AI Strain Monitoring in Breast Cancer

> Research explainer: This briefing examines verified primary research published 74 days before the briefing date. It is not a same-day research update and does not provide medical advice.

Evidence

This Research explainer examines a prospective observational study of consecutive breast cancer patients scheduled for anthracyclines and/or HER2-targeted therapies. Participants underwent transthoracic echocardiography before chemotherapy and follow-up studies approximately every 12 weeks for at least three examinations. Two experts measured global longitudinal strain (GLS) manually, and a fully automated AI-based system measured GLS automatically; the investigators compared the resulting GLS measurements and GLS-based cancer therapy-related cardiac dysfunction (CTRCD) detection. The analysis included 92 patients and 456 echocardiographic studies. pmid:42388419

The study defined GLS-based CTRCD as a relative GLS reduction greater than 15% from baseline. Expert measurements identified GLS-based CTRCD in 29 of 92 patients (31.5%), while AI-derived measurements identified it in 32 of 92 patients (34.8%); the reported comparison had P = 0.58. Agreement for the presence or absence of GLS-based CTRCD was reported as 85.9%, with Cohen’s kappa = 0.68 (95% CI 0.60–0.76; P < 0.001). pmid:42388419

Analysis — Measurement Agreement

The central finding is agreement in classification and longitudinal monitoring, rather than evidence that AI improves patient outcomes. The AI system produced lower absolute GLS values than experts, 17.7 ± 2.9% versus 18.4 ± 2.8% (P = 0.007), and the investigators described this as systematic underestimation. Yet they reported no statistically significant difference in longitudinal GLS changes between methods (P = 0.72). This combination matters because serial relative change and a fixed classification threshold can behave differently when two measurement approaches have a systematic offset. pmid:42388419

The detection-rate comparison does not establish equivalence, interchangeability in every setting, or clinical benefit. It reports a nonsignificant difference between the methods for GLS-based CTRCD detection, alongside substantial kappa agreement. The source also reported no significant difference in detection timing, using a Wilcoxon signed-rank test: median detection time was 9.1 months (IQR 6.0–12.3 months) for expert assessments and 9.0 months (IQR 5.9–12.0 months) for AI assessments (P = 0.47). Statistical significance for agreement should not be read as clinical significance, particularly where values may sit near a threshold. pmid:42388419

For oncology and cardio-oncology research readers, the study supports evaluating automated GLS as a workflow-oriented measurement approach that can track temporal relative changes and identify GLS-based CTRCD similarly to expert analysis in this cohort. It does not test whether AI-guided monitoring changes chemotherapy management, reduces cardiac events, or improves survival. The investigators specifically noted that expert review and integration with symptoms, biomarkers, and other imaging data remain essential when values fluctuate abruptly or decisions have major implications, such as modifying chemotherapy regimens. pmid:42388419

Limitations

Interpretation is constrained by the single-centre design, modest sample size, and use of specific vendors, all of which may limit generalizability. The study was designed primarily to assess measurement agreement and longitudinal tracking rather than small differences in CTRCD incidence or detection timing, and the authors stated that it may have been underpowered for very small differences in detection rates. Manual expert analysis served as the comparison standard but has variability of its own. pmid:42388419

The source also reports no formal pre-specified qualitative adjudication of all discordant cases. Automated GLS measurement can be unsuccessful with suboptimal apical views and/or insufficient endocardial visualization, which is a practical implementation limitation. Serum troponin and other biomarkers were not incorporated into CTRCD diagnosis in this study. These constraints leave uncertainty about performance across centres, vendors, image-quality conditions, and clinical workflows beyond those studied. pmid:42388419

Evidence boundary

This one-source briefing is limited to what the cited study reports. It does not establish independent confirmation, broader clinical effectiveness, or patient-specific guidance. The design, population, measurements, and follow-up described in that source define the evidence boundary. This summary provides research context and is not medical advice. The evidence should be read as a bounded report of the study rather than as a conclusion about other populations, settings, interventions, or outcomes. Any possible connection to disease mechanisms remains limited to the measurements and interpretations documented by the cited authors. Terms describing associations, responses, or biological patterns retain the meaning and uncertainty given in that source.

No inference beyond the cited source is made here.