AI Shows Promise for Stone-Linked Urinary Infections, But Validation Lags

Urinary stones and urinary tract infections make a dangerous combination. A stone can block drainage, shelter bacteria in biofilm, and turn a routine infection into urosepsis within hours, all while clinicians wait for culture results that arrive too late to guide the first critical decisions. A new narrative review published in the African Journal of Urology argues that artificial intelligence could help clinicians navigate this pathway, but only if the technology is framed as structured risk integration rather than a standalone diagnostic oracle, and only if the field confronts a stubborn gap between promising model performance and validated clinical impact.

The review, led by Andrea Carolina Ordonez-Zabala and colleagues, searched PubMed/MEDLINE, Scopus, Web of Science, and Google Scholar from database inception to 31 July 2026, supplemented by reference chaining and current clinical and reporting guidelines. The authors deliberately conducted a structured narrative review rather than a systematic review or meta-analysis, and they were explicit about why: the available studies address different populations, inputs, model tasks, validation strategies, and outcomes, making quantitative pooling inappropriate without stricter grouping. Instead, the synthesis organizes artificial intelligence applications into a clinically oriented framework spanning recognition of infected obstruction, estimation of deterioration risk, selection of empiric therapy, timing of drainage, postoperative monitoring, and governance before deployment.

A central distinction runs through the review: the difference between direct and indirect evidence. Direct evidence comes from models developed in patients with stones and a concurrent infectious target, such as kidney stone-associated urinary tract infection, infection-stone composition, urosepsis with upper urinary tract calculi, or infectious complications after stone surgery. Indirect evidence covers models for stone detection, segmentation, volumetry, metabolic recurrence, or antibiotic selection in general urinary infection populations. The authors warn that the reported performance of these adjacent tools cannot be transferred to infected obstruction or stone-associated complicated infection without dedicated validation, a caveat that guards against one of the most common overclaims in the medical artificial intelligence literature.

The most directly relevant work involves radiomics, the extraction of quantitative features from computed tomography images that go beyond what radiologists typically report in routine interpretation. Texture, density distribution, shape, and perilesional features may reflect stone composition, bacterial colonization, obstruction, or inflammation. Lu and colleagues reported that a combined radiomics-clinical model could discriminate kidney stone-associated urinary tract infection, highlighting the potential of CT-derived features to support risk classification when conventional clinical features fall short. Earlier radiomics work on urinary infection stones supports the premise that CT-derived features can contribute to noninvasive identification of infection-related stone composition, which matters because infection stones formed in association with urease-producing organisms may require different preventive and therapeutic reasoning than metabolic stones.

Yet the review is candid about the technical fragility of these approaches. Scanner protocols, reconstruction kernels, segmentation methods, contrast use, and feature extraction pipelines can all change model performance, and a model trained at one institution may not generalize to another unless preprocessing and validation are carefully standardized. Methodological critiques of radiomics in urolithiasis have emphasized that promising discrimination does not automatically imply deployability; reproducibility, segmentation stability, calibration, and external validation remain decisive. The authors conclude that radiomics should be viewed as a candidate layer in risk prediction, not as a stand-alone diagnostic test.

Perioperative prediction forms another evidence cluster. Endourological procedures such as ureteroscopy and percutaneous nephrolithotomy carry a feared risk of postoperative febrile urinary tract infection, systemic inflammatory response syndrome, and sepsis, with risk shaped by preoperative urine culture, stone size and composition, hydronephrosis, operative time, intrarenal pressure, comorbidities, prior infection, and local antimicrobial resistance. Pietropaolo and colleagues used machine learning to predict post-ureteroscopy urosepsis requiring intensive care in a multi-center European case-control study. Shen and colleagues reported a machine learning model for sepsis prediction after percutaneous nephrolithotomy. Li and colleagues developed a CT-value-based model for postoperative systemic inflammatory response after endoscopic kidney stone surgery, with SHAP-based interpretation of predictors, and Hong and colleagues proposed an artificial intelligence model based on ultrasonography and urinalysis for urosepsis in patients with upper urinary tract calculi. Together, the review suggests, these studies sketch a plausible clinical trajectory, but they still require prospective validation before routine deployment.

