CLASSIFICATION AND REGRESSION TREES (CART) ANALYSIS FOR PREDICTING INFLUENZA

Purpose: Despite the burden of influenza, the use of neuraminidase inhibiting anti-viral medication is relatively infrequent.  Rapid, cost-effective methods for determining the likelihood of influenza may help identify patients for whom antiviral medications will be most beneficial (high risk condition, ≥65 years old, and presenting for treatment within 48 hours of symptom onset).  Clinical decision algorithms are a rapid, inexpensive method to evaluate probability of influenza, but to date, most algorithms are based on regression analyses that do not account for higher order interactions.  This study used classification and regression trees (CART) modeling to estimate probabilities of influenza.

Method: 4,173 individuals ≥ 5 years of age who presented at ambulatory centers for treatment of acute respiratory illness (≤ 7 days) with cough or fever in 2011-2012 were included.  Eligible enrollees provided nasal and pharyngeal swabs for real-time, reverse transcriptase polymerase chain reaction (RT-PCR) testing for influenza, self-reported symptoms, personal characteristics and self-reported influenza vaccination status.  CART was used to develop a series of models with prediction success of sensitivity, specificity, positive predictive values (PPV), negative predictive values (NPV), and areas under the curve (AUC) calculated.

Result: 645 enrollees <65 years and 60 enrollees ≥65 years had PCR-confirmed influenza.  Antiviral medication was prescribed for 14% of those individuals.  Among nine possible clinical features, CART selected the best combination (fever, cough, fatigue, and shortness of breath and household smoke)with a sensitivity of 81%, specificity of 52%, PPV of 24%, NPV of 94% and AUC=0.68.  Limiting the sample to those 345 patients for whom antivirals are clearly recommended i.e., individuals <65 years with a high risk condition or ≥65 years, and who presented for care ≤2 days from symptom onset,  presence of fever and cough resulted in a prediction algorithm with 86% sensitivity, 47% specificity, 27% PPV, 95% NPV and AUC=0.67.

Conclusion: The algorithm based on CART recursive partitioning, among outpatients ≥5 years, was used to estimate probability of influenza with good sensitivity and high NPV, but low PPV in an influenza season with low prevalence of disease.  After further testing for seasons with higher influenza prevalence, CART may be used to exclude many who do not need antivirals, and indicate who should be considered for viral testing for confirmation of influenza.