Table 1: Random sub-sampling validation results averaged over 50 trials, for well turbidity analysis and subsequent determination of MIC and microbial drug susceptibility with very major (i.e., resistant microbes diagnosed as susceptible), major (i.e., susceptible microbes diagnosed as resistant), and minor (i.e., indeterminate/susceptible dose dependent-related misdiagnoses) error percentages for our algorithm using (a) a single exposure image and (b) a combination of three exposures.

From: High-throughput and automated diagnosis of antimicrobial resistance using a cost-effective cellphone-based micro-plate reader

Images usedWell AccuracyWell SensitivityWell SpecificityMIC AccuracyDrug Susceptibility AccuracyVery Major Error PercentageMajor Error PercentageMinor Error Percentage
(a) Single exposure (1/800)96.94 ± 1.30%98.83 ± 0.70%96.24 ± 1.92%94.89 ± 1.14%98.55 ± 0.58%0.43 ± 0.68%0.17 ± 0.13%1.23 ± 0.54%
(b) Combination of three exposures (1/1600, 1/1250, 1/800)98.21 ± 0.29%98.56 ± 0.37%98.08 ± 0.37%95.12 ± 0.87%99.23 ± 0.23%0 ± 0%0.16 ± 0.18%0.65 ± 0.20%
  1. 78 patient isolate testing plates from the UCLA Hospital Microbiology Lab and 21 blank plates without microbial content were randomly separated 50 times into training sets of 21 blank plates and 39 patient plates and blind-test sets of 39 patient plates. The threshold is determined using the training set and run on the corresponding test set, with mean and standard deviation across the 50 trials shown in the table. Combining multiple image exposures significantly increases the overall accuracy and reduces the variability for well turbidity detection, with corresponding improvements for MIC determination and drug susceptibility interpretation as well as significant reductions for very major, major, and minor errors.