A stone composition model that works best once the stone is already out

Design

retrospective cohort, multiple supervised classifiers with LASSO stability selection and SHAP explainability

Population

442 patients (median age 51, 67% male) with stones analysed as over 50% one component, 2019 to 2024

Primary outcome

dominant stone composition across five categories

Effect

clinical variables alone macro-AUC up to 0.809; adding morphological descriptors up to 0.983 with CatBoost

Machine learning models were built to predict dominant urinary stone composition — calcium oxalate monohydrate, calcium oxalate dihydrate, calcium phosphate, uric acid or cystine — from 442 patients treated or passing stones between 2019 and 2024. Predictors were demographics, comorbidities, stone metrics, procedural details and Daudon morphological descriptors, with an 80/20 split and LASSO stability selection.

Clinical variables alone gave a macro-averaged AUC up to 0.809 — reasonable. Adding the morphological descriptors pushed it to 0.983 with CatBoost. Stable predictors included age, hereditary disease type, stone density, maximal diameter and location.

The authors are unusually honest about what that means, and the honesty is the finding. The morphological features that produce the near-perfect number are intra- and postoperative descriptors — you have to see the stone to have them. So the impressive model is a decision-support tool for after the procedure, not a way to plan it. The genuinely preoperative model is the 0.809 one, and it still needs external prospective validation. In practice the clinical-only variables it relies on are already available at the CT: density, size and location, plus age and known hereditary disease.

• Use stone density, size, location and hereditary history at the CT to anticipate composition
• Do not read the 0.98 figure as preoperative performance — it needs the stone in hand
• Continue sending stones for formal analysis; the model does not replace it
• Remember uric acid stones are the ones where the preoperative prediction changes management most
• No external validation yet — do not deploy a published model on your own patients

Why it matters

It marks the line between a model that could change a preoperative plan and one that only describes a stone you already have.

Don't overread it

Retrospective, single-centre, internally validated only — the best-performing model uses descriptors unavailable before surgery.

The statistics, in plain English

Macro-averaged AUC treats every stone type equally, so the cystine class — 16 patients — counts as much as calcium oxalate monohydrate's 181. That makes the headline figure fragile: a handful of correctly classified rare stones moves it substantially. The 20 per cent validation split also comes from the same centre and period as the training data, which is internal validation, not external.

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.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.