Using machine learning of urodynamics to predict clinical outcomes in patients with spina bifida
Rebecca Fairchild, BA, Leonid Aksenov, MD, Di Martino Matias, PhD, Wilkins Aquino, PhD, James Hokanson, PhD, Jonathan Routh, MD, MPH.
Duke University, Durham, NC, USA.
Introduction: Urodynamic studies (UDS) play a fundamental role in guiding optimal neurogenic bladder management in patients with spina bifida. Findings identified on UDS correlate with the development of adverse urinary outcomes such as renal scarring, recurrent urinary tract infections (UTI), and incontinence. However, to date, associations have been based on single UDS variables and have not encompassed the multitude of parameters provided by UDS testing. Furthermore, despite the clinical relevance of accurate UDS readings, there is documented heterogeneity in UDS interpretation between providers, which may hinder optimal management. Machine learning offers the potential to address these limitations by standardizing UDS interpretation and leveraging the complexity of the data to make predictive assessments. We have previously developed machine learning models that accurately predict detrusor overactivity from UDS data, demonstrating the viability of using machine learning on UDS data. The present study aims to build on our existing work by using machine learning to predict clinical outcomes based on urodynamic data in patients with spina bifida.
Methods: This study is based on clinical and UDS data from patients at a single pediatric urology clinic who underwent UDS testing between May 2012 and September 2020. Clinical data were retrospectively extracted from medical records and included demographics, medical history, UTI history, renal laboratory tests, imaging findings, surgical history, and current management. Data extracted from UDS files included UDS indication, bladder function parameters, provider interpretation, and effect on management. Logistic regression models were generated to predict getting a symptomatic UTI within one year of UDS testing and/or hydronephrosis within two years of testing. Hydronephrosis was included if it was Society for Fetal Urology (SFU) grade II-IV unilaterally or grade I-IV bilaterally. Models were trained on 75% of the data and 25% of the data was used for performance evaluation. Performance was compared between models with both UDS and clinical features and clinical features only.
Results: This study included 305 patients with total of 509 UDS tests. In this population, there were 191 instances of patients developing a UTI within 1 year of UDS testing and 185 cases of patients developing hydronephrosis within 2 years of testing. Machine learning models that incorporated UDS data predicted the development of UTIs (AUC=0.65, [0.53-0.74] 95%CI) and hydronephrosis (AUC=0.63 [0.53-0.73] 95%CI). Models that did not include UDS data failed to predict either UTIs or hydronephrosis (Figure).
Conclusion: Machine learning models based on urodynamic data can predict the likelihood of developing future UTIs and hydronephrosis. Models are improved by the inclusion of urodynamic information. The use of machine learning models is promising and may help guide clinical decision-making to prevent adverse urinary outcomes in individuals with spina bifida.
Back to 2022 Abstracts