From laser-on time to lithotripsy duration: improving the prediction of lithotripsy duration with 'Kidney Stone Calculator' using artificial intelligence.

introduction"Kidney Stone Calculator" (KSC) helps to plan flexible ureteroscopy, providing the stone volume (SV) and an estimated duration of laser lithotripsy (eLD). eLD is calculated from in vitro ablation rates and SV. KSC's accuracy has been demonstrated with a mean difference between eLD and effective LD (EfLD) of 18.8%. We aimed to reduce the eLD-efLD difference using Machine Learning (ML).

methodsFrom the prospective multicenter KSC database, demographic and peri-operative data were anonymously extracted: SV, stone location, maximum density, anatomy, surgical expertise, ureteral access sheath, basket use, laser source, fiber diameter and settings, eLD and efLD. After normalization and splitting (training (80%), test (20%)), significant variables influencing the difference between eLD and efLD were selected through multiple linear regression (MLR). Six types of ML models were subsequently evaluated to minimize the mean absolute error (MAE) between eLD and efLD on the test group.

results125 patients were included. After normalization and MLR, MAE were significantly influenced by 14 variables (including diverticulum location, surgical expertise, laser sources, laser fiber diameters, 5 Hz frequency, 1.5 J pulse energy and eLD). eLD had the greatest positive impact on the eLD-efLD difference (2.45 (2.16-2.73), p < 0.0001)). Above the various tested ML models, the Bayesian "Automatic-Relevance-Determination" and Dense Neural Networks respectively presented the lowest and highest MAE on the test group (3.9% and 6.8%). Results did not differ among models overall (p = 0.93) and two by two.

conclusionThe difference between KSC's eLD and efLD can be five-fold reduced using ML and Artificial Intelligence, including clinically impactful factors such as the surgical technique or expertise. A clinical inference could help to externally validate our findings.