To develop and evaluate an AI model for predicting kidney stone composition when direct stone analysis is unavailable.
Approach:
Development: Researchers used data from 13,000 patients who had stone analysis and 24-hour urine testing within 90 days between September 2015 and March 2024.
Model: The model combined age, sex, and urinary measurements to predict four categories: predominantly calcium oxalate, predominantly calcium phosphate, uric acid, and other compositions.
Evaluation: Performance was assessed in an independent dataset of 1,653 patients and compared with urinary supersaturation calculations alone.
Key Findings:
The AI model correctly classified 73% of stones, compared with 33% using urinary supersaturation calculations alone; the difference was statistically significant.
The model performed better across different age and sex groups.
Interpretation:
The researchers report that AI may extract additional diagnostic information from routine laboratory measurements when a stone is unavailable for direct analysis. They state that predictions should supplement, not replace, direct stone analysis and established clinical assessments.
Limitations:
Further evaluation is needed to determine whether clinical implementation improves treatment decisions or reduces recurrence.
Conclusion:
The study suggests that existing patient and urine-test data can support more accurate estimates of kidney stone composition without additional specimen collection.