Background: Among women who present with urinary complaints, only 50% are found to have urinary tract infection. Individual urinary symptoms and urinalysis are not sufficiently accurate to discriminate those with and without the diagnosis.
Methods: We used artificial neural networks (ANN) coupled with genetic algorithms to evolve combinations of clinical variables optimized for predicting urinary tract infection. The ANN were applied to 212 women ages 19-84 who presented to an ambulatory clinic with urinary complaints. Urinary tract infection was defined in separate models as uropathogen counts of > or =10(5) colony-forming units (CFU) per milliliter, and counts of > or =10(2) CFU per milliliter.
Results: Five-variable sets were evolved that classified cases of urinary tract infection and non-infection with receiver-operating characteristic (ROC) curve areas that ranged from 0.853 (for uropathogen counts of > or =10(5) CFU per milliliter) to 0.792 (for uropathogen counts of > or =10(2) CFU per milliliter). Predictor variables (which included urinary frequency, dysuria, foul urine odor, symptom duration, history of diabetes, leukocyte esterase on urine dipstick, and red blood cells, epithelial cells, and bacteria on urinalysis) differed depending on the pathogen count that defined urinary tract infection. Network influence analyses showed that some variables predicted urine infection in unexpected ways, and interacted with other variables in making predictions.
Conclusions: ANN and genetic algorithms can reveal parsimonious variable sets accurate for predicting urinary tract infection, and novel relationships between symptoms, urinalysis findings, and infection.