Presented by: Jennifer Bjazevic MSc, MD, FRCSC
Division of Urology, Department of Surgery, Western University, London, Canada

Introduction

Preventative strategies and surgical treatment for urolithiasis depend on stone composition. However, stone composition is often unknown until the stone is passed or surgically managed. Given that stone composition likely reflects the physiological parameters during its formation, we used clinical data from stone formers to predict calcium vs non-calcium stone composition.

Materials

Stone composition, 24-hour urine results, serum biochemistry, and demographics were prospectively collected from calcium (n = 625) and non-calcium (n = 152) stone patients at a tertiary care centre metabolic stone clinic. Binary classification of calcium vs non-calcium composition was performed using a gradient boosted tree algorithm (Figure 1A). This algorithm converts multiple weak learners to strong learners to better classify stone types. Class imbalance was addressed by upsampling the minority class and hyperparameters were tuned using Bayesian optimization.

Results

,

Our model showed acceptable performance with an area under the receiver operator characteristics (AUC-ROC) curve of 0.79 (Figure 1B). The model had a good degree of sensitivity of 0.86 and a moderate degree of sensitivity of 0.56 (Figure 1C). The model demonstrated that 24-hour urine calcium and creatinine, blood phosphate and urate, and BMI were the most significant predictors of classification (Figure 1D). Sex, urine dipstick results, and blood parathyroid levels were the least important predictors in the model (Figure 1D).

Conclusion

We have demonstrated that clinical data can be used to predict stone composition, which may help urologists determine stone type and guide their management plan before stone treatment. Moreover, the model provides a better understanding of key clinical features of stone disease, which sheds light on the underlying pathophysiology. By extending machine learning algorithms, it will be possible to determine specific compositions of stones and ultimately improve medical therapy for stone formers.

Funding

None

Co-Authors

John Chmiel, BMSc, MSc
Department of Microbiology & Immunology, Western University, London, Canada

Gerrit Stuivenberg, BMSc
Department of Microbiology & Immunology, Western University, London, Canada

Jennifer Wong, MD
Division of Urology, Department of Surgery, Western University, London, Canada

Linda Nott, RN
Division of Urology, Department of Surgery, Western University, London, Canada

Jeremy Burton, BSc, MSc, PhD
Division of Urology, Department of Surgery, Western University, London, Canada; Department of Microbiology & Immunology, Western University, London, Canada

Hassan Razvi, MD, FRCSC
Division of Urology, Department of Surgery, Western University, London, Canada