An interpretation algorithm for molecular diagnosis of bacterial vaginosis in a maternity hospital using machine learning: proof-of-concept study

Allplex Bacterial vaginosis assay (Seegene, South Korea) is a molecular test for bacterial vaginosis (BV). A machine learning algorithm was devised on 200 samples (BV = 23, non-BV = 177) converting 7 identified bacterial strains polymerase chain reaction results to binary output of BV detected or not. Comparing algorithm interpretation of molecular results to the consensus Gram stain (Hay's criteria), the sensitivity was 65% [95% confidence interval (CI) 42–83%], specificity was 98% (95% CI 95–99%), positive predictive value was 83% (95% CI 58–96%), and negative predictive value was 95% (91–98%) with area under the curve of 0.82 (95% CI 0.76–0.87). For the second phase, 100 samples were processed using the 2 techniques in parallel, with the scientists blinded to the result of the other method. There was agreement 90% of the cases (n = 90/100). The samples that were called BV by the algorithm but non-BV by Gram stain all cluster with the concordant BV samples, suggesting that the molecular test was correct.