap I am trying to close through my research.