Predicting stone composition: the accurate model needs data you only have afterwards

Design

Retrospective cohort machine learning development with LASSO stability selection, multiple classifiers, 80/20 internal validation and SHAP explainability

Population

442 patients (median age 51, 67% male) with a dominant stone component above 50% on laboratory analysis, 2019 to 2024

Primary outcome

Macro-averaged one-vs-rest AUC for five-class dominant stone composition

Effect

Clinical variables alone macro-AUC up to 0.809; with morphological descriptors up to 0.983

Four hundred and forty-two patients treated endourologically or who passed a stone spontaneously between 2019 and 2024, each with a dominant component above 50% on laboratory analysis, were used to build machine learning models predicting one of five compositions: calcium oxalate monohydrate (41.0%), calcium oxalate dihydrate (24.9%), calcium phosphate (19.7%), uric acid (10.9%) and cystine (3.6%). Predictors covered demographics, comorbidities, stone metrics, procedural details and Daudon morphological descriptors, with an 80/20 split, LASSO stability selection, several classifiers and SHAP for explainability.

Clinical variables alone reached a macro-averaged AUC of up to 0.809 — good, and available before any intervention. Adding morphological descriptors lifted that to 0.983 with gradient-boosted ensembles, and morphology dominated the SHAP rankings. Stable predictors included age, hereditary disease type, stone density, maximal diameter and location.

The authors are unusually clear about the catch, and it is the whole clinical story. Morpho-constitutional descriptors are read from the stone or its surface during or after the procedure, so the 0.983 model cannot inform the decision it appears to be about — which technique to use, whether to attempt dissolution, how to counsel on recurrence. What is usable preoperatively is the 0.809 clinical model, which is decent but not decisive, and mostly restates what stone density on CT already suggests. Treat this as internal validation of an idea. It is a retrospective single cohort with a 20% held-out split and no external cohort, and cystine, at 16 patients, is too rare here for the model to have learned it reliably.

• Only the clinical-variable model, macro-AUC about 0.809, is available before treatment.
• The 0.983 model uses descriptors read during or after the procedure — it cannot guide the operative plan.
• Stone density on CT remains the single most useful preoperative pointer, and it is inside these models.
• Sixteen cystine stones is too few for any model to have learned that class properly.
• External prospective validation has not been done; do not deploy this in practice yet.

The statistics, in plain English

Macro-averaged AUC treats all five classes as equally important regardless of how common they are, which is generous to a model that has seen only 16 cystine stones — a single misclassification there moves the average far more than one among the 181 calcium oxalate monohydrate cases. The 80/20 split is internal validation: the held-out fifth comes from the same centre, era and pathway as the training data, and models routinely lose several points of AUC when tested elsewhere. That matters most for the 0.983 figure, which is high enough to raise the ordinary suspicion of overfitting, particularly when the added features are derived from the same laboratory process that defines the outcome.

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.