Deep recurrent models for forecasting infectious diseases.

Area of Science:

• Epidemiology
• Artificial Intelligence
• Public Health

Background:

• Infectious diseases pose significant global health challenges, necessitating early detection of outbreaks for effective response.
• Timely identification of unusual increases in case numbers is critical for resource allocation and intervention planning.

Purpose of the Study:

• To develop and evaluate a predictive framework using deep learning models for forecasting COVID-19 cases in Saudi Arabia.
• To detect early signs of unusual increases in COVID-19 cases by analyzing temporal patterns in search engine data.

Main Methods:

• Utilized time series data and Google Trends for search terms like "fever," "COVID," and "cough" as model inputs.
• Developed predictive models using Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Gated Recurrent Unit (GRU) neural networks.
• Trained models on preprocessed data, adjusting for time lags, and evaluated performance using Mean Square Error (MSE) and F1-score.

Main Results:

• Bidirectional LSTM achieved the highest F1-score (0.83) for predicting "COVID" surges, outperforming LSTM (0.73) and GRU (0.77).
• BiLSTM demonstrated superior performance over LSTM and GRU for early detection of "fever" and "cough" trends.
• While BiLSTM has higher computational costs, LSTM and GRU provide efficient alternatives for rapid anomaly detection.

Conclusions:

• Deep learning models, particularly BiLSTM, are effective for early anomaly detection in infectious disease trends.
• The proposed framework supports timely healthcare interventions and the development of real-time disease monitoring systems.
• Leveraging search trends alongside epidemiological data enhances the ability to predict and respond to public health emergencies.