An international study led by RISE-Health and FMUP shows that simple statistical models can outperform advanced algorithms in predicting cardiovascular risk.
An international study led by RISE-Health and FMUP has shown that regularised regression methods — a simpler statistical technique — generalise better than more complex machine learning algorithms in predicting serious cardiovascular events in patients with type 2 diabetes who have recently suffered an acute coronary syndrome (ACS); a population that remains at high cardiovascular risk in the months following the event.
The study, published in the Nature group’s journal Scientific Reports, compared – in collaboration with researchers from France, Canada and the United States – 11 different machine learning and regularised regression methods, using data from 5,107 patients in the EXAMINE clinical trial, which combines clinical information with a panel of 93 circulating biomarkers.
According to António Barros (RISE-Health/FMUP), this scientific study confirms that there is no systematic advantage of machine learning methods over regularised regression in clinical prediction models, particularly when the number of events is limited relative to the number of variables.
According to the expert from RISE-Health, the RISE Associated Laboratory and the Faculty of Medicine of the University of Porto (FMUP), “highly flexible algorithms model the randomness of the training sample as if it were a signal, whilst regularised models shrink the coefficients, become deliberately more conservative and, in turn, perform better when new patients are introduced”.
The scientific study showed that the best-predicted outcomes were cardiovascular death or hospitalisation for heart failure, rather than atherothrombotic events. “A composite outcome combining heart attack, stroke and cardiovascular death forces the model to capture distinct biological processes simultaneously, and plaque rupture at a specific moment depends on factors that a baseline blood sample does not reflect. Heart failure, on the other hand, is the outcome of a chronic and measurable process: ventricular overload, remodelling, renal dysfunction. This is where circulating markers add further information to that already provided by clinical variables,” he explained.
The study Comparative analysis of machine learning and regularised regression models for predicting cardiovascular outcomes in patients with type 2 diabetes after acute coronary syndrome, published in the journal Scientific Reports, is authored by António Barros, João Sérgio Neves, Adelino Leite-Moreira and João Pedro Ferreira, researchers at RISE-Health and FMUP. The scientific paper also features authors from the Université de Lorraine (France), the McGill University Health Centre (Canada) and the University of Connecticut (USA).