Emergency Department Demand Forecasting with LSTM

A reproducible deep-learning project for daily and true month-ahead emergency department patient-arrival forecasting using Long Short-Term Memory networks.

The project progresses from one-step forecasting to a monthly-total-aware direct multi-horizon residual LSTM that predicts a complete future month without recursively feeding predictions back into the network.

Final 2023 month-ahead result

Metric	Seasonal-7 Baseline	Total-Aware Direct LSTM	Improvement
MAE	1061.00	901.77	15.01%
RMSE	1260.34	1007.04	20.10%
MAPE	13.87%	11.38%	—
WAPE	13.11%	11.14%	15.01%
R²	0.1961	0.4867	+0.2906

The final test-set mean signed error was +464.15 visits/month, indicating residual underprediction. No post-hoc test-set correction was applied.

Daily forecasting result

Metric	Persistence t-1	Residual LSTM	Improvement
MAE	28.490	22.032	22.67%
RMSE	36.567	28.278	22.67%
MAPE	10.80%	8.53%	—
WAPE	10.71%	8.28%	22.67%
R²	0.6262	0.7765	+0.1503

Temporal design

• Training: 2019–2021
• Validation / model selection: 2022
• Final untouched test: 2023

No random time-series splitting was used.

Final model

• Historical context: 56 days
• Historical features: 9
• Forecast horizon: 31 days
• Known future calendar features: 4
• Forecasting strategy: direct multi-horizon
• Recursive feedback: no
• Residual baseline: previous 7-day seasonal pattern
• Monthly-total-aware loss weight: 0.35

Input features

Historical sequence:

1. standardized patient visits
2. 7-day rolling mean
3. 14-day rolling mean
4. 28-day rolling mean
5. 7-day rolling standard deviation
6. day-of-week sine
7. day-of-week cosine
8. day-of-year sine
9. day-of-year cosine

Known future calendar:

1. day-of-week sine
2. day-of-week cosine
3. day-of-year sine
4. day-of-year cosine

Residual formulation

Predicted demand = Seasonal-7 baseline + LSTM residual correction

Monthly-total-aware objective

Loss = Daily Huber Loss + λ × Monthly Total Loss

with λ = 0.35.

Repository structure

LSTM_Graphs/
├── README.md
├── CITATION.cff
├── requirements.txt
├── .gitignore
├── data/
├── notebooks/
├── src/
├── models/
├── results/
│   ├── metrics/
│   └── predictions/
├── figures/
│   ├── pdf/
│   ├── svg/
│   └── png/
├── tables/
├── docs/
└── kaggle_outputs/

Data

The original source dataset is intentionally not stored in this public repository.

Generated artifacts, aggregated predictions, trained models, metrics, figures, notebooks, and scripts are included where available.

Environment

• Python 3.x
• TensorFlow 2.19
• NumPy
• pandas
• scikit-learn
• matplotlib
• openpyxl
• tqdm

Install dependencies:

pip install -r requirements.txt