Objective:

To examine whether unexpected deviations from long short-term memory (LSTM)-predicted enterovirus-like (EV-like) illness activity could provide early warning of abnormal class suspensions in Taipei preschools and primary schools.

Approach:

• Study Design and Data Sources: This retrospective time-series modeling study combined daily pediatric outpatient surveillance data from 110 participating clinics with school-level class suspension records and total class counts across all 12 districts of Taipei City from January 2022 to January 2025.
• Model Development: District-specific LSTM models used a 30-day lookback window to forecast daily EV-like syndrome activity up to 14 days in advance. Large positive prediction residuals triggered alarms based on district-specific adaptive thresholds.
• Syndrome Groups: EV-like illness was the primary surveillance category and included hand-foot-and-mouth disease (HFMD), herpangina (HA), and other enterovirus-related diagnoses. HFMD and HA were also evaluated separately using their corresponding diagnostic codes.
• Outcome Assessment: Model-generated alarms were compared with abnormal class suspension days defined using district-specific suspension-rate thresholds. Performance was assessed using forecast-error and classification measures, including accuracy, sensitivity, specificity, and precision.

Key Findings:

• Enterovirus infections in Taiwan decreased by 67.1% during the COVID-19 pandemic and subsequently increased by 619.5%, according to evidence cited as background for the study.
• The LSTM framework achieved accuracies of 0.94 for preschool data and 0.96 for primary school data, with generally consistent performance across districts.
• Model-generated alarms frequently preceded periods with abnormal class suspensions, indicating that increases in pediatric outpatient EV-like illness activity may provide advance warning of suspension-related events.
• District-specific residual thresholds accounted for differences in baseline activity and variability rather than applying a single universal cutoff.
• Forecasting performance was stronger in districts with relatively stable outpatient activity and more difficult in districts with large or highly variable epidemic waves.
• When applied to HA and HFMD among preschool children, the framework achieved average accuracies of 0.87 and 0.97, respectively.

Interpretation:

District-level LSTM forecasting combined with adaptive residual thresholds identified unusual increases in EV-like syndromic activity and generally generated alarms before abnormal class suspension days. These findings suggest that routinely collected pediatric outpatient data may complement school surveillance and support earlier local preparedness. However, the observed relationship was assessed using aggregate district-level data and does not establish that model alarms predict individual school or class suspensions.

Limitations:

• The dataset did not include detailed information about affected classes or the number of children with EV-like syndromes within each suspended class; associations could therefore be assessed only at the aggregate level.
• Model performance may be influenced by data quality, missing values, and sudden external changes that cannot be fully learned from historical observations.
• The framework used a limited set of surveillance variables and did not incorporate potentially relevant meteorological, mobility, environmental, or virological data.
• The model was developed using data from Taipei City, and further validation in other settings with additional predictors is needed to assess generalizability.

Conclusion:

The district-level LSTM framework used routine pediatric outpatient syndromic data and school suspension records to capture local temporal patterns and provide early warning signals for abnormal preschool and primary school class suspension events. The findings support further evaluation of residual-based alarms as a scalable tool for local enterovirus surveillance, targeted responses, and preparedness among schools and families.