Enhancement of Risk Prediction With Machine Learning

Abstract

During the past decade, there has been a significant rise in the development of risk prediction models in cardiovascular disease. These models ostensibly offer clinicians a way of assigning risk to individual patients to allow for precision management decisions and risk-adjusted reporting of performance measures. 1 An important performance measure for percutaneous coronary intervention (PCI) is postprocedure bleeding, which is not only the most common noncardiac complication after PCI but is also associated with increased mortality, morbidity, and cost. 2 Rao et al 3 developed a 31-variable risk prediction model, developed from data from the National Cardiovascular Data Registry (NCDR), that demonstrated reasonable model performance with a C statistic of 0.77. This model, which has also been converted to a smartphone app, can facilitate decision-making regarding bleeding avoidance strategies and establishes a framework to allow for risk-adjusted provider feedback reports. Although the model has been implemented in the NCDR CathPCI registry, there are some important limitations. First, it can only incorporate covariates recorded in the registry and thus may miss some important predictors of bleeding, such as frailty. Second, some continuous covariates were transformed into and included as dichotomous variables, and third, backward stepwise selection was used to determine the final predictors, which could have eliminated important predictors. Mortazavi et al 4 examined whether machine learning methods can be used to enhance the NCDR risk prediction model of post-PCI major bleeding. 3 Machine learning methods did not demonstrate improvement in the performance of the model when some of the factors were included as dichotomous variables. 4 However, by not constraining the model to dichotomous variables and implementing machine learning algorithms, Mortazavi et al 4 were able to demonstrate an improvement in the discrimination of the model to a C statistic of 0.81, with reclassification of 5560 bleeding cases and 31 479 nonbleeding cases. These findings have important implications if machine learning algorithms can be applied to a wide range of existing risk prediction models to improve performance. In addition to the analysis by Mortazavi et al, 4 there are other examples of machine learning algorithms improving risk model performance. By adding variables and using machine learning algorithms, Huang et al 5 achieved a better area under the receiver operating characteristic curve as well as a better calibration slope for the NCDR post-PCI acute kidney injury risk prediction model. The enhanced risk prediction model reclassified nearly 15% of patients whose risks were underestimated and more than 21% of patients whose risks were overestimated. 5 A limitation of the analysis by Huang et al 5 is that the improved risk model included variables that may not be practically available at the time of PCI, such as pre-PCI left ventricular ejection fraction. Additionally, Kakadiaris et al 6 published the results of using machine ...

Multimodal deep learning for COVID-19 prognosis prediction in the emergency department: a bi-centric study

“…Machine learning (ML) enabled an effective prediction of outcomes that are not easily comprehended by other risk-predicting tools 7 , 8 . For example, XGBoost and RegCox algorithms have outperformed the widely used HAS-BLED clinical score in predicting the risk of bleeding 9 .…”

Multimodal deep learning for COVID-19 prognosis prediction in the emergency department: a bi-centric study.

“…Machine learning (ML) enabled an effective prediction of outcomes that are not easily grasped by other risk predicting tools 8, 9 . For example, XGBoost and RegCox algorithms were found to outperform the widely used HAS-BLED clinical score in predicitng the risk of bleeding 10 .…”