Building a machine learning-assisted echocardiography prediction tool for children at risk for cancer therapy-related cardiomyopathy

Abstract

Background: Despite routine echocardiographic surveillance for childhood cancer survivors, the ability to predict cardiomyopathy risk in individual patients is limited. We explored the feasibility and optimal processes for machine learning-enhanced cardiomyopathy prediction in survivors using serial echocardiograms from five centers.

Methods: We designed a series of deep convolutional neural networks (DCNNs) for prediction of cardiomyopathy (shortening fraction ≤ 28% or ejection fraction ≤ 50% on two occasions) for at-risk survivors ≥ 1-year post initial cancer therapy. We built DCNNs with four subsets of echocardiographic data differing in timing relative to case (survivor who developed cardiomyopathy) index diagnosis and two input formats (montages) with differing image selections. We used holdout subsets in a 10-fold cross-validation framework and standard metrics to assess model performance (e.g., F1-score, area under the precision-recall curve [AUPRC]). Performance of the input formats was compared using a combined 5 × 2 cross-validation F-test.

Results: The dataset included 542 pairs of montages: 171 montage pairs from 45 cases at time of cardiomyopathy diagnosis or pre-diagnosis and 371 pairs from 70 at-risk survivors who didn't develop cardiomyopathy during follow-up (non-case). The DCNN trained to distinguish between non-case and time of cardiomyopathy diagnosis or pre-diagnosis case montages achieved an AUROC of 0.89 ± 0.02, AUPRC 0.83 ± 0.03, and F1-score: 0.76 ± 0.04. When limited to smaller subsets of case data (e.g., ≥ 1 or 2 years pre-diagnosis), performance worsened. Model input format did not impact performance accuracy across models.

Conclusions: This methodology is a promising first step toward development of a DCNN capable of accurately differentiating pre-diagnosis versus non-case echocardiograms to predict survivors more likely to develop cardiomyopathy.

Keywords: Cancer survivorship; Cardiomyopathy; Echocardiography; Machine learning.

Fig. 1

Image preprocessing and montages. a Illustration of an unprocessed image with sepia colorization and annotations. Superimposed arrows on ECG indicate where frames were extracted for montage Type 1 (red) and 2 (blue). b Processed image. c Montage Type I with frames corresponding to four points in a single cardiac cycle: 1 the R wave peak; 2 one-third of the R-R interval, approximating the end of the T wave; 3 one-half the R-R, approximating the middle of the TP segment, and 4 two-thirds the R-R interval, approximating the onset of the next P wave. d Montage Type 2 with numbered frames corresponding to four points in two consecutive cardiac cycles: 1 first R wave peak; 2 one-third of the first R-R interval, approximating the end of the first T wave; 3 second R wave peak; 4 one-third of the second R-R interval, approximating the end of the second T wave

Fig. 2

Example of echocardiographic data from non-case and case patients. Interval between echocardiograms for an at-risk non-case and a cardiomyopathy (case) patients displayed in months and index diagnosis shown for the case patient. Four different data groupings for case patient echocardiograms were used to construct the deep convolutional neural networks (DCNNs) based on timing of echocardiogram relative to diagnosis: at diagnosis and pre-diagnosis echocardiograms, pre-diagnosis echocardiograms, ≥ 1 year pre-diagnosis echocardiograms, and ≥ 2 years pre-diagnosis echocardiograms. For this sample case patient, no echocardiograms were performed within 1 year prior to the index diagnosis

Fig. 3

Performance for Type 1 montage convolutional neural networks. As the number and proportion of CM + echocardiograms in the dataset decreased, model performance worsened, with falling area under the precision-recall curve (AUPRC), F1-score, and positive predictive value (PPV). As the dataset imbalance grew, negative predictive value (NPV), area under the receiving operator curve (AUROC), and accuracy, which are less helpful in an imbalanced dataset as they reflect true negatives, changed marginally