ML & Mapping Intelligence
Vorian Connect runs machine learning models in shadow mode — they predict outcomes, suggest rules, and propose LOINC mappings without ever changing a live verification verdict or result. Interface engineers review every suggestion before it becomes operational, giving you AI-assisted speed with full human control.
ML Shadow Predictions
The ML Shadow Predictions dashboard shows what Vorian's XGBoost models would do if they were making verification decisions. Every analyte in every message gets a prediction: auto (the model would pass the result through) or held (the model would flag it for manual review). These predictions run silently alongside the auto-verification pipeline and never override the live verdict.
- Stats cards show total predictions over 14 days, the auto-verify rate, average confidence across all predictions, and the number of distinct analytes tracked
- A daily trend sparkline charts the auto-verify rate over the last seven days — useful for spotting model drift after instrument maintenance or calibration events
- Filter predictions by analyte or by status (auto/held) to focus on specific test panels or investigate held predictions
- Each prediction row shows the analyte code, predicted status, confidence score, model version, timestamp, and source message ID
- The model features include the result value itself, reference range deltas, patient delta checks, QC status, sample type, dilution factor, and historical override counts — 16 features per analyte
- Predictions are stored in the
ml_predictionstable and served by a Python XGBoost REST service via the gateway
ML Rule Suggestions
Vorian Connect can analyse historical message patterns and suggest new auto-verification rules for lab review. These suggestions are generated by a Python pattern-extraction service and shown in the Rule Suggestions dashboard — they never become live rules until an interface engineer reviews and approves them.
- Each suggestion includes the analyte code, a human-readable rule description, a confidence score, and the sample count used to generate it
- Suggestions progress through a three-state lifecycle:
pending→approvedorrejected - Approving a suggestion creates a live rule in the rules table with a link back to the originating suggestion for audit traceability
- Rejected suggestions record a reason and optional feedback for model retraining
- Filter the suggestion list by status or analyte code; pagination caps at 200 suggestions per query
Test Catalog Import
Instead of manually mapping every instrument code to LOINC, upload a vendor test catalog — CSV, TSV, Excel workbook, or PDF — and let Vorian Connect suggest LOINC mappings for every test. Azure OpenAI powers the mapping intelligence using semantic embeddings and large language model reasoning.
- Supported formats:
.csv,.tsv,.txt,.xlsx, and.pdf - Select an instrument source first — the catalog is analysed in the context of that instrument's known test codes
- Each catalog row is parsed into instrument code, analyte name, units, and expected reference range
- For every row with a match, Vorian suggests a LOINC code with display name, confidence score, and source (vector search, LLM reasoning, or previous mappings)
- Rows without confident suggestions are left unchecked — you manually map those later
- Review the suggestion table, check the rows you want to apply, and click "Apply selected" — mappings are written to the source's instrument code mapping table in a single batch operation
- Requires
ml:managepermission
Unmapped Results Queue
When a new instrument code arrives without a configured LOINC mapping — either from a new instrument or a test that was missed during initial setup — it lands in the Unmapped Results Queue. The queue groups results by source, instrument code, analyte name, and unit so you can resolve entire families of unmapped results at once.
- Groups show the instrument code, analyte name, source name, unit, pending count, and when the latest occurrence was seen
- For each group, Vorian calls the ML mapping suggestion service — passing the source ID, instrument code, analyte name, and units — to get a LOINC suggestion with confidence score
- High-confidence suggestions (typically from vector embedding similarity or previous tenant mappings) appear as blue callout cards with an "Accept" button — one click maps the entire group
- Manual mapping: enter a LOINC code, display name, and unit, then click "Map all" — the mapping applies to every pending result in that group
- Ignore: mark a group as intentionally unmapped with an optional resolution note — useful for instrument codes that are placeholders, comments, or not clinically relevant
- Expand a group to see every individual result row, including the result value and source message ID (requires
results:viewpermission along withconfig:mappings)
Where It Fits
The ML & Mapping Intelligence tools span two dashboard sections. ML Shadow Predictions, Rule Suggestions, and Catalog Import live under Migration & Intelligence alongside validation tools. The Unmapped Results Queue lives under Configure alongside sources and destinations, since resolving unmapped codes is part of ongoing interface configuration.
All ML features run on Azure OpenAI (not direct OpenAI API) using the following models:
- Embeddings:
text-embedding-3-small— maps instrument codes and LOINC descriptions into a shared vector space for similarity search - Chat completion:
gpt-4o— analyses ambiguous catalog entries that vector search cannot confidently resolve
ML model predictions are viewable by users with ml:manage permission. Mapping suggestions and unmapped result management require config:mappings. Viewing individual result values within unmapped groups also requires results:view.