The review also draws a careful line between outcomes that are often conflated. Pre-existing urosepsis in a patient with an upper-tract calculus is an acute diagnostic or triage problem; postoperative febrile urinary tract infection is defined primarily by fever and urinary infection; systemic inflammatory response syndrome is a physiological response that is not synonymous with infection; and postoperative urosepsis is infection-associated organ dysfunction after intervention. Models developed for one outcome should not be assumed to predict another, even when their predictors overlap. This granularity matters clinically, because a high-risk prediction may trigger admission, broad-spectrum antibiotics, drainage, or transfer, while a low-risk prediction may support discharge or delayed intervention, so miscalibration carries real consequences in both directions.

On antimicrobial stewardship, the review highlights decision-support tools that combine patient history, prior cultures, local antibiograms, comorbidity data, and guideline logic to reduce antibiotic mismatch when integrated into prescribing workflows. A recent artificial intelligence-driven urinary infection decision-support study reported reduced antibiotic mismatches in outpatient care by integrating resistance prediction and prescribing workflows. The authors stress that this study was not limited to stone-associated infection, and that an antimicrobial model validated in general urinary infection cannot be assumed to perform equally well in obstructed, instrumented, or postoperative stone populations without local validation. Stone patients, with their frequent prior cultures, prior antibiotics, and healthcare exposure, remain a natural target for individualized antimicrobial support, but the extrapolation must stay explicit.

The implementation gap, the authors argue, is larger than the performance gap. Many models show promising accuracy in development datasets but never reach routine care, blocked by data quality problems, interoperability failures, workflow friction, lack of external validation, unclear responsibility, and limited explainability. Reporting frameworks such as TRIPOD+AI for prediction models and CONSORT-AI and SPIRIT-AI for interventional tools directly address these weaknesses by requiring clear descriptions of inputs, outputs, human-AI interaction, integration setting, and error analysis. The proposed clinical framework has five layers: data capture, risk estimation, actionable output, human oversight with clinician review and override documentation, and governance through external validation, bias audit, performance monitoring, and periodic recalibration.

For African settings, the review argues that the potential value lies less in importing complex models and more in supporting consistent decisions where specialist, imaging, and microbiology capacity varies. The World Health Organization African Region has identified inadequate infrastructure, equipment, financing, workforce, and subnational access as major constraints on diagnostic and laboratory services. Models built on routinely available variables such as temperature, diabetes, urinalysis, creatinine, hydronephrosis, prior cultures, and stone burden may be more feasible than radiomics pipelines requiring standardized CT acquisition and substantial computing infrastructure, and ultrasound-based triage may be particularly relevant where CT access is constrained, though it still requires local prospective validation and must not delay referral or decompression in suspected infected obstruction. Regional assessments document continuing gaps in antimicrobial-resistance surveillance and laboratory capacity, and East African urinary infection data show that treatment pathways and multidrug resistance vary materially across settings. The authors’ bottom line is measured: artificial intelligence in this domain is promising, not practice-changing by itself, and should for now be framed as an adjunct to clinical judgment, microbiology, imaging, stewardship, and timely urological source control, with urgent drainage never delayed when infected obstruction is suspected.

Subject of Research: Artificial intelligence applications in imaging, urosepsis prediction, and antimicrobial stewardship for stone-associated complicated urinary tract infection

Article Title: Artificial intelligence across the stone-associated urinary infection pathway: imaging, urosepsis prediction, and antimicrobial stewardship

Article References: Ordonez-Zabala, A. C., Rey-Contreras, A. V., & Arrazola-Villadiego, M. A. (2026). Artificial intelligence across the stone-associated urinary infection pathway: imaging, urosepsis prediction, and antimicrobial stewardship. African Journal of Urology, 32(1), Article 76. https://doi.org/10.1186/s12301-026-00623-0

Image Credits: AI Generated

DOI: 10.1186/s12301-026-00623-0

Keywords: artificial intelligence, machine learning, radiomics, urinary stones, urosepsis, urinary tract infection, antimicrobial stewardship, computed tomography, ureteroscopy, percutaneous nephrolithotomy, external validation, African urology