Evaluation of the BD urine culture application for artificial intelligence-assisted urine culture interpretation

Artificial intelligence-based culture reading tools can potentially accelerate reading, improving reporting consistency in image-based interpretation, which is relevant for high-volume urine cultures. This study evaluated the BD Kiestra Urine Culture Application (UCA) in a centralized microbiology laboratory. Consecutive midstream and indwelling catheter urine specimens inoculated onto BD BBL CHROMagar Orientation were analyzed by the UCA. Growth score output at different reporting thresholds was evaluated, along with expert rule-based categories constructed using the app’s growth score, purity score, and presumptive identification parameters, with blinded technologist image reads as the comparator. Of 1,304 cultures, 1,251 were evaluable for growth score comparison, and 1,239 for expert-rule assessment. The exact agreement across four growth categories (<10 ⁵ , 10 ⁵ , 10 ⁶ , ≥10 ⁷ CFU/L) was 86.2%. Using a <10 ⁶ CFU/L triage threshold, the overall agreement was 95.9%, with positive and negative percent agreement of 89.9% and 98.9%, respectively. In this dataset, 30.5% of cultures met this threshold for potential auto-release or batch-review. For presumptive identification in ≥10 ⁷ CFU/L pure or predominant cultures, UCA performed best for E. coli (PPA 98.9%). For Enterococcus spp., PPA was 89.5%, but PPV was low (25.4%), reflecting frequent false-positive presumptive calls. UCA performed well at clinically relevant thresholds and could meaningfully reduce the burden of manual reading. Most discrepancies reflected differences between total colony count outputs and morphotype-based laboratory quantitation conventions, along with missed subtle growth. Expert-rule category discrepancies were primarily related to differences in purity output and presumptive identification, particularly fine light-blue growth being called presumptive Enterococcus spp., highlighting a priority area for improvement.