# Predicting the frequency of positive laboratory submissions for porcine reproductive and respiratory syndrome in Ontario, Canada, using autoregressive integrated moving average, exponential smoothing, random forest, and recurrent neural network

## TL;DR

This study compares different models to predict the spread of a pig disease in Ontario, finding that random forest models perform best but still have limitations.

## Contribution

The study evaluates and compares time-series models for predicting PRRSV-positive submissions, identifying random forest as the most accurate.

## Key findings

- Random forest models showed the lowest prediction errors for PRRSV submissions.
- ARIMA and ETS models tended to overpredict, while RF and RNN underpredicted.
- No consistent seasonal pattern was observed in PRRSV submissions over ten years.

## Abstract

Porcine Reproductive and Respiratory Syndrome Virus (PRRSV) is endemic in many pig-producing countries and poses significant health and economic challenges. Enhanced surveillance strategies are essential for effective disease management. This study aimed to evaluate and compare the performance of different time-series modeling techniques to predict weekly PRRSV-positive laboratory submissions in Ontario, Canada. Ten years of PRRSV diagnostic data were obtained from the Animal Health Laboratory at the University of Guelph and were processed into a weekly time series. The dataset was analyzed with autoregressive integrated moving average (ARIMA), exponential smoothing (ETS), random forest (RF), and recurrent neural network (RNN) models. Two validation strategies were employed: a traditional train-test split and a simulated prospective rolling forecast. Model accuracy was evaluated using common predictive error metrics. Descriptive analysis indicated a gradual increase in PRRSV positive submissions over time, with no consistent seasonal pattern. ARIMA and ETS models generally overpredict case counts, while RF and RNN tended to underpredict them. Among the evaluated models, the RF regression model most accurately captured the underlying time-series dynamics and produced the lowest prediction errors across both validation approaches. Despite outperforming other models, the RF model’s high relative prediction errors limit its suitability for accurate forecasting of PRRSV-positive submissions in Ontario’s routine surveillance system. Further data refinement and algorithm improvements are warranted.