Forecasting Teleconsultation Demand Using an Ensemble CNN Attention-Based BILSTM Model with Additional Variables.

Area of Science:

• Health Informatics
• Artificial Intelligence
• Time Series Forecasting

Background:

• Accurate forecasting of daily teleconsultation demand is crucial for resource allocation.
• Existing models may not fully capture the complex dynamics influencing teleconsultation usage.

Purpose of the Study:

• To develop and validate a superior deep learning model for enhancing teleconsultation demand forecasting accuracy.
• To identify and integrate significant additional variables into the forecasting model.

Main Methods:

• An ensemble hybrid deep learning model, termed ECA-BILSTM, was developed, integrating Convolutional Neural Networks (CNNs), attention mechanisms, and Bidirectional Long Short-Term Memory (BILSTM).
• Additional relevant variables were identified based on teleconsultation demand characteristics and incorporated into the model inputs.
• The model's performance was evaluated using two real-world teleconsultation datasets from the National Telemedicine Center of China (NTCC).

Main Results:

• The proposed ECA-BILSTM model demonstrated significantly superior forecasting accuracy compared to benchmark models.
• Two key additional variables were identified as highly effective in improving teleconsultation demand prediction.
• The ensemble approach effectively combined the strengths of CNNs, attention, and BILSTM for robust forecasting.

Conclusions:

• The ECA-BILSTM model represents a feasible and promising approach for accurate teleconsultation demand forecasting.
• The integration of carefully selected additional variables substantially enhances prediction accuracy.
• This study provides a valuable tool for optimizing telemedicine service management and resource planning.