WCET 2024: Leveraging A Machine Learning Model To Predict Kidney Stone Composition: Implications For Treatment And Pathophysiology

(UroToday.com) To kick off the first day of the conference, Western University's Jennifer Bjazevic presented a significant advancement in the preoperative management of urolithiasis to tackle a critical issue in the treatment of kidney stones: determining the stone's composition before it is passed or surgically removed. This is essential because treatment strategies for urolithiasis are heavily dependent on the stone’s composition (i.e., calcium or non-calcium elements).


Traditionally, clinicians could only ascertain stone composition after the fact, once the stone was physically available for analysis. However, Bjazevic's team proposed a novel solution—leveraging clinical data to predict the stone's composition in advance, such as serum biochemistry, urinalysis, and medical history. By using physiological parameters that likely mirror those present during the stone's formation, the team developed a predictive model that could revolutionize how urologists approach stone treatment.

Methodologically, the study was robust and forward-thinking. The research involved a large, prospectively collected dataset from 777 patients at a tertiary care center's metabolic stone clinic. This cohort included 625 calcium stone formers and 152 non-calcium stone formers. Key data points included 24-hour urine collections, serum biochemistry, and biometric information. The dataset was split, with 80% used to train a binary gradient boosted tree model and the remaining 20% reserved for testing. To ensure the model's accuracy, the team addressed class imbalance by up sampling the minority class and fine-tuned the model's hyperparameters using Bayesian optimization (Figure 1). The model's effectiveness was validated and assessed via kappa score and feature importance was determined using SHAP (SHapley Additive exPlanation) values.

image-0.jpg 

Figure 1. Overview of machine learning workflow.

The results were promising, as the model was able to accurately predict calcium from non-calcium stones. The model's performance, measured by the area under the receiver operator characteristics (AUC-ROC) curve, achieved a respectable 0.76, proving to be the most accurate model published to date (Figure 2).

image-1.jpg 

Sensitivity and specificity stood at 0.86 and 0.73, respectively, indicating the model's solid potential for clinical application. The team investigated factors that played a crucial role in distinguishing between calcium and non-calcium stone compositions (Figure 3).

image-2.jpg 

Key predictors identified were 24-hour urine calcium levels, which indicated calcium-based stones, along with blood urate and blood phosphate levels, both of which were associated with non-calcium-based stones (Figure 4).

image-3.jpg

The implications of this study are significant. By accurately predicting stone composition based on clinical parameters, this model could enable urologists to tailor their treatment strategies before any surgical intervention. This preemptive approach not only has the potential to improve patient outcomes but also offers deeper insights into the pathophysiology of stone disease. Looking forward, Bjazevic and her team suggest that extending machine learning algorithms could further refine the ability to pinpoint specific stone compositions, ultimately leading to more personalized and effective treatments for patients suffering from urolithiasis.

During the question-and-answer session following the presentation, Dr. Scotland highlighted the elevated phosphate levels associated with non-calcium-based stones. In response, Dr. Bjazevic explained that the research team had attempted to further refine the model to distinguish between various stone types and explore the underlying causes of the observed hyperphosphaturia. However, the limited availability of data for non-calcium stone types posed a challenge, as they could not gather sufficient data points to produce reliable results, unlike the extensive data available for calcium oxalate stones. Dr. Jung then asked a follow-up question regarding the criteria used by the team to define calcium stones in their study. Dr. Bjazevic clarified that their investigation focused on stone fragments with a calcium content greater than 50%.

This presentation underscores the growing role of machine learning in urology, offering a glimpse into a future where data-driven tools could become an integral part of clinical decision-making.

Presented by: Jennifer Bjazevic, M.D., Western University

Co-Authors: Gerrit Stuivenberg, John Chmiel, Jennifer Wong, Linda Nott, Jeremy Burton, Hassan Razvi

Moderated by: Kymora Scotland, M.D., and Hae Do Jung, M.D.

Written by: Seyed Amiryaghoub M. Lavasani, B.A., University of California, Irvine, @amirlavasani_ on Twitter during the 2024 World Congress of Endourology and Uro-Technology (WCET) Annual Meeting, August 12 -16, 2024, Seoul, South Korea