Prediction of symptomatic and asymptomatic bacteriuria in spinal cord injury patients using machine learning.

backgroundIndividuals with spinal cord injuries (SCI) frequently rely on urinary catheters to drain urine from the bladder, making them susceptible to asymptomatic and symptomatic catheter-associated bacteriuria and urinary tract infections (UTI). Current identification of these conditions lacks precision, leading to inappropriate antibiotic use, which promotes selection for drug-resistant bacteria. Since infection often leads to dysbiosis in the microbiome and correlates with health status, this study aimed to develop a machine learning-based diagnostic framework to predict potential UTI by monitoring urine and/or catheter microbiome data, thereby minimising unnecessary antibiotic use and improving patient health.

resultsMicrobial communities in 609 samples (309 catheter and 300 urine) with asymptomatic and symptomatic bacteriuria status were analysed using 16S rRNA gene sequencing from 27 participants over 18 months. Microbial community compositions were significantly different between asymptomatic and symptomatic bacteriuria, suggesting microbial community signatures have potential application as a diagnostic tool. A significant decrease in local (alpha) diversity was noted in symptomatic bacteriuria compared to the asymptomatic bacteriuria (P < 0.01). Beta diversity measured in weighted unifrac also showed a significant difference (P < 0.05) between groups. Supervised machine learning models were trained on amplicon sequence variant (ASVs) counts and bacterial taxonomic abundances (Taxa) to classify symptomatic and asymptomatic bacteriuria with a repeated tenfold and leave-one-out participant (LOPO) type of cross-validation approaches. Combining urine and catheter microbiome data improved the model performance during repeated tenfold cross-validation, yielding a mean area under the receiver operating characteristic curve (AUROC) of 0.95 (95% CI 93-0.97) and 0.83 (95% CI 0.79-0.89) for ASVs and taxonomic features in the independent held-out test set, respectively. The LOPO cross-validation yielded a mean AUROC of 0.87 (95% CI 0.85-0.89) and 0.79 (95% CI 0.77-0.82) for ASVs and taxa features, respectively. These results suggest the potential of microbiome features in differentiating symptomatic and asymptomatic bacteriuria states.

conclusionsOur findings demonstrate that signatures within catheter and urine microbiota could serve as tools to monitor the health status of SCI patients. Establishing a classification system based on these microbial signatures could equip physicians with alternative management strategies, potentially reducing UTI episodes and associated hospital costs, thus significantly improving patient quality of life while mitigating the impact of drug-resistant UTI. Video Abstract.