PURPOSE

To identify predictive factors for satisfactory treatment outcome of the patients with IC/BPS using urine biomarkers and machine-learning models.

METHODS

The IC/BPS patients were prospectively enrolled and provide urine samples. The targeted analytes included inflammatory cytokines, neurotrophins, and oxidative stress biomarkers. The patients with overall subjective symptom improvement of ≥ 50% were considered to have satisfactory results. Binary logistic regression, receiver-operating characteristic (ROC) curve, machine-learning decision tree, and random forest models were used to analyze urinary biomarkers to predict satisfactory results.

RESULTS

Altogether, 57.4% of the 291 IC/BPS patients obtained satisfactory results. The patients with satisfactory results had lower levels of baseline urinary inflammatory cytokines and oxidative biomarkers than patients without satisfying results, including interleukin-6, monocyte chemoattractant protein-1 (MCP-1), C-X-C motif chemokine 10 (CXCL10), oxidative stress biomarkers 8-hydroxy-2'-deoxyguanosine (8-OHDG), 8-isoprostane, and total antioxidant capacity (TAC). Logistic regression and multivariable analysis revealed that lower levels of urinary CXCL10, MCP-1, 8-OHDG, and 8-isoprostane were independent factors. The ROC curve revealed that MCP-1 level had best area under curve (AUC: 0.797). In machine-learning decision tree model, combination of urinary C-C motif chemokine 5, 8-isoprostane, TAC, MCP-1, and 8-OHDG could predict satisfactory results (accuracy: 0.81). The random forest model revealed that urinary 8-isoprostance, MCP-1, and 8-OHDG levels had the most important influence on accuracy.

CONCLUSION

Machine learning decision tree model provided a higher accuracy for predicting treatment outcome of patients with IC/BPS than logistic regression, and levels of 8-isoprostance, MCP-1, and 8-OHDG had the most important influence on accuracy.