Deep learning estimation of LVEF and GLS from echocardiographic videos to determine cancer therapy-related cardiac dysfunction

Abstract

Cancer therapy-related cardiac dysfunction (CTRCD) surveillance relies on left ventricular ejection fraction (LVEF) and global longitudinal strain (GLS), measures that often require dedicated software and trained personnel. Video-based deep learning may reduce these barriers by estimating LVEF/GLS directly from routine echocardiographic cine loops; PanEcho is a pretrained model designed for this task. To evaluate whether PanEcho pretrained weights, fine-tuned on apical four-chamber (A4C) videos, can estimate LVEF and GLS and support exam-level CTRCD assessment in a longitudinal cardio-oncology cohort. We analysed 185 patients contributing 665 longitudinal echocardiographic A4C examinations with reference LVEF and GLS. Videos were standardised using an EchoNet-Dynamic-style preprocessing pipeline to match PanEcho's input domain. Patient-level splits comprised 130/28/27 patients (469/96/100 examinations) for training/validation/test. Starting from pretrained PanEcho weights, we fine-tuned models to predict LVEF and GLS; results are reported for the default configuration selected after exploring multiple parameter settings. Test-set performance (n=100 for LVEF and GLS) was summarised with MAE, RMSE, and Pearson r. Reference CTRCD labels were derived from measured baseline-to-follow-up LVEF/GLS changes per the 2021 ICOS consensus (aligned with 2022 ESC cardio-oncology guidance). CTRCD was inferred using measured baseline values with predicted follow-up values and evaluated in follow-up examinations with complete baseline measurements, reference labels, and corresponding predictions. On the test set, LVEF estimation achieved MAE 3.85%, RMSE 4.78%, and Pearson r=0.503 (Fig. 1); GLS estimation achieved MAE 1.89%, RMSE 2.43%, and Pearson r=0.417 (Fig. 2). For CTRCD, the inference was evaluable in 480 follow-up examinations (50 reference-positive) and achieved sensitivity 50.0%, specificity 91.4%, accuracy 87.1%, and ROC AUC 0.714. Fine-tuned PanEcho applied to EchoNet-Dynamic-style-preprocessed A4C videos provided clinically interpretable estimates of LVEF and GLS and supported feasible exam-level CTRCD assessment when measured baseline values were retained. To strengthen generalisability and longitudinal fidelity, future analyses should incorporate multi-view inputs, mitigate class imbalance, validate performance in independent external cohorts, and align follow-up using recorded study dates rather than exam-order suffixes.Fig 1.Test-set LVEF: reference vs predictedFig 2.Test-set GLS: reference vs predicted