Predicting heart failure with data

What if an algorithm could flag a patient’s potential to suffer heart failure based on routine information found in medical records, such as blood pressure or cholesterol readings? 

A machine learning algorithm that can predict the potential for a patient to experience heart failure has been developed by Toronto Metropolitan University (TMU) professor Aziz Guergachi of the Ted Rogers School of Information Technology Management and an international team of collaborators. This research builds on professor Guergachi’s extensive work using machine learning and medical records to predict the likelihood of diseases such as Type 2 diabetes. 

Professor Guergachi and his team set out to create an algorithm that would be practical and applicable to a broad population using routinely collected data to enable early identification of patients at risk of heart failure in primary care settings. The algorithm determines the heart failure prediction by comparing nine of the patient’s routine biomarkers, such as body mass index, blood pressure, fasting glucose or total cholesterol, taken over time and entered into their medical records. 

Other attempts at training models to predict heart failure have been based on biomarkers from tests that are less common, he explains. “We want to be able to scale up quickly, and that’s one of the reasons we need to build up on what’s already available in the electronic medical records,” he said. 

Training the algorithm 

The team developed their machine learning algorithm using inclination analysis, training it on a dataset of almost 700 patients. Inclination analysis treats the system being analyzed – in this case, the heart – as a black box and focuses on the trends shown in the data points collected over time – for this algorithm’s purposes, the routine biomarkers – to make predictions. This study represents a novel application of inclination analysis to heart failure prediction. The team’s algorithm achieved an accuracy of 89 per cent in predicting a patient’s potential for heart failure. 

Professor Guergachi notes one of the main challenges in using these biomarkers for this purpose is navigating the variations in the data collection intervals. For instance, sometimes this data is collected over the long term, such as for annual physicals, or several times over a short term, such as during a visit to the emergency room. The researchers were able to address these intervals in the algorithm.  

Next steps 

Professor Guergachi is discussing the potential to commercialize and deploy these disease prediction algorithms in real-world settings with industry partners. He and his colleagues plan to expand their ongoing machine-learning-powered prediction research to other types of disease, and he has filed a provisional patent for a different disease prediction algorithm.

This work builds upon his ongoing collaboration with Dr. Karim Keshavjee, a medical doctor and research scientist from the University of Toronto, on disease prediction through algorithms. To develop the heart failure prediction algorithm, the pair also collaborated with professor Alessia Paglialonga of the Institute of Electronics, Computer and Telecommunication Engineering at the National Research Council of Italy, graduate students Federica Guida and Marta Lenatti, and postdoctoral fellow Alireza Khatami. 

This study represents a novel application of inclination analysis to heart failure prediction. The team’s algorithm achieved an accuracy of 89 per cent in predicting a patient’s potential for heart failure.

Read more about the team’s heart failure prediction findings in “Characterization of Inclination Analysis for Predicting Onset of Heart Failure from Primary Care Electronic Medical Records” (external link)  in the journal Sensors. 

This research was supported by the Natural Sciences and Engineering Research Council of Canada and the Canada Foundation for Innovation.