Table 1 The characteristics and result of three studies that compared GLM-based algorithms

Simulated/ Real Dataset

Simulated

Real & Simulated

Simulated

Type of Surveillance System

Weekly Health Surveillance

Daily Syndromic Surveillance

Daily Syndromic Surveillance

Measures to Assess the Performance of the Algorithms

FPR^a, POD^b, POD1week^c, Sensitivity, Specificity, PPV^d, NPV^e, F₁

Sensitivity, Specificity, PPV, Timeliness

POD, Sensitivity, Specificity, PPV, Timeliness

Number of Algorithms Included

21

6

3

List of Algorithms Included

Farrington Flexible, Original Farrington, CDC (historical limits), CUSUM, CUSUM Rossi, CUSUM GLM, CUSUM, GLM Rossi, Bayes 1, Bayes 2, Bayes 3, RKI 1, RKI 2, RKI 3, GLR Negative Binomial, GLR Poisson, EARS C1, EARS C2, EARS C3

OutbreakP, periodic Poisson regression, periodic negative binomial regression

Three control chart-based algorithms commonly referred to as C1, C2, and C3, GLM, EWMA9, EWMA4

RAMMIE, EARS, Farrington Flexible

Result of Performance Measures

Farrington Flexible^f: (1.0%, 99.0%, 43.3%, 34.0%, 20.5%, 95.0%, 58.3%, 0.30)

Original Farrington: (2.3%, 97.7%, 56.9%, 45.5%, 29.0%, 95.4%, 45.0%, 0.35)

Periodic Poisson GLM: (3.3%, 96.8%, 67.8%, 56.6%, 35.6%, 95.8%, 42.3%, 0.39)

GLR Poisson: (15.15%, 84.5%, 75.5%, 60.3%, 45.9%, 95.9%, 16.4%, 0.24)

GLM model was more sensitive than the other algorithms and detected 54% (95% CI = 52%–56%) of the simulated epidemics when run at an alert rate of 0.01

Amongst the algorithm variants that have a high specificity (i.e. > 90%), Farrington Flexible has the highest sensitivity and specificity, but it is not most timely