Risk Classification for Interstitial Cystitis/Bladder Pain Syndrome Using Machine Learning Based Predictions

Abstract

OBJECTIVESTo improve diagnosis of IC/BPS(IC) we hereby developed an improved IC risk classification using machine learning algorithms. METHODSA national crowdsourcing resulted in 1,264 urine samples consisting of 536 IC (513 female, 21 male, 2 unspecified), and 728 age-matched controls (318 female, 402 male, 8 unspecified) with corresponding PRO pain and symptom scores. In addition, 296 urine samples were collected at three academic centers: 78 IC (71 female, 7 male) and 218 controls (148 female, 68 male, 2 unspecified). Urinary cytokine biomarker levels were determined using Luminex assay. A machine learning predictive classification model, termed the Interstitial Cystitis Personalized Inflammation Symptom (IC-PIS) Score, that utilizes PRO and cytokine levels, was generated and compared to a challenger model. RESULTSThe top-performing model using biomarker measurements and PROs (AUC=0.87) was a support vector classifier, which scored better at predicting IC than PROs alone (AUC=0.83). While biomarkers alone (AUC=0.58) did not exhibit strong predictive performance, their combination with PROs produced an improved predictive effect. CONCLUSIONSIC-PIS represents a novel classification model designed to enhance the diagnostic accuracy of IC/BPS by integrating PROs and urine biomarkers. The innovative approach to sample collection logistics, coupled with one of the largest crowdsourced biomarker development studies utilizing ambient shipping methods across the US, underscores the robustness and scalability of our findings.