Machine learning models provide modest accuracy in predicting clinical impact of porcine reproductive and respiratory syndrome type 2 in Canadian sow herds.

Abstract

Objective: To determine the predictive potential of the open reading frame 5 nucleotide sequence of porcine reproductive and respiratory syndrome (PRRS) virus and the basic demographic data on the severity of the impact on selected production parameters during clinical PRRS outbreaks in Ontario sow herds.

Methods: A retrospective longitudinal study of clinical outbreaks in Ontario sow herds at various points between September 5, 2009, and February 5, 2019, was conducted using herds as units of analysis. Data were gathered from study sow farms in Ontario at the start of each clinical outbreak. Six machine learning models and 2 different genetic input structures of open reading frame 5 sequences were utilized to predict the impact on abortion and preweaning mortality.

Results: Extreme boosting machine learning models with genetic data represented through 2-dimensional multiple correspondence analysis had the highest accuracy when predicting clinical outcomes (60.8% [SD = 12.4%] and 74.4% [SD = 13.2%]) for abortion and preweaning mortality outcomes, respectively. The mean sensitivity of classifying outbreaks with a high impact on abortion was 50%, with a specificity of 89.2%. The mean sensitivity of classifying outbreaks with high preweaning mortality was 56.2%, with a specificity of 85.2%.

Conclusions: The data and methods utilized herein exhibited improvement in accuracy over the baseline; however, this increase was not sufficient to warrant field implementation.

Clinical relevance: Predictive models based on observed data could assist practitioners in linking the genetics of the PRRS virus with clinical impact in clinical settings. Models trained in this study show promise for PRRS clinical impact prediction.