Spina bifida (SB) is the most common permanently disabling neurological birth defect in the United States. Neurogenic bladder (NGB) is the most common urological sequela of SB, and NGB has been shown to increase a patient’s risk of urinary tract infections (UTIs), pyelonephritis, hydronephrosis, and renal damage.^1,2

To prevent these adverse urological outcomes, clinicians rely on urodynamic testing (UDS) as part of effective SB management. UDS has been employed for decades to detect abnormal voiding and urine storage patterns. More specifically, UDS is used to detect neurogenic detrusor overactivity (NDO), detrusor sphincter dyssynergia, and bladder noncompliance causing dangerously high bladder storage pressures.^1-3 UDS thus forms the basis of modern, proactive bladder management.¹ However, despite the seeming clinical importance of UDS, there are inconsistencies and significant heterogeneity in UDS performance and interpretation; these inconsistencies limit the utility of UDS in terms of management and predictive capacity of long-term urological outcomes.⁴ For example, despite high levels of UDS expertise across centers participating in the CDC-funded UMPIRE (Urologic Management to Preserve Initial REnal function) protocol study sites—which utilize a UDS-based risk stratification to determine medical management for infants with myelomeningocele—there was noted to be only 50% agreement with the initial local UDS interpretation.⁵ As the cornerstone of NGB assessment, these variations in interpretations lead to life-altering changes in treatment of individuals with SB. The complexity of UDS data and inconsistency in interpretation make it an apt application for machine learning.

Machine learning is a data science technique in which models are iteratively trained to recognize complex patterns in data. There are many forms of machine learning, which can be alternatively referred to as artificial intelligence, deep learning, or multiple other similar techniques and concepts. Within urology, machine learning has been leveraged across a wide variety of topics, including urologic oncology, sexual health, UTI prediction and diagnostics, and imaging interpretation.^6,7 Recent efforts by our group and others have established models that detect NDO or accurately categorize NGB risk categories with a reasonably high degree of accuracy.^8-10 Although machine learning models based on UDS data are viable, they have not yet been used to predict the most clinically relevant outcomes such as UTI, hydronephrosis, or, most importantly, renal damage. As such, our group has focused our recent efforts on using machine learning techniques to build more predictive models for the identification of UTI and hydronephrosis. As an exciting step in the right direction, we have recently shown that combining UDS and clinical data into merged models was found to perform better than clinical models alone, with the best models predicting worsening hydronephrosis noting an AUC of 0.70 (unpublished data).

These results are certainly promising, but much work still remains to be done before machine learning-based models are ready for day-to-day clinical use. Further potential applications such as integration of imaging data and bladder appearance, standardization of compliance calculation metrics, and refinement of NDO definitions (for example, based on the amplitude of slope of the contraction curve) based on UDS variables will help to further define risk stratification based on both known and previously unrecognized UDS tracing data. Limitations also exist within machine learning models in general; a particularly fundamental limitation remains that most current UDS machine learning models are based on supervised learning techniques, where human judgment and interpretation are assumed to represent the “truth.” However, this is the exact same human judgment that we know to be problematic based on lack of consistency in UDS interpretation!⁴ Further, UDS machine learning models are unable to account for what humans don’t know to look for, the so-called “unknown unknowns” of data science. Lastly, it is imperative to keep in mind what data inputs or variables are used in training these predictive models. The concept of “garbage in, garbage out” remains a fundamental truth of prediction modeling; in other words, if machine learning algorithms are trained on bad data, they are doomed to make bad predictions.

UDS, despite its shortcomings, remains our best tool to identify and prevent renal deterioration in NGB, particularly in children and adults with SB. However, UDS interpretation is not objectively reproducible, and it thus is in serious need of improvement. Machine learning has the potential to help accomplish this goal by augmenting our understanding of UDS data and streamlining interpretation. While it is not yet ready for prime time, the future of machine learning in UDS is bright. Stay tuned.