Automated evaluation of cardiac contractile dynamics and aging prediction using machine learning in a Drosophila model

Abstract

The Drosophila model has proven tremendously powerful for understanding pathophysiological bases of several human disorders including aging and cardiovascular disease. Relevant high-speed imaging and high-throughput lab assays generate large volumes of high-resolution videos, necessitating next-generation methods for rapid analysis. We present a platform for deep learning-assisted segmentation applied to optical microscopy of Drosophila hearts and the first to quantify cardiac physiological parameters during aging. An experimental test dataset is used to validate a Drosophila aging model. We then use two novel methods to predict fly aging: deep-learning video classification and machine-learning classification via cardiac parameters. Both models suggest excellent performance, with an accuracy of 83.3% (AUC 0.90) and 77.1% (AUC 0.85), respectively. Furthermore, we report beat-level dynamics for predicting the prevalence of cardiac arrhythmia. The presented approaches can expedite future cardiac assays for modeling human diseases in Drosophila and can be extended to numerous animal/human cardiac assays under multiple conditions.

Keywords: Drosophila heart model; Machine learning; aging prediction; cardiovascular disease; medical segmentation.

Figure 1:. Proposed deep learning pipeline.

One pipeline enables frame-level segmentation of heart walls, while two alternate pipelines are used for age classification. For the segmentation task, users upload videos for analysis, perform segmentation using a trained neural network, select validated parameters, and extract beat-level parameters. The middle pipelines enable age-prediction of hearts via calculated cardiac statistics, further enabling transparent modeling of physiological parameters. Lastly, the bottom-most pipeline details an age-prediction task using a neural network, thus providing an alternate method for age-prediction.

Figure 2:. Representative output of neural network.

(a) Select frames extracted from cardiac video, with overlaid mask from neural network denoting heart walls. (b, c) Using annotated frames, an annotated mechanical mode (M-Mode) image is generated. (d, e) Detected heart masks are used to calculate average stationary diameter on a per frame resolution. (b, d, f) represent generated M-Mode, time-series beating pattern, and calculated cardiac parameters for a representative 1w male heart. (c, e, g) represent generated M-Mode, time-series beating pattern, and calculated cardiac parameters for a representative 5w male heart. Green (start of DI) and red (end of DI) lines are annotated in (b, c) and as gray periods in (d, e). Times of max contraction (red) and max relaxation (green) is annotated by x indicators in (d, e).

Figure 3:. Deep learning recovers aging phenotypes in Drosophila cardiac model.

(a) Diastolic diameter, in micron, calculated by our model (left), SOHA (right), and agreement between two datasets (right). (b) Systolic diameter, in micron, calculated by our model (left), SOHA (right), and agreement between two datasets (right). (c) Radial contractility (fractional shortening), in percentage, calculated by our model (left), SOHA (right), and agreement between two datasets (right). (d) Heart rate, in Hertz, calculated by our model (left), SOHA (right), and agreement between two datasets (right). (e) Beating dysrhythmia (arrythmia index) calculated by our model (left), SOHA (right), and agreement between two datasets (right). All error bars report ± SEM. Age-dependent statistics compared with paired t-Test statistics. Statistics are calculated via the use of a restricted ROI, selected by a trained user.

Figure 4:. Contractile dynamics detected by deep learning.

(a) Stroke volume for male and female aging groups is depicted, in picolitres. A significant reduction in stroke volume is exhibited with aging in both genders. (b) Cardiac output visualized via integration of per-beat stroke volumes and normalized by beat-times, measured in pL s-1. Aging depicts a strong reduction in cardiac output. (c) Time to peak contraction (negative) heart-wall velocity visualized for aging. (d) Time between peak contraction (negative) and relaxation (positive) velocities visualized for aging. Both (c) and (d) reflect no significant change with aging. All error bars report ± SEM. Age-dependent statistics compared with paired t-Test statistics.

Figure 5:

Machine learning classification of aging data. Logistic classification models enable accurate prediction of fly age from model-calculated cardiac statistics. (a) Heatmap of test-dataset predictions, indicating high agreement. (b) Average ROC-curve of all folds from presented model, reporting a mean AUROC of 0.85. (c) Physiological feature relevancies for predicting fly age, as determined by SHAP study. SHAP study suggests large dependence on fractional shortening (FS) in aging.

Figure 6:. Deep learning classification of aging videos (n = 497).

Deep learning models enable accurate prediction of fly age from only video data. (a) Heatmap of test-dataset predictions, indicating high agreement. (b) Average ROCcurve of all folds from presented model, reporting a mean test accuracy of 83.3% and mean AUROC of 0.900. (c) Model class log-likelihood distribution for investigated test predictions.