See the unusual journey.
Keep the decision human.
Local, explainable anomaly analysis for specimen-handling metadata. Fit a reference from your chosen examples, then inspect unusual transit, queue and custody-scan patterns.
| Record / route | Transit | Missing scans | Analysis |
|---|
Small, inspectable, local.
A three-nearest-neighbor model learns reference distances per route. Features are transit minutes, queue minutes and missing-scan counts, scaled by a robust spread estimate. A leave-one-out 95th-percentile distance sets the review threshold. Routes with fewer than eight reference examples stay unscored.
This is unsupervised distance-based anomaly detection. It has not been clinically validated. Reference selection, sample size and distribution changes can strongly affect results.
Your data stays in this browser
JSON files are processed locally. Only IDs, routes and the three numeric features are retained. Do not import identifiable patient information, even in IDs or routes. Unknown fields are dropped, not exhaustively de-identified.
Prior work & AI disclosure
Adapted from Amrit Lahari's MED-BLACKBOX. Its Splunk integration and hosted-model agent are prior work, retained in source; this new browser app runs its own local kNN model. OpenAI Codex assisted with code, tests and documentation. New UnivaBio work: route-specific model, reference inspector, strict import, review packet and interface.
Source, setup and validation ↗