Lower urinary tract urinary symptoms (LUTS), such as urinary urgency, frequency, and incontinence, affect the majority of the population at some point in the lifespan, causing substantial morbidity, yet few receive effective care. Acccurate diagnosis and treatment is usually dictated by the dominant symptom, however, the sizeable symptomatic overlap between disease categories and subjectivity of language used to describe symptoms leads to high rates of misdiagnosis. We hypothesized that more specific and homogeneous LUTS diagnoses are characterized not by specific pathognomic features, but by patterns of existing symptoms indicative of unique causes of convergent symptomatologies. To improve care and diagnostic accuracy, we sought to employ a data-driven approach to LUTS categorization using machine learning to generate diagnostic groupings based on patient-reported clinical data, creating a novel tool for diagnosis for patients with voiding complaints.

Questionnaire responses in a Development Dataset of 514 female subjects were used for model development using agglomerative hierarchical clustering. Resulting phenotypic clusters were assigned a clinical identity consistent with recognized causes of voiding dysfunction by consensus of three urologic specialists. A random forest classifier trained to assign unseen patients into these phenotypes was then applied to an independent cohort of 571 unselected, consecutive women presenting for urologic care to confirm reproducibility of that diagnostic algorithm. In this Validation Dataset, symptomatic questionnaires were used by the ML LUTS classifier to categorize subjects into the defined phenotypes. After association of phenotypes with accepted urologic diagnoses by specialist consensus, we examined concordance of the assigned phenotypes with coded diagnoses and specialist-assigned treatments.

Hierarchical clustering identified 4 major clusters and 9 specific phenotypes of LUTS capturing the overlapping symptoms inherent in typical patients. The derived algorithm recognized both common uncomplicated diagnoses (i.e., pelvic organ prolapse, overactive bladder) and several underrecognized diagnostic categories (i.e., myofascial pelvic pain).  In the Validation Dataset, a real-world cohort of care-seeking women, the ML LUTS classifier classified patients into the 9 phenotypic patterns representative of a broad spectrum of LUTS. The characteristic patterns of co-existing symptoms were congruent with the population used to train the classifier. Application of the ML LUTS classifier also facilitated improved recognition of often overlooked pelvic complaints, such as fecal incontinence. Diagnoses were consistent with coded diagnoses with 70% accuracy, but were in greater agreement with treatments assigned by specialist providers. These treatments correlated well with presumed etiologies of the ML phenotypes; for example, pelvic floor physical therapy was most commonly prescribed for myofascial pelvic pain while prolapse repair surgeries were exclusive restricted to the pelvic organ prolapse group.

We successfully applied machine learning algorithms to the diagnostic classification of women with a wide range of symptoms presenting for urologic care. This classification generated logical, phenotypic groups based on validated patient-reported symptoms alone. Symptomatic patterns could be grouped into four general clusters: minimal/mild symptoms, urogenital pain, storage urinary complaints, and pelvic floor disorders. Validation of these clusters revealed high reproducibility in an independent cohort with an accuracy of 71%. These groups are analogous to the general clinical categories currently used: most patients presenting to urogynecology clinics will be diagnosed with either incontinence, genitourinary pain, or pelvic organ prolapse. As unsupervised machine learning brings no assumptions to cluster derivation, agreement of the overall diagnostic categories with well-accepted clinical categorization validates the ability of data-driven methods to derive clinically meaningful diagnostic categories. There are, however, limitations to this simplistic categorization. Each cluster, as seen for the currently used symptom complexes such as overactive bladder (OAB) and interstitial cystitis/bladder pain syndrome (IC/BPS), likely encompasses multiple pathophysiologies requiring different treatments, limiting their utility in providing personalized, effective treatment. To circumvent these limitations, several groups have tried to subclassify urinary symptoms or genitourinary pain, but have typically examined only one symptom cluster in isolation (e.g., overactive bladder or genitourinary pain). While providing insight into the patterns of LUTS, many patients present with multiple urinary symptoms that do not perfectly fit these pre-established diagnoses. In addition, most of these classification approaches require detailed information (patient demographics, physical exam findings, imaging, genetic or biochemical markers, or other diagnostic testing results) unavailable or unfamiliar to most practitioners outside of specialized clinical settings. To achieve greater utility in a broad range of real-world clinical settings, a clinical decision support tool needs to account for the overlapping symptoms and co-existing pelvic organ prolapse that complicates our current diagnostic schema without requiring extensive clinical information that is difficult for non-specialty providers to obtain. To overcome this obstacle, broad inclusion of all patients consecutively presenting for urogynecologic care combined with unsupervised clustering using patient complaints alone allowed us to derive nine unique phenotypes encompassing the range of overlapping symptoms without bias. Distinction between groups was based on unique combinations of symptoms rather than individual, pathognomonic features.49 These nine phenotypic diagnoses included the range of common urologic diagnoses (stress urinary incontinence, urgency urinary incontinence/OAB, mixed urinary incontinence, IC/BPS), but also incorporated several less common, emerging pathologies that are frequently underrecognized in patients with LUTS (myofascial urinary frequency syndrome, myofascial pelvic pain, non-urologic pelvic pain). The classifier also distinguished between subjects with mixed urinary incontinence in whom a correctable, anatomic cause (pelvic organ prolapse) to their symptoms should be suspected, which may influence treatment choices. Lastly, the classifier was capable of recognizing highly impactful symptoms like fecal incontinence, which are frequently unaddressed as patients are often too embarrassed to express them. Thus, these resulting groups captured the ranges of coexisting symptoms while still accounting for the complicated symptomatic overlap of real-world patients, something no other ML categorization system has done thus far. These diagnostic groups were concordant with the subspecialist coded diagnoses for patients 70% of the time, but correlated even more frequently with the treatment assigned to the patient by the specialist. This discordance between diagnostic codes and clinical behavior supports the utility of the ML LUTS classifier to understand the subtler patterns of LUTS in real-world patients.

We describe the generation of a machine learning algorithm relying only on validated patient-reported symptoms for accurate diagnostic classification. Algorithm-based assignment of unseen subjects into LUTS categories demonstrated good reproducibility of the phenotypes and their symptomatic patterns in an independent care-seeking population.
Given a growing physician shortage and increasing challenges for patients accessing specialist care, this type of digital technology holds great potential to improve the recognition, diagnosis, and treatment of functional urologic conditions. This novel LUTS classification algorithm can be utilized to assign treatment plans without the need for either sub-specialist evaluation, to which access can be limited, or physical examination, which can be challenging for patients in underserved areas. While future prospective work with larger, multi-institutional cohorts is needed, with refinement, this approach is capable of increasing both the equity and rapidity of access to effective urologic care.

Continence 12S (2024) 101377
DOI: 10.1016/j.cont.2024.101377