Using community syndromic surveillance to anticipate enterovirus-related school class suspensions: A real-time LSTM approach.

OBJECTIVES

This study aimed to develop a district-level early warning framework using a long short-term memory (LSTM) model with adaptive thresholds to detect abnormal enterovirus-like (EV-like) syndrome activity from clinic-based surveillance data and assess its ability to identify aberrations earlier than the onset of class suspension events in Taipei.

METHODS

Daily counts of EV-like syndrome cases and preschool and primary school class suspension records (2022-2025) were analyzed at the district level. LSTM models forecasted EV-like syndrome activity 14 days in advance using a 30-day lookback window. Abnormal signals were identified using residual-based anomaly detection with district-specific adaptive thresholds. The model performance was evaluated using mean absolute error, root mean square error, sensitivity, specificity, and precision.

RESULTS

The LSTM model achieved accuracies of 0.94 for preschool data and 0.96 for primary school data in Taipei City, with consistent performance across districts. The proposed LSTM-based framework effectively captured temporal changes in disease activity and generated alarms generally earlier than abnormal class suspension days.

CONCLUSION

A district-level LSTM model with adaptive threshold detection offers a scalable approach for early warning of EV-like syndrome activity, supporting timely local responses, targeted prevention strategies, and improved preparedness for schools and families.