2026

Hospital Data Forecasting

Hospital activity forecasting system for Pitié-Salpêtrière using SARIMA model. Analysis of 132 monthly observations (2012-2022), COVID-19 impact modeling, and interactive Streamlit dashboard with Normal/Crisis modes.

SARIMA forecast of monthly ER visits for 2023-2024, with the 2012-2022 history and the 95% confidence interval

fig. 01 · SARIMA forecast of ER visits for 2023-2024, 95% confidence interval

Hospital Data Forecasting is a predictive analysis project for the emergency department activity of Pitié-Salpêtrière Hospital (Paris). The system uses a SARIMA(1,1,1)(1,1,1,12) model trained on 132 monthly observations (2012-2022) to forecast activity over 24 months (2023-2024) with 95% confidence intervals. The project includes an in-depth COVID-19 impact analysis, crisis scenario simulation, and an interactive Streamlit dashboard with temporal filtering and crisis intensity adjustment. Team project of 4 for Epitech Digital School.

Ten years of visits

Heatmap of monthly ER visits, one row per year from 2012 to 2022 and one column per month, from pale yellow to dark red

fig. 02 · Monthly ER visits by month and year, 2012-2022

The series holds 132 monthly observations (2012-2022), rebuilt from the PSL-CFX annual reports. The month-by-year view shows the steady rise and the summer dip that returns every year.

COVID-19 impact

COVID-19 waves isolated in the time series

fig. 03 · The COVID-19 waves isolated in the series, to model their impact

The COVID-19 waves are isolated in the series to model their exceptional impact without breaking the seasonality learned on the rest of the period.

SARIMA diagnostics

Four SARIMA residual diagnostic charts: standardized residuals, histogram with density, normal Q-Q plot and correlogram

fig. 04 · Residual diagnostics of the SARIMA(1,1,1)(1,1,1,12) model

SARIMA(1,1,1)(1,1,1,12) model optimized for monthly seasonality. Residuals are checked (histogram, Q-Q plot, correlogram) before keeping the model: R² = 89%, MAE = 892 visits, MAPE = 6.2%.

Normal / Crisis mode

Streamlit dashboard in Crisis mode: mode selector and intensity slider in the sidebar, key indicators and a chart comparing the normal scenario to the crisis scenario over 2023-2024

fig. 05 · Streamlit dashboard, Crisis mode, simulation at intensity 1.35

Streamlit dashboard with two modes, Normal and Crisis, temporal filtering by year and quarter, and an adjustable crisis intensity that recomputes the scenario against the baseline forecast.

Challenges

Building the dataset from PSL-CFX annual reports
Modeling seasonality and the exceptional COVID-19 impact
Creating reliable forecasts with confidence intervals for hospital planning
Simulating crisis scenarios with adjustable intensity

Solutions

Data pipeline with extraction and cleaning from annual reports
SARIMA(1,1,1)(1,1,1,12) model optimized for monthly seasonality
Interactive Streamlit dashboard with Normal and Crisis modes
20+ visualizations (Plotly, Seaborn) with technical report and presentation

Results

R² = 89%, MAE = 892 visits, MAPE = 6.2%
24-month forecasts with 95% confidence intervals
Interactive dashboard with temporal filtering and crisis simulation
Complete deliverables: dashboard, technical report, implementation plan

Technologies

Python · SARIMA · Streamlit · Pandas · Plotly · Scikit-learn · Statsmodels