External validation of machine learning models predicting global longitudinal strain from conventional echocardiography in patients with cancer

Global longitudinal strain (GLS) is a well-established prognostic marker for the early detection of cancer therapy-related cardiac dysfunction (CTRCD). We previously developed machine learning (ML) models to predict reduced GLS (Low-GLS, defined as absolute GLS < 16%) from conventional echocardiographic parameters in patients with cancer. This study aimed to externally validate the performance and generalizability of the previously developed ML models in an independent cohort. This multicenter study included patients from the Tokyo Metropolitan Tama Medical Center (TMC, n = 1484) and Shizuoka Cancer Center (SCC, n = 141) who underwent echocardiography with GLS measurements before or after anticancer chemotherapy. The exclusion criterion was left ventricular ejection fraction < 50%. Low-GLS was predicted using 25 conventional echocardiographic parameters. The previously developed ML models (Random Forest, Extra Trees, and CatBoost) were directly applied, without retraining, to an independent external validation cohort from SCC. A conventional logistic regression model was included as a reference for comparison. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), and other metrics. In the internal validation cohort, the ML models demonstrated discriminative comparable performance (AUCs 0.722–0.748), whereas logistic regression achieved an AUC of 0.729. In the external validation cohort, performance was generally preserved; the Random Forest model demonstrated the highest discriminative ability (AUC 0.772), while the remaining models achieved AUCs of 0.725–0.746. Logistic regression achieved an AUC of 0.738. ML models trained on conventional echocardiographic parameters retained moderate predictive ability for reduced GLS upon external validation.