Kidney stone disease is a prevalent and costly condition [1]. Large kidney stones are often treated with percutaneous nephrolithotomy (PCNL) [2]. In addition to the planned outcome of stone removal (ie, stone-free status), PCNL has a number of potential complications, including a need for blood transfusion, postoperative infection, and visceral injury [3]. Several scoring systems have been built in attempts to predict outcomes for individual patients [4]. More recently, (supervised) machine learning (ML) techniques have been used to build models for predicting outcomes of PCNL [5], [6], [7]. In comparison to statistical methods, ML can handle highly nonlinear relationships by allowing a computer to predict outcomes on the basis of algorithmic rather than statistical methods (eg, logistic regression). This provides superior accuracy in comparison to traditional statistical methods, especially for rare outcomes. However, the models generated to produce these results can be highly complex and in effect a “black box”. To try and demonstrate which variables contribute to a particular outcome, metrics such as Shapley weighting are used to “explain” individual ML predictions [8].

To date, the largest PCNL data set used for ML involved 134 cases [6] and only four outcomes have been described: stone-free status, need for adjuvant treatment, need for stent insertion, and need for blood transfusion. Aminsharifi et al [6] demonstrated that ML using support vector machines had superior accuracy to traditional nomograms (Guy’s stone score and the Clinical Research Office of the Endourological Society [CROES] PCNL nomogram).

None of the currently published models have been validated. There are three ways to perform validation, from which prognostic accuracy statistics are generated: internal, external, and temporal validation. Internal validation (often termed the “test” set) simply represents a small subset (usually 20–30%) of the total dataset. External validation uses an external data set. Temporal validation is a form of external validation for which new data are collected from the same source as the training set but in a different (ideally later) time period [9].

To facilitate better personalised prediction of postoperative outcomes, we used a large national database to develop ML models for prediction of seven important PCNL outcomes: stone-free status, need for transfusion, need for intensive care, visceral injury, need for adjuvant treatment, postoperative infection, and postoperative complications. After temporal validation, the best-performing models were used in a web-based application to facilitate individualised predictions.