Statistical methods for determining epidemic threshold in endemic disease surveillance: a scoping review and benchmark

Background An epidemic threshold is a statistically defined level of disease activity used in public health surveillance to signal a potential outbreak when incidence rates exceed this value. The timely detection of outbreaks is vital, but there is no consensus on the optimal statistical method for setting these thresholds. This study systematically reviewed and benchmarked state-of-the-art methods to identify the most effective approaches. Method A scoping review were conducted in PubMed, Embase, and Scopus up to July 2024 to identify outbreak detection and threshold determination methods. The review mapped methodological characteristics and applications in infectious disease surveillance, identifying 20 distinct approaches. Of these, 18 commonly used methods were selected for benchmarking. Performance evaluation was conducted using simulated epidemic scenarios to assess sensitivity, timeliness, and alarm stability. Empirical surveillance data of Dengue and Hand, Foot, and Mouth Disease in Ho Chi Minh City (2016–2024) were used to examine the practical ability of selected algorithms to generate alerts and detect unusual transmission patterns. Results The scoping review screened 30,699 records and included 55 studies, identifying 20 outbreak detection methods, of which 18 widely used algorithms were selected for benchmarking. Simulation results demonstrated marked heterogeneity in performance: control chart–based methods, particularly CUSUM and its variants, consistently achieved the earliest detection but frequently generated prolonged alarms, whereas Farrington-based and CDC approaches showed better control of alarm duration after incidence returned to baseline. Bayesian methods demonstrated greater adaptability under complex epidemic patterns, especially in multi-peak scenarios. Application to real surveillance data confirmed similar alerting patterns, with notable differences in signal persistence and consistency across diseases and spatial levels. Conclusions Detection performance varied substantially by algorithm and epidemic pattern. Control chart methods provided high early sensitivity but were prone to persistent alarms, while Farrington-based approaches offered better control of prolonged signaling. Bayesian methods demonstrated greater adaptability to complex epidemic dynamics, particularly in multi-peak settings. Algorithm selection for routine surveillance should therefore balance early sensitivity with signal stability, especially when applied to finer spatial resolutions.