Reconciliation UI

A visual comparison interface that presents the source of truth (e.g., the existing EHR record) directly adjacent to the proposed changes (e.g., the AI-extracted list). This layout highlights specific textual, structural, or coding differences to accelerate validation.

• Inline highlighting: Additions, deletions, and modifications are color-coded (e.g., green for additions, red for deletions, yellow for modifications).
• Field-level alignment: Each medication attribute (name, dose, frequency, route) is vertically aligned to enable rapid scanning.
• Context preservation: The original source document snippet is displayed alongside the structured data to provide evidence without leaving the interface.

The interface visually encodes the model's calibrated probability for each extracted field, allowing reviewers to prioritize their attention on high-uncertainty items. This directly combats alert fatigue by preventing cognitive desensitization to low-value notifications.

• Color-coded badges: High-confidence fields (>95%) are marked green and may be collapsed; low-confidence fields (<70%) are marked red and expanded by default.
• Uncertainty indicators: Specific tokens or spans that contributed to low confidence are underlined with a wavy pattern.
• Sort and filter controls: Reviewers can sort the reconciliation queue by aggregate uncertainty score or filter to show only discrepancies.

For each discrepancy, the UI provides discrete, unambiguous action buttons that allow the reviewer to resolve the conflict with a single click or keystroke, minimizing cognitive load and review burden.

• Accept AI: Overwrites the existing record with the AI-proposed value.
• Reject AI: Retains the existing record value and discards the AI suggestion.
• Manual Override: Opens an inline editor for the reviewer to type a corrected value.
• Defer/Escalate: Flags the item for a senior reviewer via an adjudication workflow.
• Keyboard shortcuts: Each action is mapped to a single key (e.g., 'A' for Accept, 'R' for Reject) to enable high-throughput review.

Every AI-generated suggestion is directly linked to the exact sentence, paragraph, or table cell in the original unstructured source document. This source attribution enables rapid evidence verification and builds reviewer trust.

• Click-to-navigate: Clicking on a proposed medication navigates the source viewer to the originating text, which is temporarily highlighted.
• Multi-source traceability: If a conclusion is derived from multiple documents (e.g., a discharge summary and a specialist note), all sources are listed with their contribution weight.
• Copy citation: Reviewers can copy a formatted citation of the source evidence for audit documentation.

When a reviewer corrects a specific error—such as fixing the spelling of a drug name or adjusting a span offset—the system automatically applies that same correction to identical or semantically similar errors across the entire batch. This correction propagation mechanism maintains consistency and drastically reduces redundant manual effort.

• Pattern recognition: The system identifies repeated extraction errors using fuzzy matching and normalized entity IDs.
• Preview before apply: A summary of all downstream changes is presented for reviewer approval before propagation.
• Undo capability: Propagated corrections can be rolled back in a single action if an error in the correction logic is identified.

The interface maintains a chronological, tamper-proof record of every user interaction and system change. This audit trail provides a complete chain of custody for clinical data modifications, which is essential for compliance verification and medicolegal defense.

• Immutable event log: Records the reviewer identity, timestamp, action taken, previous value, and new value for every resolution.
• Session replay: Allows compliance officers to reconstruct a reviewer's exact screen state and sequence of actions for any past session.
• Exportable reports: Audit data can be exported in standard formats (CSV, JSON) for integration with external governance, risk, and compliance (GRC) systems.

A visual comparison interface that highlights specific textual, structural, or coding differences between an AI-generated output and a human-corrected version. In medication reconciliation, a diff view might display the AI-extracted list on the left and the existing EHR record on the right, with color-coded inline highlights marking additions, deletions, and modifications. This pattern reduces cognitive load by eliminating the need for reviewers to manually scan two separate documents. Effective implementations use character-level or token-level differencing to surface subtle discrepancies like dosage changes or frequency updates that would otherwise be missed.

A feature that directly links an AI-generated clinical statement or code to the exact sentence or paragraph in the original medical record. When a reconciliation UI flags a discrepancy between an AI-derived medication and the EHR record, source attribution enables the reviewer to click through and instantly verify the originating text. This eliminates the need to search through lengthy clinical notes and provides an auditable chain of evidence for every data point. Strong implementations highlight the source span on hover and support bidirectional navigation between the structured output and unstructured source.

A granular annotation task where a human reviewer adjusts the start and end character offsets of a highlighted medical entity in unstructured text. In reconciliation workflows, span correction is critical when the AI correctly identifies a medication but misaligns its boundaries—for example, capturing 'metformin 500mg' when the full entity is 'metformin 500mg twice daily'. The UI must support drag-to-resize highlighting or precise offset input fields. Accurate span correction directly improves downstream FHIR resource mapping and ensures extracted data faithfully represents the source document.

The percentage of clinical documents or transactions processed entirely by AI without any human intervention, a key metric for measuring automation efficiency. In reconciliation UI design, the STP rate directly informs interface complexity: high-STP pipelines require lightweight exception-only review screens, while low-STP workflows demand full-featured reconciliation interfaces. Tracking STP over time reveals whether model improvements are translating to reduced review burden. A well-designed reconciliation UI should display real-time STP dashboards segmented by document type, clinical domain, and error taxonomy.

A mechanism that automatically applies a single human correction to identical or semantically similar errors across a batch or downstream dataset. When a reviewer resolves a medication discrepancy in the reconciliation UI, correction propagation ensures that the same error—such as a consistently mislabeled generic name—is fixed everywhere it appears without repeated manual intervention. This requires fuzzy matching algorithms and a confidence-scored propagation queue that lets reviewers approve or reject suggested cascading fixes, maintaining consistency while preventing over-correction.

The systematic process of identifying, analyzing, and correcting mismatches between AI-extracted clinical data and the source document or between two independent human reviews. A reconciliation UI must present discrepancies in a structured, side-by-side format with clear accept/reject/merge controls. Common resolution actions include:

• Accept AI output and overwrite the existing record
• Reject AI output and retain the original EHR data
• Merge by selecting fields from both sources
• Manual edit to create a corrected composite entry Each action must be logged in an immutable audit trail for compliance.