A data driven clinical algorithm for differential diagnosis of pertussis and other respiratory infections in infants.

Background: Clinical criteria for pertussis diagnosis and clinical case definitions for surveillance are based on a cough lasting two or more weeks. As several pertussis cases seek care earlier, a clinical tool independent of cough duration may support earlier recognition. We developed a data-driven algorithm aimed at predicting a laboratory confirmed pertussis.

Methods: We enrolled children <12 months of age presenting with apnoea, paroxistic cough, whooping, or post-tussive vomiting, irrespective of the duration of cough. Patients underwent a RT-PCR test for pertussis and other viruses. Through a logistic regression model, we identified symptoms associated with laboratory confirmed pertussis. We then developed a predictive decision tree through Quinlan's C4.5 algorithm to predict laboratory confirmed pertussis.

Results: We enrolled 543 children, of which 160 had a positive RT-PCR for pertussis. A suspicion of pertussis by a physician (aOR 5.44) or a blood count showing leukocytosis and lymphocytosis (aOR 4.48) were highly predictive of lab confirmed pertussis. An algorithm including a suspicion of pertussis by a physician, whooping, cyanosis and absence of fever was accurate (79.9%) and specific (94.0%) and had high positive and negative predictive values (PPV 76.3% NPV 80.7%).

Conclusions: An algorithm based on clinical symptoms, not including the duration of cough, is accurate and has high predictive values for lab confirmed pertussis. Such a tool may be useful in low resource settings where lab confirmation is unavailable, to guide differential diagnosis and clinical decisions. Algorithms may also be useful to improve surveillance for pertussis and anticipating classification of cases.