AI IDs 5 kinds of heart failure to guide risk prediction and treatment

Heart failure affects many millions globally but can be caused by multiple factors, requiring different treatments. Now, researchers have trained multiple machine learning models using a large, population-based dataset to identify five subtypes of heart failure, which may better inform treatment, patient education, and the prediction of future risk factors.

‘Heart failure’ is an umbrella term used to describe when the heart doesn’t pump effectively enough to meet the body’s needs for blood and oxygen. It can be caused by several underlying factors affecting the treatment of the condition. Risk factors for heart failure include coronary artery disease and heart attacks, diabetes, high blood pressure, being overweight and obese and disease of the heart valves.

Traditionally, the different types of heart failure are classified according to a person’s left ventricular ejection fraction (LVEF), the amount of blood the heart’s left ventricle pushes out with each contraction. But a 2018 Swedish machine-learning study found that LVEF didn’t predict heart failure survival rates.

Now, researchers from University College London have used four machine learning models to develop a framework for determining heart failure subtypes that might better inform treatment and determine future risk.

The researchers looked at anonymized electronic health record data from more than 300,000 UK patients who’d been diagnosed with heart failure over a span of 20 years. The data were taken from two large primary care datasets representative of the UK population.

“We sought to improve how we classify heart failure, with the aim of better understanding the likely course of the disease and communicating this to patients,” said Amitava Banerjee, the study’s lead author. “Currently, how the disease progresses is hard to predict for individual patients. Some people will be stable for many years, while others get worse quickly.”

To avoid bias that might arise from using one machine learning model, the researchers used four models to separate heart failure cases into groups. After being trained using segments of the data, the models discerned five subtypes based on 87 of a possible 635 factors, including age, symptoms, the presence of other conditions, medications the patient was taking, health parameters like blood pressure, and test results such as kidney function. The subtypes were validated using a separate dataset.

The five subtypes were clustered according to specific characteristics. ‘Early onset’ included young people with a low rate of risk factors. ‘Late onset’ were older, female, prescribed few medications and had cardiovascular disease. ‘Atrial fibrillation-related’ included people with atrial fibrillation – a condition where the heart beats irregularly – or disease of the heart valves. The ‘Metabolic’ subtype included overweight people with a medium rate of risk factors but a low rate of cardiovascular disease. And ‘Cardiometabolic’ included overweight people on a high number of prescribed medications, with a high rate of risk factors and cardiovascular disease.

The researchers found that the risk of dying in the year following diagnosis differed between subtypes. At one year, the all-cause mortality risks were highest for those in the atrial fibrillation-related subgroup (61%), followed by late onset (46%), cardiometabolic (37%), early onset (20%) and metabolic (11%).

The researchers say the study’s findings can be used to improve the treatment of heart failure.

“Better distinctions between types of heart failure may also lead to more targeted treatments and may help us to think in a different way about potential therapies,” Banerjee said.

The researchers developed an app based on their machine-learning approach that doctors can use to determine what subtype a person falls into. It can be used to guide patient education and improve prediction of future risk.

“The next step is to see if this way of classifying heart failure can make a practical difference to patients – whether it improves predictions of risk and the quality of information clinicians provide, and whether it changes patients’ treatment,” said Banerjee. “We also need to know if it would be cost-effective. The app we have designed needs to be evaluated in a clinical trial or further research, but could help in routine care.”

The study was published in the journal The Lancet Digital Health.

Source: University College London