Beyond the black box: interpretability, accountability, and responsible clinical integration of AI-driven heart rate variability models—a narrative review

Background Heart rate variability (HRV) is a widely used digital biomarker reflecting autonomic regulation and has been associated with diverse cardiovascular, critical care, and stress-related outcomes. In parallel, AI and machine-learning methods have expanded rapidly in HRV-based prediction, often achieving strong predictive performance. However, clinical translation remains constrained by limited interpretability, unclear accountability, and challenges in workflow integration, particularly for black-box models used in high-stakes settings. Methods This narrative, concept-driven review examined conceptual, methodological, clinical, and governance dimensions of interpretability in AI-driven HRV prediction. A structured literature search was performed in PubMed/MEDLINE, Scopus, and Embase databases. Results The analysis showed that interpretability is not a binary property but varies by model design and deployment context. Post-hoc explainability methods may increase transparency, yet they can also be unstable, incomplete, or misleading, with potential to increase automation bias. Clinical adoption is further limited by signal-quality variability (ECG vs. wearable PPG), insufficient external validation, workflow misalignment, and unclear medico-legal responsibility. A four-step pragmatic implementation framework is proposed: data governance and signal integrity; robust model development and validation; workflow-compatible clinical integration with human oversight; and continuous post-deployment monitoring and governance. Conclusion HRV-AI systems should be treated as socio-technical interventions. Responsible adoption requires proportional transparency, explicit accountability structures, and lifecycle governance beyond predictive accuracy alone.