An efficient hybrid approach for classifying cardiovascular disease risks DenseNet and MobileNet

The existing machine learning models on CVD prediction suffer in aspects such as poor generalization regarding heterogeneous data like clinical, demographic, and biochemical indicators, cannot support automatic learning of multilevel-feature interactions, and possess bias which reduce accuracy. The existing deep learning models suffer with unimodal architecture approaches, overfitting in case of small datasets, and limited robustness across diverse datasets. There are risk factors considered such as blood pressure, cholesterol levels, diabetes, chest pain, fatigue, irregular heartbeats, etc. from the primary dataset such as Kaggle CVD dataset (Images and clinical data). To overcome these, a hybrid architecture with CNN, DenseNet + MobileNet is needed with data preprocessing which removes noise and inconsistency, followed by parallel learning using DenseNet and MobileNet. In these, DenseNet process hierarchical structures, address underfitting whereas MobileNet adds lightweight, mitigates overfitting and heavy workloads, and together support fusion, ensures 99.25% accuracy and lowest loss 0.02 when compared against the other models. The fused representations are then passed to classification, for severity prediction. From the results evaluated, it is observed that the proposed system outperforms than other approaches considered.