JMIR Publications

ABSTRACT

Background:

Blockchain-supported verification can establish that a clinical record has diverged from a previously committed state, but the resulting alert does not explain which fields changed, how they changed, or their potential relevance. Additional analysis is therefore required to transform integrity evidence into interpretable information.

Objective:

This study aimed to transform blockchain-supported integrity alerts over versioned openEHR records into interpretable and prioritised information by combining deterministic analysis with locally hosted LLM assistance for alert interpretation and free-text categorisation.

Methods:

All 369 alerts registered between 1 September 2025 and 28 February 2026 for diagnostic procedure requests at a Portuguese local health unit were analysed. These comprised 306 integrity alerts, generated when the record changed without a corresponding version increment, and 63 log alerts. The blockchain-committed and current database states were decoded and compared deterministically. Integrity alerts were assigned a concern level according to the highest-level changed field using a four-level scheme based on the Critical Data Elements concept. A locally hosted llama3.1 model operated only on validated information to interpret detected changes and categorise clinician-authored free-text justifications using a closed nine-category taxonomy.

Results:

Accounting for the asymmetric structure of the committed and current states reduced the apparent number of changes among the 306 integrity alerts from 2,179 to 625 and prevented 922 fields from being incorrectly represented as cleared. The concern-level scheme classified 3 alerts (1.0%) as Critical and 303 (99.0%) as High. Among the High-level alerts, 232 contained only a date in the non-conformity note, 5 contained no associated note, and 66 contained free clinical text. All 66 received a predefined category label, with manual inspection identifying one remaining misclassification after refinement. Because discrepancy detection, field-level comparison, and concern-level classification remained deterministic, LLM-generated interpretations could be checked against the corresponding field-level results.

Conclusions:

Blockchain-supported integrity alerts require additional processing before their underlying changes can be meaningfully interpreted. In the proposed framework, deterministic processing establishes the discrepancy, identifies the changes, and supports concern-level prioritisation, while the locally hosted LLM assists with interpretation and free-text categorisation. This separation preserves auditable integrity evidence while using language-model capabilities only after the underlying changes have been established.