Predictive intelligence calibrated to your hospital
PrismML trains models on your patient population, scores them nightly, and writes risk scores, propensity predictions, and operational forecasts back to DataCloud -- enriching every patient profile, care gap, and campaign with ML intelligence. No third-party ML cloud. No patient data leaving your infrastructure.
The Problem with Generic Clinical Benchmarks
A readmission model trained on a US academic medical centre dataset does not know that your patient population skews diabetic, that your ED peaks on Friday evenings, or that patients admitted from a particular feeder clinic have a higher complication rate. Generic models give you a number. PrismML gives you a model that has learned your hospital.
How PrismML Enriches Your Data Platform
PrismML reads from DataCloud, trains on your data, scores your patients, and writes enriched predictions back -- where they surface on every Patient360, care gap, and campaign. The closed loop from data to intelligence to action.
ML Platform Capabilities
Not just training and scoring. PrismML manages the full model lifecycle: experimentation, governed promotion, production monitoring, drift detection, and continuous improvement.
Experimentation & Model Selection
Run experiments across multiple ML frameworks and hyperparameter configurations. Cross-validation, temporal splits, and out-of-time evaluation. Compare candidates side by side. The best model wins on metrics, not opinion.
Training Pipeline
Training jobs pull directly from DataCloud snapshots -- no data export, no format conversion. Schedule training on a cadence or trigger after a data update. Each run produces a versioned artifact with full evaluation metrics.
Batch Inference Engine
Nightly scoring runs against the full patient or entity population. The inference engine reads the production model, scores the latest DataCloud snapshot, and writes enriched predictions back -- available on Patient360 before morning rounds.
Promotion Gates
Models don't reach production without passing governance gates: absolute performance threshold, improvement over the current production model, and calibration quality. Shadow scoring compares candidates against production before promotion. Rollback is one operation.
Drift Monitoring & Auto-Retraining
Feature drift and score drift are tracked on every scoring run. Population monitoring detects cohort shifts. When drift exceeds configured thresholds, retraining triggers automatically with cooldown periods to prevent unnecessary churn.
Explainability (SHAP)
Every tabular model produces feature importance alongside its predictions. Clinicians and ops managers see exactly which signals drove a specific risk score -- not just the number. Explainability is a first-class output, not a post-hoc add-on.
Calibration & Fairness
Probability calibration ensures that a 30% readmission risk actually means 30%. Fairness analysis checks for demographic parity, equalized odds, and subgroup calibration -- healthcare AI must be equitable, not just accurate.
Model Registry & Versioning
Every trained model is versioned with its training dataset reference, evaluation metrics, and hyperparameter configuration. Promote through development, staging, and production states. Complete audit trail of every model lifecycle event.
DataCloud-Native Integration
PrismML connects through the same data transport that powers NovaHub. Feature engineering stays in DataCloud. PrismML reads snapshots, trains or scores, and writes enriched data back -- all within your infrastructure boundary. No new pipelines to maintain.
What Gets Scored
Clinical risk, patient engagement, financial outcomes, and operational planning -- each grounded in your DataCloud data, each enriching your patient profiles and powering downstream actions.
Clinical Risk
Readmission probability at discharge. In-hospital mortality risk for active admissions. Length-of-stay prediction for bed planning. Disease progression and adverse event risk for chronic populations. Scores are patient-level, updated nightly, and surface on Patient360.
Patient Engagement
Appointment no-show probability for scheduling optimization. Patient disengagement and churn risk for care programme management. Preferred communication channel and optimal contact timing -- learned from each patient's own engagement history.
Financial & Revenue Cycle
Insurance claim denial probability before submission -- flagging claims for coding review before they reach the payer. High-cost patient identification for care management targeting. Scores integrate with revenue cycle workflows as a pre-submission quality layer.
Operational Planning
Facility and department-level demand forecasting for staffing, procurement, and capacity planning. Predictions account for day-of-week patterns, seasonal cycles, and branch-level historical volume. Entity is department or facility, not individual patient.
The Closed Loop: From Scores to Patient Outcomes
ML predictions don't sit in a dashboard. They drive action.
How ML Scores Enrich DataCloud
Scored predictions write back to DataCloud as enriched fact data. The clinical intelligence layer picks them up automatically -- risk scores become patient flags, propensity scores feed cohort definitions, channel preferences optimize journey routing.
MLOps Without the Overhead
Most hospital data teams that want ML end up running three separate systems: a feature store, a training platform, and an inference service. Each with its own pipeline, its own failure modes, its own deployment process.
PrismML collapses this. DataCloud is the feature store. PrismML handles training and inference. Scored output goes back to DataCloud and surfaces through NovaHub. The only thing your team manages is which models are in production and when training runs.
Model deployment is a state change in the registry -- development to staging to production. A rollback is the same operation in reverse. Every inference run is traceable to the model version, the training dataset, and the scoring timestamp.
Enterprise ML Governance
Healthcare ML requires more than accuracy. PrismML enforces governance at every stage of the model lifecycle.
Promotion Gates
3-gate governance
Models must pass performance thresholds, beat the current production model, and meet calibration standards before deployment. No shortcuts.
Shadow Scoring
Pre-promotion testing
Candidate models score alongside production models before promotion. Compare outputs on real data before making the switch.
Fairness Analysis
Equitable AI
Demographic parity, equalized odds, and subgroup calibration checks. Healthcare AI must be equitable across patient populations -- not just accurate on average.
Multi-Tenant Isolation
Per-hospital models
Complete model isolation per hospital. One hospital's training data never influences another's models. Tenant-scoped registry, artifacts, and scoring runs.
Add Predictive Intelligence to Your DataCloud
PrismML is available as an add-on to existing DataCloud deployments. If your data platform is already running, PrismML can be operational in weeks -- training on your data, scoring your patients, enriching your profiles.