tuva-ml-models/medicare-cost-encounter-model-prospective

Medicare Prospective PMPM and Encounter Prediction Models

Model ID: Medicare LDS 2023 Prospective PMPM + Encounter Model Bundle Model Types: XGBoost regressor bundle (Tweedie for PMPM, Poisson for PMPC) Dataset: 2023 CMS Limited Data Set (LDS) 5% Sample Target Level: Member month (annualized for evaluation)

---

What the Models Predict

This repository contains a bundle of prospective regression models that forecast next-year healthcare cost and utilization. Models are trained using prior-year demographics and lagged clinical features (e.g., condition flags, HCCs, CMS chronic indicators) to predict outcomes for the subsequent prediction year.

The models predict outcomes at three levels of granularity:

1. Total PMPM: A model predicting total paid amount PMPM (Per Member Per Month).
2. Cost & Utilization by Encounter Group: Models that predict PMPM paid amount and PMPC (per-member-per-count) for broad encounter groups (INPATIENT, OUTPATIENT, OFFICE_BASED, OTHER).
3. Cost & Utilization by Encounter Type: Models that predict PMPM and PMPC for granular service types (e.g., ACUTE_INPATIENT, TELEHEALTH, HOME_HEALTH, DIALYSIS).

Note on prediction type: These are prospective models — they use features from the baseline period to predict the following year’s outcomes. All core models output a rate (PMPM or PMPC) that is multiplied by member months to yield annualized values for reporting and evaluation.

---

Intended Use

This bundle supports forward-looking financial and operational workflows:

• Benchmarking: Comparing observed vs. predicted annual costs to identify efficiency opportunities.
• Healthcare research: Studying drivers of future cost and utilization in Medicare populations.
• Resource planning: Anticipating demand across encounter groups and service types.
• Care management: Identifying members likely to have high cost or utilization in the next year.
• Prospective forecasting: Budgeting and premium setting using next-year cost forecasts.

---

Model Performance

These metrics reflect performance on a held-out evaluation split from the 2023 CMS LDS data produced by the training pipeline in this repo. Models predict PMPM/PMPC rates; metrics are computed on annualized values (rate × member months). See prosp_model_eval_metrics_train.csv for full details across all targets.

Annual Cost (PMPM) — Summary

• Overall: R² 0.185, MAE% 97.06%
• Inpatient: R² 0.085, MAE% 146.33%
• Outpatient: R² 0.158, MAE% 102.83%
• Office-based: R² 0.070, MAE% 91.49%
• Other: R² 0.023, MAE% 114.80%

Annual Utilization (PMPC) — Encounter Groups

• Inpatient: R² 0.195, MAE% 134.95%
• Outpatient: R² 0.419, MAE% 77.94%
• Office-based: R² 0.315, MAE% 65.48%
• Other: R² 0.271, MAE% 83.18%

Note: The complete list of encounter type metrics (e.g., ACUTE_INPATIENT, TELEHEALTH) are included in prosp_model_eval_metrics_train.csv along with calibration factors and prediction-to-actual (P/A) ratios.

---

Files Included

• 2023_prospective_models_bundle.pkl — Bundle with trained XGBoost models, per-target calibration factors, and feature list.
• pmpm_encounters_prospective_train.py — Training pipeline (prospective): builds PMPM and PMPC models, logs metrics, and writes the bundle.
• pmpm_encounters_prospective_predict.py — Inference + evaluation script: loads the bundle, scores the prediction year, and logs metrics/predictions.
• prosp_model_eval_metrics_train.csv — Evaluation metrics exported from training (annualized MAE, MAE%, R², P/A ratios, calibration factors) for each target.
• prosp_encounter_model_feature_frequency.csv — Feature prevalence (non-zero rates) in training features used by the bundle.
• prosp_encounter_model_hyperparams.csv — Tuned hyperparameters used for key PMPM targets (when tuning enabled).

---

Understanding Model Artifacts

This repository includes CSV outputs summarizing what the prospective models learned and how they were trained.

Feature Frequency (prosp_encounter_model_feature_frequency.csv)

• Positive rate of each engineered feature in the training set (share of rows where the feature > 0).
• Useful to monitor input drift and sparsity when scoring new populations.

Hyperparameters (prosp_encounter_model_hyperparams.csv)

• Stores tuned hyperparameters for targets where tuning was enabled (e.g., OVERALL PMPM). Includes objective, tweedie_variance_power, tree depth, regularization, and n_estimators.

Evaluation Metrics (prosp_model_eval_metrics_train.csv)

• One row per target × metric with annualized MAE, MAE%, R², prediction-to-actual ratio, calibration factor, and counts of evaluation rows.
• Includes summary rows for groups (e.g., PMPM: INPATIENT/OUTPATIENT) and detailed encounter types (e.g., PMPC: ACUTE_INPATIENT).

---

Quick Start

The training and inference scripts are designed to run in a Snowflake environment using Snowpark.

Prereqs:

• Snowflake account and permissions to read MEDICARE_LDS_FIVE_MULTI_YEAR.BENCHMARKS.PERSON_YEAR_PROSPECTIVE.
• A stage for model artifacts: @MEDICARE_LDS_FIVE_PERCENT.BENCHMARKS.MODEL_STAGE.
• Python deps: pandas numpy scikit-learn xgboost snowflake-snowpark-python.

Train (writes metrics to Snowflake and uploads the bundle to the stage):

• Run pmpm_encounters_prospective_train.py within a Snowpark session. It trains PMPM/PMPC models (groups and types), computes calibration factors on train, evaluates on a held-out split, and writes 2023_prospective_models_bundle.pkl to the stage.

Predict + evaluate (scores the prospective prediction year and logs metrics/predictions):

• Run pmpm_encounters_prospective_predict.py. It downloads the bundle, scores the configured SCORING_YEAR, applies calibration, logs evaluation metrics, and writes per-member predictions with BENCHMARK_KEY.

All per-target metrics, calibration factors, and feature frequencies are exported in the CSVs included in this repo for reference.