The pathogenesis of kidney stone formation is complex. Treatment is predicated on appropriately addressing the environmental and metabolic risk factors that lead to stone formation. Because of this, patient data are usually multidimensional and collected across several modalities (ie, electronic health records [EHRs], laboratory results, imaging systems, and surgical recordings). This complicates clinical interpretation and management strategies that require a multifaceted approach, including medical, dietary, and surgical interventions. Novel machine learning (ML) methods offer a unique solution for evaluating clinical information as they enable flexible model fitting for high-dimensional, complex patient data. These methods could refine interpretation and evaluation of clinical data to improve both medical and surgical kidney stone treatment (Figure).

Risk Prediction and Patient Stratification

One of the most well-developed areas in the application of ML to nephrolithiasis is risk prediction. ML models have been trained to predict stone recurrence, spontaneous passage of ureteral stones, and postoperative complications. For example, algorithms such as random forests and gradient boosting machines have been used to assess stone-free rates following shock wave lithotripsy or ureteroscopy based on stone size, location, density, and patient characteristics. A study by Nguyen et al¹ demonstrated the potential of ML models to identify individuals at risk of reduced quality of life following stone treatment, enabling health care providers to allocate targeted support and prevent stone recurrence. To do this, they utilized EHR data from 3206 kidney stone patients to develop the Wisconsin Stone Quality of Life Questionnaire Machine-Learning Algorithm, which automatically stratified health-related quality of life scores with good accuracy.

Improving Imaging and Diagnosis

ML has shown utility in radiology, where convolutional neural networks have been developed to assist with stone detection and characterization on CT scans.² Automated stone burden quantification, including for determining stone volume, an otherwise time-consuming task, is another area where ML has the high potential to impact patient care. Similarly, EHR-based ML models can potentially improve the identification of patient-specific metabolic abnormalities that contribute to stone formation. For example, Kavoussi et al³ leveraged EHR information to associate EHR features with 24-hour urine parameters. By training and comparing several models, they found that ML models could associate EHR-derived features with 24-hour urine chemistries, with BMI, age, and gender being the highest prioritized features for prediction. This could enable identification of high-risk stone formers as well as earlier dietary or pharmacological treatment. Similarly, Abraham et al⁴ sought to develop an ML model to predict stone composition based on EHR and 24-hour urine features. They demonstrated using ML models to noninvasively predict stone composition, which could lead to earlier, directed therapy. Finally, ML models can improve classification of acute care needs during symptomatic stone events. For example, Bejan et al⁵ found that the large language models improved the classification of symptomatic stone events compared with International Classification of Diseases codes alone. Models have also demonstrated good discrimination in identifying patients admitted with kidney stones into acute care units based on EHR data.⁶ This demonstrates the feasibility of automated tools for triaging patients for disease-specific care.

Optimizing Surgical Decision-Making

ML models could also be applied to improve kidney stone surgery by leveraging vast amounts of clinical and imaging data toward a more patient-specific approach. Selecting the optimal surgical treatment can be challenging and involves weighing the risks and benefits of surgical approach with patient factors. By leveraging computer vision techniques (ie, a subfield of ML that enables interpretation of clinical data), kidney stone imaging can be objectively analyzed and associated with outcomes. For example, ML models have been developed both for improving targeting during shock wave lithotripsy, as well as predicting successful treatment outcomes.⁷ Similarly, ML models have been applied to patients undergoing percutaneous nephrolithotomy to facilitate percutaneous access and assess the risk of complications. This could enable surgeons to more confidentially and appropriately personalize treatment options.

More recently, implementation of ML models during ureteroscopy shows potential in identifying and tracking kidney stones endoscopically, while giving kidney stone–related information. Despite being the most frequently performed surgical procedure for kidney stones, ureteroscopy is limited by a small field of visibility that is easily impacted by bleeding and debris. Deol et al⁸ demonstrated that computer vision models can accurately segment kidney stones during different tasks of ureteroscopy across 136 videos. Additionally, the group compared the model to the performance of surgeons and found similar performance. Not only could these models be used to help with stone localization and navigation during treatment, but they can also be used to train more complex models for stone classification and surgical characterization. For example, Leng et al⁹ developed a similar segmentation model to enable classification of endoscopic stone composition, which could inform laser settings or earlier medical intervention. Similarly, Cabo et al¹⁰ used segmentation models to predict surgical experience during endoscopic kidney stone surgery, which could be used as an educational tool.

Challenges and Considerations

The integration of ML models into kidney stone disease management demonstrates the potential to transform kidney stone treatment. It is only a matter of time before these models will noticeably change care delivery for kidney stone disease. However, many of the models are still early in development. Many published models are based on small cohorts from single-institution data, limiting generalizability. Rigorous evaluation and external validation are still needed prior to widespread clinical use. Data standardization is another challenge, as ML models require large, well-annotated datasets. Finally, regulatory and ethical considerations will need to be addressed, particularly as ML tools begin to influence treatment decisions directly. Despite these challenges, ML models have the potential to revolutionize kidney stone care and bring precision medicine to our patients.