Pointwise Reliability of Machine Learning Models: Application to Cardiovascular Risk Assessment

Machine learning has made significant advances in many areas, particularly in the healthcare domain. However, despite the advances, the implementation of these models in clinical scenarios is still limited due to several challenges, including the lack of trust. Standard performance measures, such as sensitivity, specificity and confidence intervals can be used to evaluate the reliability of a model, but these are overall performance metrics and do not provide insight into the performance of individual instances. Moreover, these estimates are typically calculated during the training phase and are not easily generalized to new, unseen instances, occurring in the deployment phase. As result, besides the prediction outcome, the existence of a measure of reliability in the prediction of individual estimations would add a layer of security, increasing trust in human-AI interaction, as well as it could also be helpful to support the improvement of the model. This study proposes a reliability measure, combining density and local fit principles, to estimate the confidence of individual predictions in the deployment phase. When applied to a machine learning model in the cardiovascular risk assessment context, the method demonstrates the ability to distinguish between reliable and unreliable predictions, as well as aiding in the stratification of the population.