Bridging innovation and caution: A critical review of AI-Driven diagnostics for lumbar disc herniation

Background Artificial intelligence (AI) has demonstrated substantial potential in medical imaging, particularly in the detection and characterisation of lumbar spine pathology. In lumbar disc herniation diagnostics, numerous models report high performance under controlled conditions. However, the extent to which these systems are clinically ready for real-world deployment remains unclear. Objective This study aims to critically evaluate the current state of AI in lumbar disc herniation diagnostics through the lens of clinical readiness, with a focus on methodological rigor, clinical applicability, and regulatory alignment. Methods We conducted a critical narrative review of the literature on AI applications in lumbar spine MRI, complemented by a structured domain-based appraisal framework designed to improve transparency and consistency of evidence synthesis rather than to perform formal systematic review methodology. Main domains included dataset representativeness, ground-truth definition, validation strategies, calibration and uncertainty reporting, handling of heterogeneity, and alignment with clinical decision-making processes. Additional considerations included explainability, trustworthiness, synthetic data usage, and emerging regulatory frameworks such as the European Health Data Space (EHDS) and the EU AI Act. Results While many AI systems demonstrate high diagnostic performance, most are developed and evaluated using retrospective, single-centre datasets with limited external validation. Reporting of calibration, uncertainty, and subgroup performance remains inconsistent, and heterogeneity in imaging protocols and labelling strategies constrains comparability across studies. Explainability techniques are widely applied but often lack clinical validation, and their contribution to trustworthiness is limited when considered in isolation. Synthetic data approaches show promise in addressing data scarcity and privacy constraints but require rigorous evaluation of fidelity, bias, and clinical utility. Across studies, there is a consistent gap between technical performance and clinically meaningful outcomes, with limited evidence from prospective or real-world evaluations. Conclusions Despite rapid technological advances, current AI systems for lumbar disc herniation diagnostics remain insufficiently validated for routine clinical use. Achieving clinical readiness requires a shift from performance-centric evaluation toward comprehensive frameworks that integrate robust validation, uncertainty quantification, fairness, and regulatory compliance. Future research should prioritise prospective, multi-centre studies with clinically relevant endpoints and explicit consideration of human–AI interaction to enable safe and effective deployment in real-world settings. This review adopts a critical narrative approach supported by structured evidence mapping and domain-based appraisal to evaluate the clinical readiness of AI systems across methodological, clinical, ethical, and regulatory dimensions. These structured components are intended to strengthen transparency and critical interpretation while preserving the interpretative nature of a narrative review.