Improving ED admissions forecasting by using generative AI: An approach based on DGAN.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Generative Deep Learning synthesizes realistic data, valuable for challenging data acquisition scenarios.
• Patient admissions forecasting is crucial for hospital Emergency Department resource management.

Purpose of the Study:

• To employ the DoppelGANger algorithm for enhancing patient admissions forecasting in hospital Emergency Departments.
• To assess the effectiveness of synthetic time series data generated by DoppelGANger in improving predictive model performance.

Main Methods:

• Utilized the DoppelGANger algorithm for generating synthetic time series data, conditioned on unique attributes.
• Implemented a Train-Synthetic-Test-Real framework for validating synthetic data.
• Augmented original datasets with synthetic data to improve the Prophet forecasting model's accuracy.
• Applied the methodology to datasets with varying training and testing periods (4-year train/1-year test and 3-year train/2-year test).

Main Results:

• The DoppelGANger-augmented Prophet model outperformed the baseline Prophet model in forecasting accuracy (reduced SMAPE).
• Specific improvements in SMAPE were observed: 7.30 to 6.99 (4-year set) and 22.84 to 7.41 (3-year set) for forecasting.
• Data replacement and augmentation tasks further reduced SMAPE values, surpassing models trained solely on real data.
• Consistent performance improvements were noted across different data aggregations.

Conclusions:

• Generative algorithms, like DoppelGANger, can effectively extend training datasets to enhance predictive models for Emergency Department admissions.
• Improved forecasting accuracy facilitates more efficient hospital resource allocation and patient management strategies.