# Draft Abstract

## Title Option 1

Development of an Explainable Machine Learning Pipeline for Urodynamic Classification and Structured Report Generation

## Title Option 2

Explainable AI for Pressure-Flow-Based Classification of Lower Urinary Tract Dysfunction: A Public Dataset Proof-of-Concept

## Abstract

**Background:** Lower urinary tract dysfunction is commonly assessed with pressure-flow studies, but AI efforts in this domain are limited by scarce public datasets, incomplete explainability, and the absence of standardized raw-trace benchmarks.  
**Objective:** To develop a modular, explainable proof-of-concept pipeline for urodynamic classification and structured clinician-facing report drafting using a public tabular dataset.  
**Methods:** We used a public CC BY 4.0 Mendeley dataset of 527 patients undergoing pressure-flow studies for lower urinary tract symptoms. The released MATLAB structures were flattened into a reproducible feature table containing demographic and pressure-flow summary variables. Because the annotation matrix contains overlapping diagnoses, we formulated the task as multilabel one-vs-rest classification rather than exclusive multiclass diagnosis. Logistic Regression, Random Forest, and gradient-boosted tree models were evaluated with multilabel-stratified train-test splitting and 5-fold cross-validation. To address feature leakage risk, we compared a full-feature analysis with a leakage-reduced analysis that removed the clearest label-proximal summary indices, specifically effective infused capacity and compliance. Explainability was assessed with tree SHAP when feasible, and model outputs were translated into templated clinician-reviewable report drafts.  
**Results:** In the current proof-of-concept run, the best model in both settings was a Random Forest. In the full-feature analysis, it achieved subset accuracy 0.495, macro F1 0.611, weighted F1 0.748, and macro AUROC 0.920. In the leakage-reduced analysis, the best model achieved subset accuracy 0.417, macro F1 0.574, weighted F1 0.696, and macro AUROC 0.868. The decline in performance after removal of the clearest label-proximal summaries suggests that part of the predictive signal in the public benchmark is embedded in engineered summary variables near established diagnostic constructs.  
**Conclusion:** This work demonstrates a modular, explainable, clinician-in-the-loop AI-UDS proof-of-concept rather than a clinically validated diagnostic system. The primary contribution is a transparent pipeline architecture and a leakage-aware evaluation framework that can support future IRB-enabled retrospective dataset creation, trace-level and PDF-derived data ingestion, and prospective clinical validation.

## Short Limitations Statement

This abstract should be interpreted in the context of a public tabular feature release rather than trace-level waveform data, overlapping multilabel annotations, and a leakage-reduced benchmark that is conservative but not equivalent to a fully primitive-variable analysis.

## Short Future Work Statement

The next scientific step is development of IRB-enabled retrospective datasets linking raw UDS traces, PDF reports, clinician review, and downstream outcomes, followed by external and prospective validation.
