Nurse Unit Census Prediction

Multi-horizon census forecasting for hospital nurse units. Five model types trained per unit on 21 months of hourly admit/discharge/transfer data, delivering 1-, 4-, 24-, and 72-hour predictions to three operational dashboards.

99.7%
±2 patient accuracy at 1-hour horizon · validation set
View dashboards → Model cards Methodology
10
Active nurse units
8
Forecast horizons (1h–72h)
5+1
Model types + ensemble
154
Total bed capacity (max observed)
138
Latest house-wide census
87.6%
±2 accuracy at 72-hour horizon
34 / 34
Pytest cases passing — view suite
Operational architecture
GitHub Actions crondaily 12:15 UTC
→
Synthetic ADT generatorscripts/generate_synthetic_hour.py
→
Forecast pipelineRF · LightGBM · LSTM · ARIMA · Prophet · Ensemble
CSV exportsoutputs/tableau/
→
Dashboard rebuildscripts/build_dashboards.js
→
Git commit + pushgithub-actions[bot]
GitHub Pagesstatic dashboards (this site)
↘
Tableau Publicrefreshes from raw.githubusercontent.com

In production, the synthetic ADT step is replaced by an ETL job pulling the live data warehouse; everything downstream is unchanged.

Featured: Executive Census Summary

Executive Census Summary preview
House-wide capacity utilization, 72-hour forecasts per unit, capacity alerts. View live →
Quick links

Model cards

Six models

Per-model architecture, hyperparameters, training data, validation accuracy across all eight horizons, strengths and limitations.

Methodology

Pipeline · features · evaluation

Data preprocessing, leakage-safe feature filtering, per-unit per-horizon training strategy, evaluation metrics and split.

Test suite

34 cases · pytest

Per-test catalog covering data integrity, leakage prevention, chronological splits, metric correctness, model fits, and ensemble weighting. Data-free subset runs in CI.

Source code

GitHub · MIT licensed

Full repository: pipeline source, tests (34 cases), GitHub Actions workflow, configuration, and this static site.

Last regenerated: 2026-05-12 14:51 UTC · Daily refresh via GitHub Actions