EUCARDIA: A web-based platform for the CVD prediction using ML techniques in the Greek population.
Statistical models used to estimate the cardiovascular disease (CVD) risk often present methodological constraints leading to overestimate or underestimate the total CVD risk. The aim of this study was to develop and implement a web-based Machine Learning (ML) platform to predict the personalized 10-year CVD risk for the Greek population.
The retrospective study included clinical and demographic data from 3,290 CVD-free participants. The CVD risk prediction model was based on two classifiers. The first was a binary classifier to estimate the occurrence of CVD, and the second was a multiclass classifier designed to replicate SCORE2 risk stratification categories. The selection of appropriate algorithms for integration into the platform was based on the evaluation metric ROC-AUC. To support the clinicians, the platform was integrated with scientific libraries to retrieve the most relevant literature based on the features that most influence the model decision-making.
The Voting Ensemble algorithm was selected for the binary classifier, achieving an AUC-ROC of 0.78. For the multiclass classifier, the selection algorithm was Stacking Ensemble, which yielded a 0.97 AUC-ROC. The comparison between ML and statistical model HellenicScore showed that the binary classifier was better in all metrics except accuracy in which HellenicSCORE had a higher value. The CVD risk prediction model and the integration with scientific libraries were successfully developed and deployed as a web-based platform.
The pilot run of the platform showed that it could be used as a reliable tool for CVD risk assessment, outperforming the traditional statistical models.
The retrospective study included clinical and demographic data from 3,290 CVD-free participants. The CVD risk prediction model was based on two classifiers. The first was a binary classifier to estimate the occurrence of CVD, and the second was a multiclass classifier designed to replicate SCORE2 risk stratification categories. The selection of appropriate algorithms for integration into the platform was based on the evaluation metric ROC-AUC. To support the clinicians, the platform was integrated with scientific libraries to retrieve the most relevant literature based on the features that most influence the model decision-making.
The Voting Ensemble algorithm was selected for the binary classifier, achieving an AUC-ROC of 0.78. For the multiclass classifier, the selection algorithm was Stacking Ensemble, which yielded a 0.97 AUC-ROC. The comparison between ML and statistical model HellenicScore showed that the binary classifier was better in all metrics except accuracy in which HellenicSCORE had a higher value. The CVD risk prediction model and the integration with scientific libraries were successfully developed and deployed as a web-based platform.
The pilot run of the platform showed that it could be used as a reliable tool for CVD risk assessment, outperforming the traditional statistical models.
Authors
Ploussi Ploussi, Petrou Petrou, Manginas Manginas, Demetrakopoulos Demetrakopoulos, Panagiotakos Panagiotakos, Adaloglou Adaloglou, Tokanis Tokanis, Kyritsis Kyritsis, Kollias Kollias, Tournis Tournis, Siasos Siasos, Efstathopoulos Efstathopoulos
View on Pubmed