All work
Applied AI · Clinical Decision Support

Clinical
Risk Engine

Calibrated malignancy risk scoring for biopsy triage — designed for the clinician, not around them.

Biopsy
Validate
Ensemble
Calibration

01 · Problem

A probability isn't a decision.

Pathologists triage biopsy cases under heavy cognitive load. A raw model probability — even a confident one — doesn't tell them when the model is uncertain.

The cases that matter most are the ambiguous ones, sitting on either side of the decision boundary. Those need a flag, not just a number.

High cognitive load

Mis-triage is expensive

Ambiguity needs a flag

02 · Solution architecture

Calibration is a first-class layer.

Raw ensemble probability isn't a clinical signal. Calibration makes the number actionable.

The spine

FNA vector

30 features

Validation

schema + range

Ensemble

GBM + RF voting

Calibration

isotonic + Wald CI

Reads downstreamTriage payloadAmbiguity flagCohort %ileAudit log

FNA vector

30 cell-nucleus features per biopsy slide.

Validation

Schema enforcement and range checks before inference.

Ensemble

Gradient boosting + random forest voting, SHAP attribution.

Calibration

Isotonic calibration over the WDBC training cohort. The ambiguity flag fires downstream when the 90% CI straddles 0.5.

03 · Live triage

One case at a time.

Pick a case, see the calibrated probability, watch the ambiguity flag fire when the model is unsure.

The case

+5 more features applied silently

The inference

p(malignancy)

0.02

CI 0.010.03

Worst concave points−2.23
Worst radius−1.26
Worst perimeter−1.21
Worst area−1.02

The triage

Ambiguity flag · dark

Standard review queue

Cohort position · 4th %ile

1. Case Pick a case — the surrogate loads its biopsy feature vector.

Interactive prototype · calibrated surrogate of the production ensemble.

04 · Impact

From probability to decision.

A calibrated probability with an ambiguity flag turns a raw model output into something a clinical workflow can route on.

0.99

AUC (calibrated ensemble)

0.041

Brier loss post-calibration

22ms

per-case inference latency

Operator audiencesPathology labsTelemedicine networksClinical research

05 · What's next

From a triage signal to a learning loop.

The system improves with every clinician decision it observes.

Roadmap

FHIR ingestion

Pull FNA observations from PACS/LIS directly.

Roadmap

Drift monitoring

Watch input distributions, auto-flag calibration drift.

Roadmap

Human-in-the-loop

Clinician overrides feed the calibration retrain queue.

Triage that knows what it doesn't know.

Happy to walk through the calibration approach, the ambiguity policy, and what FHIR integration would look like.