Artificial intelligence can predict better than real doctors when patients with serious heart disorders are likely to die. That's according to a paper published this week in Radiology.
A team of medics and computer scientists, led by the MRC London Institute of Medical Sciences (LMS) at Imperial College London, created what is claimed to be the first computer program that uses machine learning to study heart disease.
Pulmonary hypertension, a condition that increases the level of pressure exerted on the arteries that supply oxygen to the lungs, is dangerous if left untreated. It affects up to 7,000 people in the UK, and a third of patients die of heart failure within five years of diagnosis.
The risk of death is typically calculated by radiologists by continuously monitoring heart function through measurements taken by hand. The AI software, we're told, can analyze MRI scans and other readings within seconds, and make a prediction almost instantly: armed with this advice, doctors should be able to pull together better treatment plans faster.
“The computer performs the analysis in seconds and simultaneously interprets data from imaging, blood tests and other investigations without any human intervention. It could help doctors to give the right treatments to the right patients, at the right time,” said Tim Dawes, coauthor of the paper and a researcher at University College London.
The team's paper shows that patients have to undergo cardiac magnetic resonance imaging to assess their heart function. The scans are projected onto a virtual three-dimensional model to map the direction and magnitude of pressure building in the heart’s right ventricle. The patients each have to walk for six minutes, and the distance covered is added to the mix of data analyzed by the software along with the heart models.
Linear regression, a supervised machine learning technique, is used to track the relationships between all of these variables linked to heart performance to estimate the risk of heart failure as the disease progresses.
Patients are sorted by the AI into risk classes of "very high", "high," "moderate" and "low" risk. Patients at "very high risk," for example, have a 40 per cent chance of survival after five years, compared to around 90 per cent for "low risk" patients.
Graph of survival prediction for patients split into different categories over five years ... Source: O'Regan et al.
The study, approved by a research ethics committee beforehand, was based on MRI scans from 256 NHS patients who gave their written consent for their medical records to be used. A third of the patients died.
"Applying machine learning" to data "obtained from cardiac MR imaging allows more accurate prediction of patient outcomes" in pulmonary hypertension cases, the study concluded.
"A supervised machine-learning survival model that includes three-dimensional cardiac motion provides incremental prognostic benefit when compared with conventional imaging and hemodynamic, functional, and clinical markers. Machine learning by using cardiac MR imaging should be evaluated as a tool to guide patient management."
Researchers plan to verify results by testing the software on patient data taken from a different hospital. One limitation in the design of the software is that people with heart disease can die from a related illness, thus skewing the training data given to the AI. The team said it is aware it needs to keep this in mind when teach the program.
Overall, the goal is to improve the software to make better predictions on patient survival, and also help doctors offer the best treatment plans for pulmonary hypertension and other heart conditions as soon as possible. ®