Key result
Machine learning models show moderate CTRCD prediction but remain limited by methodological flaws and poor validation.
Why the study?
Reliable baseline risk prediction models for cancer therapy-related cardiac dysfunction are needed to guide prevention, but the readiness of machine learning models in cardio-oncology is unexplored.
Are machine learning-based models reliable and methodologically robust for predicting baseline cancer therapy-related cardiac dysfunction risk?
Systematic Review (n=7)
Are machine learning-based models reliable and methodologically robust for predicting baseline cancer therapy-related cardiac dysfunction risk?
Effect estimate: AUC 0.65-0.88
Current machine learning models for predicting baseline CTRCD risk show moderate discrimination but are methodologically underdeveloped and require rigorous external validation before clinical deployment.
ML models for CTRCD risk prediction remain unsuitable for clinical use; underscores need for rigorous designs, calibration reporting, and external validation.
Background Cancer therapy-related cardiac dysfunction (CTRCD) is a prevalent complication with adverse clinical implications. Major guidelines have emphasised the need for reliable baseline risk prediction to guide prevention strategies. While machine learning (ML) has shown promise in cardiovascular medicine, its readiness for application in cardio-oncology remains unexplored. Aim To systematically review and critically appraise the performance, methodological quality, and clinical applicability of ML-based models for predicting baseline CTRCD risk. Methods A search on Embase and Medline was conducted up to April 2025. Eligible studies used at least one ML algorithm to estimate CTRCD risk prior to cancer treatment. Data were extracted using an adapted version of the CHARMS checklist. Risk of bias (ROB) and reporting quality were assessed using PROBAST and TRIPOD+AI respectively. Results Seven studies were included. Five developed novel models using multiple predictors, and two repurposed or validated existing AI-enhanced electrocardiography tools. While their levels of discrimination were at least moderate (AUC 0.65–0.88), calibration was not reported in any of the studies and only one study included a small external validation cohort. ECG-only models showed modest performance, whereas models integrating clinical and imaging predictors performed better but were methodologically constrained. Four studies scored high ROB, one low, and two unclear. Common limitations included small datasets with low number of events, limited transparency in reporting, and lack of reproducibility measures. Conclusions Current ML-based models for predicting CTRCD are underdeveloped and not ready for clinical deployment. The present findings underscore the need for more rigorous and standardised study designs, with greater emphasis on methodological transparency with clearly-defined adjudicated outcomes and external validation to improve reproducibility.
No takes yet. Share an insight, caveat, or question.
Andres et al. (2026) conducted a systematic review in Cancer therapy-related cardiac dysfunction (CTRCD) (n=7). Machine learning-based risk prediction models was evaluated on Model discrimination (AUC) (AUC 0.65-0.88). Machine learning models for predicting baseline CTRCD risk showed moderate discrimination (AUC 0.65-0.88) but are currently underdeveloped, methodologically constrained, and lack external validation.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: