A dual-modality computational pipeline integrating machine learning and deep neural networks for modelling Drosophila cardiac ageing and disease

High-speed optical microscopy of Drosophila hearts has become an essential approach for modeling cardiovascular disease and aging. However, manual evaluation of cardiac function is time-consuming, subjective, and error-prone, highlighting the need for reliable computational frameworks. Here, we present an end-to-end dual-modality pipeline that combines segmentation-derived cardiac parameters with ensemble machine learning and raw video-based representations with deep convolutional neural networks. This integrated approach enables automated and quantitative phenotyping of cardiac function across both aging and cardiovascular disease contexts. We evaluated the performance of the pipeline across three biological problems relevant to cardiovascular systems biology: distinguishing between young (1-week) and aged (5-week) hearts, classifying healthy hearts from those carrying the lamin C mutation (LamC ^G489V ), and identifying phenotypes associated with Ogdh knockdown. Ensemble learning with Random Forest, XGBoost, and Logistic Regression applied to segmentation-derived features achieved accuracies of 82.3% ± 0.6% for aging classification, 79.0% ± 0.4% for lamin C mutant detection, and 98.5% for Ogdh knockdown identification. In parallel, video-based modelling using a custom convolutional neural network and a residual 3D network (R3D-18) yielded accuracies of 89.1% ± 0.5%, 74.0% ± 1.2%, and 93.9%, respectively. By integrating classical feature-based learning with modern deep architectures, this work provides a scalable computational framework for modelling Drosophila cardiac aging and disease. More broadly, the study demonstrates how multimodal machine learning can enhance quantitative phenotyping, minimize human error in analysis, and expedite the application of invertebrate models in systems-level investigations of cardiovascular health and aging.