Re-anamnesis as a Mandatory Post-Imaging Diagnostic Node: Clinical Validation of the ISWB Sequential Bayesian Model in Three Neuroradiology Cases with Posterior Gains of 0.84–0.92 and EU AI Act Art. 14 Compliance Mapping
Background: Current AI-assisted radiology workflows lack a formalized mechanism for mandatory post-imaging clinical re-evaluation. The ISWB (Ishikawa-Small- World-Bayesian) model addresses this gap by structuring the diagnostic pathway as a weighted directed acyclic graph G = (V, E, W) with ten sequential nodes (v1-v10). Methods: We applied the ISWB model to three retrospective neuroradiology cases: frontal haematoma (Case 1), meningioma (Case 2), and Denis I vertebral fracture (Case 3). For each case, we computed prior probability P0(H) from epidemiological prevalence and calculated posterior gains Delta-P at each diagnostic node using sequential Bayesian updating. Node v4 (re-anamnesis) was implemented as a mandatory post-imaging step requiring an Independent Clinical Review (ICR >= 90% SLA) before therapeutic decision. Results: Re-anamnesis (v4) produced posterior probability gains of Delta-P = 0.90 (Case 1), Delta-P = 0.84 (Case 2), and Delta-P = 0.92 (Case 3). In Case 3, a single targeted re-anamnesis question yielded 47% of the total information gain achievable by full MRI, demonstrating the high diagnostic efficiency of structured post-imaging clinical review. AI integration at node v3 (compatible with published architectures such as NVIDIA NV-Reason-CT) provided structural pattern recognition without replacing mandatory clinical judgment. Conclusions: The ISWB model provides a validated sequential framework for AI-assisted neuroradiology that formally satisfies EU AI Act Art. 14, paragraph 4(b) requirements for independent clinical judgment prior to acting on AI output. The mandatory re-anamnesis arc (v3->v4) represents a reproducible, quantifiable safety mechanism applicable to clinical AI deployment. Within the HIC/HITL/HOTL taxonomy, ISWB is structurally defined as a Human-in-the- Loop (HITL) architecture, with v4 as the mandatory gate preventing autonomous diagnostic closure.
Authors
- Francesco Matozza (ORCID: https://orcid.org/0000-0003-4014-0529)
- Ernesto Matozza
Institutions
- European Society of Radiology (AT)
- MNM Consulting (France) (FR)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24210049
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- preprint