Machine learning predictors of post-radiotherapy cardiotoxicity in a Latin American breast cancer cohort: Clinical risk factors outperform cardiac dosimetry.

@article{aguirrebravo2026machine,
  title = {{Machine learning predictors of post-radiotherapy cardiotoxicity in a Latin American breast cancer cohort: Clinical risk factors outperform cardiac dosimetry.}},
  author = {Víctor Aguirre-Bravo and Rodrigo El Far and María Herrera Ruiz and Sophia Gálvez Santelices and Pablo Munoz-Schuffenegger and Tomás Merino and Sergio Diez Domingo and Marcelo E Andia},
  journal = {Journal of applied clinical medical physics},
  year = {2026},
  month = oct,
  doi = {10.1002/acm2.70862},
  url = {https://europepmc.org/article/MED/42876042},
  note = {CC BY},
}

TY  - JOUR
TI  - Machine learning predictors of post-radiotherapy cardiotoxicity in a Latin American breast cancer cohort: Clinical risk factors outperform cardiac dosimetry.
AU  - Víctor Aguirre-Bravo
AU  - Rodrigo El Far
AU  - María Herrera Ruiz
AU  - Sophia Gálvez Santelices
AU  - Pablo Munoz-Schuffenegger
AU  - Tomás Merino
AU  - Sergio Diez Domingo
AU  - Marcelo E Andia
PY  - 2026
DA  - 2026/10/01
JO  - Journal of applied clinical medical physics
DO  - 10.1002/acm2.70862
UR  - https://europepmc.org/article/MED/42876042
AB  - Background  Most evidence on post-radiotherapy cardiotoxicity (CDT) derives from Global North cohorts, and the relative contribution of baseline clinical risk versus cardiac dosimetry remains poorly characterized in Latin American populations.Purpose  To examine, in a Chilean breast cancer cohort, whether baseline clinical risk factors or cardiac dose-volume metrics better predict post-RT cardiac events, using supervised and unsupervised learning.Methods  Retrospective cohort of 42 breast cancer patients treated with thoracic radiotherapy (RT, median heart dose 2.6 Gy; median follow-up 84.4 months). CDT was defined per CTCAE v5.0 ( ≥ Grade 1), excluding Grade 1 events coded under the non-specific term "Cardiac disorders - Other, specify", which on review were incidental imaging findings. Dose-volume histogram (DVH) parameters from seven cardiac substructures were extracted. Firth penalized logistic regression was used for prediction, principal component analysis (PCA) and kernel PCA for exploratory clustering, and bootstrap 0.632+ for optimism correction. The 2023 American Heart Association (AHA) PREVENT score was evaluated post hoc.Results  Ten patients (23.8%) developed CDT and 16 (38.1%) experienced cardiovascular events. No DVH differences were found between CDT+ and CDT- groups across any cardiac structure (all p>0.05 ), although the design is powered only for large effects. By contrast, IMRT delivered significantly higher cardiac doses than 3D-CRT across multiple substructures, without a differential CDT rate. Pre-existing cardiovascular disease was the strongest predictor (crude OR = 6.50, p=0.020 ; adjusted OR = 2.08, 95% CI 1.03-4.16, p=0.040 ). The median time to the first CTCAE-graded event was 48.4 months. The clinical-only model reached an area under the curve (AUC) of 0.783 versus 0.604 for dosimetry alone; the combined model corrected to 0.649 (95% CI 0.204-0.933). The AHA PREVENT score showed no significant risk gradient (trend p=0.203 ) but was markedly less susceptible to overfitting than DVH-only models (optimism 0.089 vs. 0.194).Conclusions  In this low-cardiac-dose cohort, baseline clinical variables outperformed cardiac dose-volume metrics in predicting post-RT cardiac events, although the design was not powered to exclude a dose-effect relationship at these exposures. Registry-based baseline risk assessment is therefore the more robust starting point in this setting, and IMRT breast RT planning warrants multi-substructure cardiac dosimetry monitoring.
SN  - 1526-9914
ER  -