TITLE:
Development and Validation of an Objective Risk Scoring System for Assessing the Likelihood of Virus Introduction in Porcine Reproductive and Respiratory Syndrome Virus-Free Sow Farms in the US
AUTHORS:
Derald J. Holtkamp, Hui Lin, Chong Wang, Dale D. Polson
KEYWORDS:
Porcine Reproductive and Respiratory Syndrome (PRRS); Group Lasso; Logistic Regression; Risk Scoring System; Area under the ROC Curve
JOURNAL NAME:
Open Journal of Veterinary Medicine,
Vol.3 No.2,
June
6,
2013
ABSTRACT:
The lack of validated tools to predict how
long sow farms will remain PRRS virus-free following successful elimination of
the virus has deterred veterinarians and producers from attempting to eliminate
the PRRS virus from sow farms. The aim of this study was to use the database of
PRRS Risk Assessments for the Breeding Herd in PADRAP to develop and validate
an objective risk scoring system for predicting the likelihood of virus
introduction in PRRS virus-free sow farms in the US. To overcome the
challenges of dealing with a large number of variables, group lasso for
logistic regression (GLLR) was applied to a retrospective dataset of PRRS Risk
Assessment for the Breeding Herd surveys completed for 704 farms to develop the
risk scoring system. The validity of the GLLR risk scoring system
was then evaluated by testing its predictive ability on a dataset from a
long-term prospective study of 196 sow farms to assess risk factors associated
with how long PRRS virus-free sow farms remained PRRS virus-free. Receiver
operator characteristic(ROC) curves were estimated to compare the performance
of the GLLR risk scoring system to the risk scoring system based on expert
opinion (EO), currently used in the PRRS Risk Assessment for the Breeding Herd, for
predicting whether herds remained PRRS virus-free for 130 weeks. The GLLR risk
scoring system (AUC, 0.76; 95% CI, 0.67 - 0.84)
performed significantly better than the EO risk scoring system (AUC, 0.36; 95%
CI, 0.27 - 0.46) for predicting whether to sow farms in the
prospective study survived for 130 weeks (p 0.001). Dividing
farms into 3 risk groups (low, medium and high) using a low and high cutoff
values for the GLLR risk score was informative as the differences in the KM
survival curves for the 3 groups were both clinically meaningful and
statistically significant. The GLLR risk scoring system used in conjunction
with the PRRS Risk Assessment for the Breeding Herd survey delivered through
PADRAP appears to have the potential to help veterinarians predict the
likelihood of virus introduction in PRRS virus-free sow farms in the US.