Comparative evaluation of time series models for predicting influenza outbreaks: application of influenza-like illness data from sentinel sites of healthcare centers in Iran

• Published in: BMC research notes Vol. 12; no. 1; p. 353
• Main Authors: Tapak, Leili, Hamidi, Omid, Fathian, Mohsen, Karami, Manoochehr
• Format: Journal Article
• Language: English
• Published: England BioMed Central Ltd 24-06-2019; BioMed Central; BMC
• Subjects: Artificial neural networks; Care and treatment; Comparative analysis; Diagnosis; Disease Outbreaks - statistics & numerical data; Epidemics; Forecasts and trends; Health education; Health Facilities; Humans; Influenza; Influenza, Human - epidemiology; Iran; Iran - epidemiology; Methods; Models, Statistical; Neural network; Outbreak; Prevention; Public health; Public health surveillance; Random Forest; Research Note; Sentinel surveillance; Support vector machine; Time Factors; Time series analysis; Iran; Outbreak; Neural network; Support vector machine; Public health surveillance; Influenza; Random Forest
• ISSN: 1756-0500; 1756-0500
• DOI: 10.1186/s13104-019-4393-y

Forecasting the time of future outbreaks would minimize the impact of diseases by taking preventive steps including public health messaging and raising awareness of clinicians for timely treatment and diagnosis. The present study investigated the accuracy of support vector machine, artificial neural-network, and random-forest time series models in influenza like illness (ILI) modeling and outbreaks detection. The models were applied to a data set of weekly ILI frequencies in Iran. The root mean square errors (RMSE), mean absolute errors (MAE), and intra-class correlation coefficient (ICC) statistics were employed as evaluation criteria.
It was indicated that the random-forest time series model outperformed other three methods in modeling weekly ILI frequencies (RMSE = 22.78, MAE = 14.99 and ICC = 0.88 for the test set). In addition neural-network was better in outbreaks detection with total accuracy of 0.889 for the test set. The results showed that the used time series models had promising performances suggesting they could be effectively applied for predicting weekly ILI frequencies and outbreaks.