Introduction

Patients with heart disease who require open heart surgery must be informed about the risks related to the surgical procedure. In cases where the risk is predicted to be particularly high either the patient may refuse, or the surgeon may decide not to undertake the operation. Currently, heart surgeons in the UK are quoting the risk using a calculation, the so-called 'EuroSCORE'. EuroSCORE has now been shown to be no longer accurate as it tends to overestimate the actual risk. This project aims to use machine learning for accurate risk prediction in patients undergoing open heart surgery.

Explaining the science

Machine learning approaches are increasingly used for prediction in health research and care as they have the potential to overcome limitations of regression models by including pairwise and higher order interactions and modelling nonlinear effects. Whilst calibration drift over time is well documented amongst logistic regression models for hospital mortality, the susceptibility of competing modelling methods to performance drift has not been well studied.

Project aims

The purpose of the present research is to develop a new more precise risk calculator using a large dataset of information routinely collected from all patients undergoing surgery in the UK. The research will apply the most advanced mathematical methods, so-called machine learning to identify patients who are likely to have a successful operation.

Applications

Big data and machine learning techniques have the potential to revolutionise the monitoring and improvement of cardiac surgical-outcome quality. The Society for Cardiothoracic Surgery (SCTS) recognises the urgent need to replace the EuroSCORE with a more precise risk prediction model and has endorsed the present project. Following a successful project, the SCTS will evaluate the opportunity to replace the EuroSCORE and implement the new score in all the cardiac units in the UK.

This project will support the SCTS in this process by delivering a software package in R that will allow users to adopt the model. The risk calculator will be built as a Shiny application. Accurate clinical risk-prediction will apply to three different objectives with relative beneficiaries:

  • To assess patient risk, which physicians and patients can then factor in to healthcare decisions including the choice between conventional vs. new trans-catheter therapeutic strategies.
  • To stratify risk for determination of inclusion criteria in clinical trials (researchers).
  • To assess and compare healthcare outcomes among providers (benchmarking).

Collaborators