Reassessing the management of uncomplicated urinary tract infection: A retrospective analysis using machine learning causal inference

Noah C Jones, Ming‐Chieh Shih, Elizabeth Healey, Chen Wen Zhai, Sonali D Advani, Aaron Smith–McLallen, David Sontag, Sanjat KanjilalmedRxiv 2024 · DOI

Abstract

Our study provides support for the continued use of first-line antibiotics for the management of uncomplicated UTI. Our results also provide proof-of-principle that automated feature extraction methods for OMOP formatted data can emulate manually curated models, thereby promoting reproducibility and generalizability.

Introduction

Urinary tract infections (UTIs) are a ubiquitous clinical challenge, affecting approximately half of all women during their lifetime and serving as a primary driver for outpatient antibiotic prescriptions. Current clinical practice is largely guided by the 2011 Infectious Diseases Society of America (IDSA) recommendations. However, in the decade since their publication, the medical landscape has evolved significantly, characterized by emerging pathogen resistance, shifting patient demographics, and changing prescribing patterns.

A common clinical trend involves the use of "second-line" agents, such as fluoroquinolones (e.g., Ciprofloxacin), under the assumption that they offer superior potency or are necessary to combat suspected resistance to first-line treatments. This practice persists despite the higher risk profile of these drugs. This study evaluates whether the 2011 guidelines remain effective in contemporary practice or if a paradigm shift is required.

Methods

We conducted a retrospective cohort study utilizing the Independence Blue Cross claims database, encompassing approximately 3 million individuals in the Philadelphia region from 2012 to 2021. The final cohort consisted of 57,585 UTI episodes among 49,037 non-pregnant adult females. Data were standardized using the Observational Medical Outcomes Partnership (OMOP) Common Data Model to ensure reproducibility.

Study design timeline illustrating the extraction of static and time-windowed features relative to UTI diagnosis (T0) and subsequent treatment and outcome monitoring.

To address the inherent selection bias in real-world clinical data, we employed machine learning-based causal inference. We used Inverse Probability of Censoring Weighting (IPCW) to account for patient attrition and Propensity Score Weighting to adjust for confounding by indication. Three machine learning architectures—Logistic Regression, Random Forests, and Light Gradient Boosted Machines (LightGBM)—were evaluated, with LightGBM demonstrating superior performance.

The causal inference pipeline, showing cohort definition, the parallel building of treatment and censoring models using machine learning, and the final computation of average treatment effects.

A key methodological innovation was the comparison of two feature engineering strategies:

  • Expert-Derived: 245 features manually curated by infectious disease specialists.
  • Automated (OMOP-learn): Over 140,000 features automatically extracted from the database.

Results

Efficacy and the "Stronger" Drug Fallacy

Contrary to the perception that second-line drugs are more effective, our analysis revealed that first-line antibiotics (Nitrofurantoin and TMP-SMX) actually resulted in lower treatment failure rates. Patients receiving first-line therapies had a 2.1% lower absolute risk of a 30-day revisit compared to those on second-line fluoroquinolones. The disparity was even more pronounced when compared to alternative $\beta$-lactams, where first-line agents showed a 6.6% lower risk of failure. This may be attributed to the higher genetic barrier to resistance for first-line agents compared to the single-gene mutations that often confer fluoroquinolone resistance.

Safety and Adverse Events

Adverse event profiles were generally comparable across treatment groups. While first-line agents were associated with a marginal increase in skin rash (+0.4%), they demonstrated a significant safety advantage regarding acute kidney injury, with a 2.7% lower risk compared to alternative treatments.

Validation of Automated Feature Engineering

A significant finding for the field of clinical informatics was the performance of the automated OMOP-learn models. These models produced results nearly identical to those derived from expert-curated feature sets. This suggests that complex causal inference in clinical research can be effectively semi-automated, facilitating faster and more scalable evidence generation.

Discussion

While these results strongly support the continued use of 2011 IDSA guidelines, certain limitations must be noted. The study population, drawn from a single insurer in Southeast Pennsylvania, may not fully represent global resistance patterns or diverse socioeconomic backgrounds. Furthermore, the exclusion criteria for "uncomplicated" UTIs resulted in a cohort with a lower comorbidity profile than the general population.

Future research should focus on replicating this methodology across larger, more geographically diverse datasets to confirm the generalizability of these findings and to further refine automated feature extraction pipelines for clinical decision support.

Summary updated

Abstract

Our study provides support for the continued use of first-line antibiotics for the management of uncomplicated UTI. Our results also provide proof-of-principle that automated feature extraction methods for OMOP formatted data can emulate manually curated models, thereby promoting reproducibility and generalizability.
BibTeX
@misc{jones2024reassessing,
  author = {Noah C Jones and Ming‐Chieh Shih and Elizabeth Healey and Chen Wen Zhai and Sonali D Advani and Aaron Smith–McLallen and David Sontag and Sanjat Kanjilal},
  title = {{Reassessing the management of uncomplicated urinary tract infection: A retrospective analysis using machine learning causal inference}},
  year = {2024},
  doi = {10.1101/2024.08.18.24312104},
  howpublished = {medRxiv},
}
Figure