Artificial intelligence (AI) has been in the headlines over recent months but its impact on healthcare is already being felt – with some studies finding impressive results for AI programs trained to spot skin cancer, diagnose brain tumours mid-operation, and interpret radiology images.
Another study we can add to the list is a recent paper from a team led by researchers at the University of Western Australia involving cardiovascular complications after non-cardiac surgery. In the study, researchers tested whether ‘neural networks’ or computer systems modelled on our human brain could be used to predict complications in people who had surgery.
The study included the information of 25,000 people who had undergone these non-cardiac surgeries and their data was captured as part of the Vascular Events in Non-Cardiac Surgery Patients Cohort Evaluation Study (VISION). All these people had their level of troponin T (a protein found in heart muscle which rises if there’s damage) measured postoperatively. Using this information, as well as routinely collected data such as a patient’s age, weight, ethnicity, and more in-depth measurements like intra-operative heart rate and blood pressure, researchers wanted to better understand whether these neural networks could predict whether someone would have complications. Each participant’s postop outcomes (complications, death) were also known to researchers, so they could test the accuracy of the model’s predictions.
The study found that the multi-layer neural network was good at predicting which people were likely to have heart problems or even die following surgery. It could predict postop myocardial injury with 70% accuracy and forecast which patients would die with 89% accuracy.
The researchers believe these findings prove neural networks can be used to predict post-operative non-cardiac surgery outcomes effectively. They say such a model might be used by a treating healthcare team to determine which patients are at the highest risk of negative outcomes and to allocate resources on that basis. They suggest the models may become more accurate when continually given data that are more relevant to particular areas or subgroups of people.