Moderate evidence
More
AI predicted kidney stone type correctly 73% of the time, versus 33% for the standard urine test
Study title: A Validated Artificial Intelligence Model Versus Urinary Supersaturation Alone for Predicting Kidney Stone Composition
- Evidence type
- Controlled or validated comparison
- Sample
- n = 1,653
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2883cite or check it
Study title: A Validated Artificial Intelligence Model Versus Urinary Supersaturation Alone for Predicting Kidney Stone Composition
Researchers built an AI model to predict what a kidney stone is made of, using patient demographics plus results from a 24-hour urine supersaturation panel. On held-out data, it identified stone composition correctly 73% of the time, compared with 33% for the EQUIL2 supersaturation algorithm used in routine clinical practice.
Read the study abstract
Objective To develop and validate an artificial intelligence (AI) model that can identify kidney stone composition more accurately than the EQUIL2 urine supersaturation (SS) algorithm. Patients and methods Patients with both a 24-hour urine SS panel and a kidney stone composition analysis conducted within 90 days of each other between September 24, 2015, and March 15, 2024, were identified (n=13,000 unique patients). Kidney stones were classified as either majority calcium oxalate (CaOx), majority calcium phosph…
The gain
In a validation set of 1,653 paired measurements from patients, an AI model identified kidney stone composition with 0.73 accuracy, versus 0.33 for the EQUIL2 urine supersaturation calculation used in routine care.
In brief
The model was developed on about 13,000 unique patients who had both a 24-hour urine panel and a stone composition analysis within 90 days of each other between 2015 and 2024. It was then validated on 1,653 paired measurements collected from March 2024 to March 2025.
Stones were sorted into majority calcium oxalate, majority calcium phosphate, uric acid or other. The authors report the model beat EQUIL2 in every stone category and within various subgroups, and also outperformed previously published AI models using the same urine panel.
Main points
- The study used records from about 13,000 unique patients with a 24-hour urine supersaturation panel and a stone analysis within 90 days of each other.
- Data for model development were collected between September 24, 2015, and March 15, 2024.
- The model was validated on a held-out set of 1,653 paired measurements collected from March 16, 2024, to March 15, 2025.
- Stones were classified as majority calcium oxalate, majority calcium phosphate, uric acid or other.
- Accuracy in validation was 0.73 for the AI model versus 0.33 for supersaturation alone.
- The authors report the model outperformed previously published AI models in predicting calcium oxalate, calcium phosphate and uric acid stones.
The full story
The AI model used demographic characteristics plus urinary supersaturation panel results as inputs. The comparison baseline was supersaturation for calcium oxalate, hydroxyapatite, brushite and uric acid, calculated with EQUIL2 as in routine practice and expressed on the delta Gibbs free energy scale.
Training data came from patients with both a 24-hour urine panel and a stone composition analysis within 90 days, spanning September 2015 to March 2024. Validation used a later, separate period, March 2024 to March 2025, giving 1,653 paired measurements.
The authors conclude the model outperformed EQUIL2-calculated supersaturation in all stone categories and various subgroups, as well as earlier published AI models that use the same 24-hour urine panel.
Sources
1 on file- Mayo Clinic Proceedingsdoi:10.1016/j.mayocp.2026.06.027
Talk it over
The AI model identified kidney stone composition with 0.73 accuracy, versus 0.33 for EQUIL2. If you were a patient with a stone, how would that difference affect which test result you wanted?
AI predicted kidney stone type correctly 73% of the time, versus 33% for the standard urine testhttps://truvace.com/item/a-validated-artificial-intelligence-model-versus-urinary-supersaturation-alone-f
The AI model identified kidney stone composition with 0.73 accuracy, versus 0.33 for EQUIL2. If you were a patient with a stone, how would that difference affect which test result you wanted?
Answer in the threadThe model was trained on records from about 13,000 patients and checked on 1,653 paired measurements from March 2024 to March 2025. How far does a later test period from the same records settle whether it works elsewhere?
Answer in the threadStones were sorted into majority calcium oxalate, majority calcium phosphate, uric acid or other. Which patients or stone types might be poorly served if results come mostly from this group?
Answer in the thread
If you have more time
Look up how stone composition is confirmed today. What would you want to know about the wait or cost before a model like this changed practice?
Answer in the threadThe authors report the model beat earlier published AI models using the same 24-hour urine panel. What further comparison would show it works in a different clinic?
Answer in the thread
ace


What readers say