PathwayIQ

Clinical decision-support intelligence

Calibrated readmission
risk, built for review.

PathwayIQ surfaces calibrated 30-day readmission risk estimates for diabetic patients - grounded in explainable inference, probability calibration, and continuous drift monitoring. Designed to support clinician judgment, not replace it.

Open assessment interface Calibrated on 100,000+ diabetic patient encounters
Model
Calibrated XGBoost
Explainability
SHAP feature attribution
Calibration
Isotonic regression
Monitoring
Automated drift detection
Inference architecture
Input
Patient clinical data
Frontend assessment interface
↓
Inference
FastAPI inference service
Calibrated XGBoost pipeline
Risk stratification engine
↓
Reliability
SHAP explainability layer
Prediction logging & monitoring
Drift detection system
↓
Lifecycle
Automated retraining workflow
Versioned model registry
AI system reliability & monitoring
Calibration reliability Probability reliability layer

Calibrated probability estimates align predicted readmission risk with observed clinical outcomes. Isotonic regression post-processing ensures a predicted risk of 30% corresponds to a clinically meaningful 30% observed rate — not an arbitrary model score.

Isotonic
Calibration method
Brier ↓
Score improved
3 bands
Risk stratification
Reliability diagram - calibrated vs uncalibrated probability
Reliability diagram
Place calibration_curve.png
in /static/ to display
Drift monitoring

Feature distribution shifts are detected automatically. When monitored clinical inputs deviate beyond set thresholds, a retraining signal is raised.

Incoming data
→
Shift detection
→
Threshold check
→
Retrain trigger
Monitored features
number_emergency number_inpatient time_in_hospital number_diagnoses
Retraining workflow

When drift is confirmed, a retraining pipeline activates, evaluates the new candidate, and logs the artifact to the model registry.

Model drift detected
Retraining pipeline activated
New model version generated
Artifact logged to registry
Observability & traceability

Every inference is logged with a request ID, timestamp, input snapshot, and prediction metadata. Latency is tracked to ensure production-grade responsiveness.

142 ms
Avg. API latency
4
Monitored features
✓
Request traceability
✓
Versioned model tracking
Clinical assessment interface

The following workflow simulates real-time clinician-facing risk assessment. Adjust the patient parameters below and the system will return a calibrated probability estimate alongside the top clinical risk factors driving the prediction.

Patient details
0 = youngest · 9 = oldest decade
-
—
—
What's driving this
Predictions and explanations are intended to support clinician review and discharge planning workflows. Final intervention decisions remain under physician oversight.
SHAP feature attributions reflect model behavior patterns and should not be interpreted as causal medical reasoning. Explanations may vary under correlated clinical variables and are intended only as supportive transparency signals.