Cardiovascular disease (CVD) is a leading cause of death and disability, with aging being a major risk factor. Lifestyle factors exacerbate these risks. The Drosophila melanogaster model offers a powerful tool for studying CVD and aging due to conserved pathways between Drosophila and human hearts. Many genetic and environmental risk factors for human heart disease also impact Drosophila, making it a valuable model. While advanced imaging techniques provide high-resolution videos of Drosophila hearts, analysis using current methods, such as Semi-Automatic Optical Heart Analysis (SOHA) software, is time-consuming and requires manual intervention. This research aims to develop automated, machine learning-based methods for analyzing these high-resolution videos, enabling faster and more efficient assessment of cardiac function in Drosophila models.

Existing methods for analyzing Drosophila heart function, like SOHA, rely on manual selection of points of interest, limiting throughput. While some studies have used 3D convolutional architectures for segmentation of Drosophila heart images from optical coherence microscopy (OCM) or optical coherence tomography (OCT), similar techniques for standard high-resolution optical microscopy are lacking due to the complexity of heart morphology. Previous machine learning applications to Drosophila cardiac analysis have been limited in the range of parameters quantified and the rigor of the aging or disease models tested. This study addresses these gaps by developing an automated analysis platform for high-speed, high-resolution optical microscopy recordings.

The researchers developed three machine learning pipelines: (1) heart wall detection via semantic segmentation using a 2D attention-based UNet architecture; (2) age prediction using segmentation-calculated cardiac statistics and logistic classification; and (3) age prediction directly from raw video frames using a convolutional neural network. The UNet architecture segmented heart walls frame-by-frame, enabling calculation of cardiac parameters like diastolic diameter, systolic diameter, fractional shortening, heart rate, and arrhythmia index. For age prediction, both cardiac statistics and raw video data were used as input to separate models. The study used a training set of 54 hearts and a validation set of 177 hearts. The OGDH knockdown model for DCM analysis used a similar deep-learning approach, comparing cardiac parameters in control and OGDH knockdown Drosophila.

The deep learning model accurately segmented Drosophila heart walls, allowing for the calculation of a comprehensive suite of cardiac parameters. The model successfully captured age-related changes in these parameters, showing a significant decrease in fractional shortening (cardiac function) and an increase in arrhythmia index with age in both male and female flies. There was a high degree of agreement between the deep learning results and those obtained using SOHA software. The machine learning models achieved high accuracy in predicting fly age based on both cardiac statistics (79.1% accuracy, AUC 0.87) and raw video data (83.3% accuracy, AUC 0.90). The application of the deep-learning approach to a Drosophila DCM model, created by knocking down OGDH, revealed significant cardiac dilation and reduced cardiac performance, consistent with features of DCM. This confirmed the applicability of the method beyond aging models.

This study demonstrates a significant advancement in the automated analysis of Drosophila cardiac function. The deep learning approach overcomes the limitations of manual analysis methods, providing a more efficient and accurate way to assess cardiac dynamics. The high accuracy of age prediction using both cardiac parameters and video data underscores the physiological relevance of the extracted features. The successful application of the method to a DCM model showcases its versatility and potential for studying a wide range of cardiovascular diseases in Drosophila. The findings are relevant to the field of cardiovascular research, providing a powerful new tool for studying cardiac function and disease in a model organism.

This research provides a novel, automated machine learning platform for analyzing Drosophila cardiac function. The high accuracy of the deep learning models in predicting fly age and in characterizing cardiac dysfunction in a DCM model demonstrates the method's value for studying cardiovascular disease. The platform offers a significant advancement in Drosophila cardiac research, accelerating the process of data analysis and facilitating broader applications in the study of aging and disease.

The study's limitations include the relatively small size of the labeled dataset for training the segmentation model, which might affect the generalizability of the results. Further validation with larger and more diverse datasets is needed. The current age prediction models are limited to predicting ages within the 1-week and 5-week ranges. Future work could address these limitations by expanding the dataset and exploring more advanced model architectures.