Solace Diagnostics
A triage model that flags anomalies in lab results before a technician reviews them.
Industry
Healthcare
Timeline
14 weeks
Team
2 ML engineers, 1 clinical advisor
Category
AI Development
Where things stood
Solace's lab technicians were manually scanning every result for anomalies, which worked but meant a rare abnormal pattern could sit in a queue behind routine results for hours.
How we got there
Clinically-informed labeling
Worked with Solace's clinical advisor to label training data against criteria technicians actually use, not just historical outcome data.
Anomaly triage model
Built a model that flags likely anomalies for priority review rather than replacing technician judgment on any result.
Shadow-mode validation
Ran the model silently alongside technician review for six weeks before it influenced any queue ordering, comparing its flags against real outcomes.
What changed
6 wks
Shadow-mode validation before go-live
Priority queue
Flagged anomalies reviewed first, not first-in-first-out
0
Model decisions made without technician review
Tech stack
Python
The language for the anomaly-detection pipeline and labeling tools.
TensorFlow
Trains the triage model on lab results labeled against technician-defined criteria.
AWS SageMaker
Hosts the model and serves flags into the technician review queue.
PostgreSQL
Stores lab results alongside model flags for the shadow-mode comparison against real outcomes.
scikit-learn
Handles preprocessing and baseline comparisons used to validate the deep model's lift.
Docker
Packages the model service identically across the lab's environments.
