Total Spend
Expected paid amount per member over customizable time horizons
Predictive Models
Built for Tuva Analytics Teams
In-Warehouse Predictive Intelligence
Built by healthcare data engineers for analytics teams, Illuminate Predictive Models removes the need to master ML pipelines, point-in-time feature engineering, or model deployment. Instead of spending six figures on vendor risk scores or months building custom ML infrastructure, you get production-ready predictions that run inside your existing dbt workflow.
- No external ML platform required: runs inside your existing dbt workflow
- No opaque scoring logic: transparent features, diagnostics, and model metadata
- No selection bias: trained on your population, your cost structure, and your data completeness
What You Can Predict
Out-of-the-box spend and utilization models, plus configurable targets for your own workflows.
Inpatient Utilization
Predicted encounter rates for acute inpatient admissions
Emergency Department Visits
ED encounter probability and expected frequency
SNF Utilization
Skilled nursing facility encounter predictions
Custom Targets
Fully configurable target policy for any encounter type and time horizon
Overview
Illuminate Predictive Models makes it easy to train and deploy healthcare risk models without having to build ML infrastructure from scratch or depend on opaque third-party vendors. We train gradient-boosted models directly in your data warehouse on your own claims data, producing calibrated spend and utilization predictions as dbt tables with no external infrastructure required.
Comprehensive Feature Engineering
- Demographics: Age, sex, race, state, enrollment tenure, and cold-start indicators
- Utilization History: Paid amounts and encounter counts across 3/6/12-month lookback windows by encounter type
- Chronic Conditions: CMS chronic condition assignments from both claims mart and raw diagnosis codes
- HCC Risk Scores: Hierarchical Condition Category assignments normalized across payers and plan versions
Calibrated Probability Outputs
- Count Thresholds: P(Y >= k), the probability of at least 1, 2, 3, or 5 encounters in a given category
- Spend Percentiles: P(spend in top k%), the probability a member falls in the top 1% or 5% of spenders
- Isotonic Calibration: Predictions calibrated to match aggregate actuals for reliable population-level estimates
Clinical and Operational Insights
- Point-in-time feature construction with no lookahead bias or data leakage
- Person-level train/test splits that prevent information leakage from overlapping monthly windows
- Claims lag adjustment to account for incomplete recent claims data
- Feature importance and fill-rate diagnostics to catch data quality issues early
- Model registry with signature-based reuse to avoid unnecessary retraining
Purpose-Built for Tuva Users
- Runs entirely within your dbt workflow with no Jupyter, Airflow, or external ML platforms
- Trained on your population, your cost structure, your data completeness, with no selection bias
- Separate models per data source for multi-payer environments
- PHI-safe summary exports for non-technical stakeholders
- Versioned model artifacts with full audit trail
Differentiation: Build vs Vendor vs Illuminate
| Feature | Build In-House | Vendor Risk Scores | Illuminate Predictive Models |
|---|---|---|---|
| Training Data | Your own claims population, but requires substantial engineering investment | National averages that may not match your data | Your own claims population with no selection bias |
| Infrastructure | Pipeline orchestration, model hosting, serving, and monitoring all owned by your team | Separate ML platform, API integrations, or file transfers | Runs in your warehouse via dbt with zero external dependencies |
| Calibration | Must be designed and maintained internally | Requires manual adjustment factors for your population | Automatically calibrated to your actuals |
| Transparency | High if your team invests in diagnostics and documentation | Black-box scores with limited explainability | Full feature importance, fill rates, and diagnostics |
| Customization | Flexible but costly to build and maintain | Fixed model outputs, vendor-controlled roadmap | Configure targets, horizons, features, and thresholds via dbt vars |
| Updates | Dependent on internal roadmap and staffing | Annual or semi-annual vendor refresh cycles | Retrain anytime on fresh data with a single dbt run |
| Integration | Custom data products required for activation and BI | CSV drops, API calls, or proprietary formats | Native dbt tables in your warehouse, ready for downstream analytics |
Quickstart Path
- Add Tuva and illuminate_predictive_models to packages.yml and run dbt deps.
- Set minimal vars in dbt_project.yml (for example ml_enabled: true).
- Run dbt run --select package:illuminate_predictive_models.
- Validate outputs in your ML schema before downstream operationalization.
Core Output Contract
| Output Table | Description |
|---|---|
train_model_registry | Train/reuse status, artifact URI, diagnostics, and model metadata for the current run |
predict_values | Predicted values by person, anchor month, target definition, and prediction horizon |
predict_probabilities_long | Threshold and percentile probability outputs, including P(Y >= k) and spend top-percent probabilities |
train_metrics_long | Train/test evaluation metrics, including MAE, RMSE, R2, AUC, Brier, and logloss |
Bring Predictive Modeling Into Your Existing Tuva Workflow
Keep your data, logic, and operational analytics in one place. Illuminate Predictive Models helps your team move from retrospective reporting to proactive risk targeting without adding a separate ML platform.
Book a Demo