- Design
- Retrospective cohort with machine-learning model development and internal validation
- Population
- 442 patients with analysed urinary stones, France, 2019–2024
- Primary outcome
- Dominant stone composition (five classes)
- Effect
- Clinical-only macro-AUC up to 0.809; with morphology up to 0.983 (CatBoost)
Stone composition guides prevention, but formal analysis is not always done. This French retrospective study of 442 patients (2019 to 2024) built machine-learning models to predict the dominant component — calcium oxalate monohydrate (41%), dihydrate (25%), calcium phosphate (20%), uric acid (11%) or cystine (4%).
Models using clinical data alone — age, hereditary disease, stone density on CT, size and location — reached a macro-AUC of up to 0.81. Adding standardised Daudon morphological descriptors, recorded during or after retrieval, raised it to 0.98.
The practical message is less about the algorithm than about morphology. Examining the stone under a stereomicroscope carries most of the information, and the CT density already on the scan gives a useful pre-operative hint — a low-density stone suggests uric acid, which dissolves with urinary alkalinisation. India's stone belt produces large volumes of stones that are never analysed.
- Send every retrieved stone for analysis, ideally including morphology as well as composition.
- Note Hounsfield density on the pre-operative CT; a low-density radiolucent stone suggests uric acid.
- Consider urinary alkalinisation for suspected uric acid stones before intervening, unless obstruction or infection needs drainage first.
- Tailor prevention (fluids, citrate, diet) to the composition rather than giving generic advice.
- Screen young patients and those with recurrent or bilateral stones for cystinuria and other hereditary causes.
Why it matters
Composition determines prevention, yet most stones in routine practice are discarded unanalysed.
Don't overread it
A single-centre retrospective model without external validation; the 0.98 model uses information only available after retrieval.
The statistics, in plain English
Macro-AUC averages how well the model separates each stone type from the rest. The jump from 0.81 to 0.98 shows that stone appearance carries most of the predictive information, not the machine-learning method.
Read the rest in the app
You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for urology, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free