Transformer-Based Health Record Summarisation with Auditable Source Attribution: A Reproducibility-Centred Evaluation

Abstract

Clinical summarisation systems can sound convincing while omitting decisive events or attributing facts to the wrong encounter. To address this, the study will measure whether source-linked transformer summaries reduce unverifiable statements without sacrificing important clinical detail. The research design centres on a blinded assessment of de-identified clinical narratives paired with evidence passages and structured omission/error rubrics. The choice of evaluation units and exclusion criteria is informed by the broader methodological record [3–7]. The analysis will contrast ordinary abstractive summaries, retrieval-supported summaries and provenance-constrained summaries. The reproducibility extension records dataset versions, model settings, exclusion rules and provenance before any held-out evaluation. Enterprise-storage sources are used to specify constraints on data availability and provenance [1], with complementary infrastructure evidence defining practical recovery safeguards [2]. These citations are not treated as clinical effectiveness evidence. Assessment will focus on claim-level factual accuracy, critical omission rate, evidence traceability and clinician review time. The supporting evidence base motivates the chosen comparator conditions and robustness criteria [8–12]. Critical sensitivity checks address ambiguous documentation, conflicting observations, retrieval gaps and variations in note length. Further systems and modelling literature supplies context for the proposed sensitivity analyses [13–17]. An independent replication exercise would repeat the main analysis from the documented artefacts and quantify disagreements. The expected output is a reusable evaluation framework with transparent uncertainty and reporting rules, informed by wider reproducibility literature [18–23].