Researchers compared the accuracy of statistical models developed by artificial intelligence with those developed by heart experts in predicting heart disease deaths.
When a doctor is planning treatment for a patient they need to consider the likely outcomes for that patient, or their “prognosis”. Statistical models can help doctors evaluate a particular patient’s risk level and prognosis and tailor the most appropriate treatment. Up to now, these prognostic models have been based on expert medical research to identify key risk factors in a particular illness. This process requires a lot of research effort and time. With the increasing use of electronic health records, there is now a huge amount of patient data available.
Researchers at the Francis Crick Institute, London, UK set out to use artificial intelligence (AI), to harness some of this data to develop a“machine-generated” model for predicting the risk of heart disease deaths. They compared the accuracy of the AI model with traditional prognostic models developed by heart experts. They published their results in PLOS ONE.
The researchers used data from 80,000 patients on the CALIBER platform which links four sources of routinely-collected electronic health records in England – primary health care records, hospital discharge records, a national audit of myocardial ischemia records, and death registrations. Using this data in a machine-learning approach, they developed an AI prognostic model for heart disease based on around 600 predictors of risk. They compared the accuracy of this AI model to a conventionally-derived model based on 27 risk factors selected by heart experts.
The AI derived model was better at predicting patient outcomes than the conventionally derived model. The AI model also identified novel predictors of risk – for example, the number of home visits by a doctor. This could be a useful indicator of how ill the patient is but had not been previously identified in conventional prognostic models.
This study demonstrated that it is possible to use machine learning to develop prognostic tools. AI models can be as accurate as conventionally-derived models and are less time-consuming to develop. This approach could be used to develop models for other diseases and has the potential to greatly influence how patient care is managed in the future.
Written by Julie McShane, Medical Writer
- Steele AJ, Denaxas SC, Shah AD, et al. Machine learning models in electronic health records can outperform conventional survival models for predicting patient mortality in coronary artery disease. PLOS One 13(8):e0202344. Doi:10.1371/journal.pone.0202344.
- AI beats doctors at predicting heart disease deaths. The Francis Crick Institute EurekAlerthttps://www.eurekalert.org/pub_releases/2018-09/tfci-abd090418.php