Governed aggregates · advisory-only prediction
Analytics & Predictive Modelling Centre
Move from service performance and governed datasets to transparent, human-reviewed predictive action in one workspace.
Service Intelligence
Filter governed aggregate measures, inspect a single signal, then open the workspace that owns the response.
Loading governed measures…
Diagnosis to Treatmentdays
Treatment Status
Stage Distribution
Owned Priorities
Facility Performance HeatmapSelect a cell to drill down
Governed Data & Definitions
Understand the cohorts, versioned datasets, indicator definitions, and quality profile feeding analytics and models.
Cohorts
| Cohort | Patients | Version | Status |
|---|
Published Datasets
| Dataset | Type | Rows | Status |
|---|
Governed Definitions
| Indicator | Version | Use |
|---|
Variable Quality
Continue the Task
Predictive Workflow
Transparent model training, testing, and patient scoring
This workspace shows the real authenticated user, the approved canonical dataset, the latest training and deployment state, and the exact output returned during scoring.
Clinical validity is not established by deployment aloneSelect a model to inspect intended use, validation, calibration, clinical utility, equity, privacy and monitoring evidence. Missing evidence is shown explicitly.
0 / 12
Clinical Prediction Queue
Loading registry summary...
InsightsPatientsCoordination
1
AuthenticateResolve the current session from the bearer token and enforce facility or global scope.
2
Train and testBuild the approved canonical dataset, train on 80%, test once on the held-out 20%, and register a research candidate.
3
Score and reviewScore one canonical patient record, store the output, and keep review decisions visible.
Clinician capture that matters
Why review was triggeredMissed visit, toxicity concern, treatment interruption, recurrence suspicion, referral handoff, or severe risk change.
What action was takenAcknowledged with action, acknowledged without action, or disagreed with written reasoning before the alert can close.
What happened nextObserved outcome, admission, MDT escalation, repeat labs, tracing, or palliative transition to close the loop.
Data scientist oversight that matters
Trustworthiness of the deployed modelDataset hash, deployment freshness, validation metrics, staleness, and drift-sensitive facility patterns.
Readiness of the input dataFeature completeness, missing-input degradation, label scarcity, failed mappings, and governed canonical source coverage.
Defensibility of the outputTraining trace, top drivers, explainability, persisted review outcomes, and disagreement patterns requiring recalibration.
Pipeline Stages
Session and Scope
Request Trace
Pending Reviews
Risk Stratification
Governance
Model Comparison — registry versions
Model Readiness Assessment
Eligibility, missingness, leakage risk, and subgroup coverage by predictive use case.
Readiness Matrix
Model-Ready Dataset Builder
Generate an approved dataset shell from the FHIR-aligned canonical cervical model before training.
Build Dataset
Dataset generation is advisory and metadata-first. Training should only proceed from approved datasets.
Dataset Catalog
Data Audit
Pre-training quality checks and fixes
Issues and Applied Fixes
Risk Worklists
Action-oriented queues for follow-up, treatment delay, referral, diagnostic delay, and data-quality risk.
Follow-up and Appointment Risk
Treatment / Referral / Diagnostic Risk
Monitoring & Drift
What the data scientist needs to watch before recommending retraining or threshold changes.
Missing Input and Coverage
Operational Monitoring
Prediction-to-Action Effectiveness
Bias and Fairness Assessment
Approved subgroup slices only. Advisory activation still requires governance review when fairness concerns are flagged.
Fairness Monitoring
Cancer Prediction Model Laboratory
Train, inspect, and compare reproducible experiments from the cleaned canonical cancer dataset.
Guided ML experiment
Choose algorithms, define the answer, then compare evidence
The platform builds X from de-identified clinical predictors and y from the selected outcome. Every selected model receives the same stratified 80% training and untouched 20% testing split.
X · questions / predictors
Clinical facts used to ask the model a question. Direct identifiers are excluded.
y · answer / target
Choose the outcome the algorithms must learn to estimate.
Row-level X and y remain server-side.
Train / test split
80% train · 20% test
Training Trace
Training Output
Pending Model Approvals
A trained candidate is never used for live patient scoring until someone other than the trainer approves it here.
Explainability
Feature importance and SHAP-style direction effects
Global Importance
Latest Patient Explanation
Clinician Response Capture
Capture the real-world judgement that makes the advisory defensible and clinically useful.
What Must Be Captured
Current Review Context
Response Workflow
Trigger reasonDocument whether the advisory was triggered by missed follow-up, toxicity, treatment delay, recurrence suspicion, referral discontinuity, or a severe change in risk band.
Clinical responseCapture one of the governed responses: acknowledged with action, acknowledged without action, or disagree with documented reasoning.
Resulting actionRecord the concrete next step: MDT review, admission, lab repeat, tracing, treatment hold, transfusion, palliative referral, or continued plan.
Outcome closureClose the loop with an observed coded outcome and a short note so disagreements and overrides can feed monitoring.
Patient Review
Clinical score uses only canonical `cdr_cervical_*` data for the patient reference you select.
Canonical Patient Lookup
Manual edits on this screen do not drive prediction. The score uses the approved canonical cervical record linked to the patient reference.
One-Screen Clinical Summary
Pre-Cycle Safety Gate
Risk Advisory Badge
--
No advisory yet
Scoring Transparency
Confidence & Freshness
Clinical Recommendations
Proposed Next Action
Score a patient to see governed next actions.
Clinician Review Loop
Capture the clinician response before closing any advisory. Model output is advisory-only and cannot act autonomously.
No score selected for review.
Prediction Audit Trail
Every predictive export, setting update, patient-level view, review decision, and outcome record remains append-only.
Recent Predictive Audit Events
Model Settings
Thresholds and governance settings are logged and apply to future predictions only.
Current Settings
Update Setting