Hierarchical Multi-modality Prediction Model to Assess Obesity-Related Remodelling
Résumé
The diagnosis of cardiovascular illnesses uses multiple modalities in order to obtain a complete and as robust as possible assessment of the heart. However, when addressing distinct pathologies, not all information might be needed in order to achieve a confident-enough diagnosis.
We propose a probabilistic machine learning method to identify the patients for which the acquisition of more complex data would be useful. We hypothesise that there exists a hierarchical relationship between modalities: echocardiography is more accessible and has a lower economical cost than other modalities (like magnetic resonance imaging (MRI)). The framework consists of two classifier models, each predicting the illness from the echocardiographic and MRI views, and a sample-weighting model that combines both predictions. This weighting model is used to decide which individuals will not need an MRI acquisition additional to the echocardiographic examination.
We illustrated this on a dataset of asymptomatic individuals with an echocardiographic study (N = 480), a subset of those also includes a MRI (N = 159). We analyse the effect of being overweight on cardiac geometry. We identified that the type of remodelling depended on blood pressure: overweight combined with high blood pressure resulted in an increase of ventricular mass, while only size changes were preserved for low-pressure individuals. With our method, we established that boundary cases of the former group could be correctly classified after incorporating MRI, while it was not the case for the latter.
Origine : Fichiers produits par l'(les) auteur(s)