Use of Internally Validated Machine and Deep Learning Models to Predict Outcomes of Percutaneous Nephrolithotomy using data from the BAUS PCNL audit

Abstract

Background

Machine (ML) and Deep learning (DL) are subsets of artificial intelligence that use data to build algorithms. These can be used to predict specific outcomes. To date there have been a few small studies on post-PCNL outcomes.

Objective

We aimed to build and internally validate ML/DL models for post-PCNL transfusion and infection using a comprehensive national database.

Design

Machine Learning study using prospective national database. Eight machine learning models for 11 outcomes using 43 predictors. Models were ‘complete-case’ analyses.

Setting

National database

Participants

Patients undergoing PCNL in the UK between 2014-2019.

Outcome Measurements

Diagnostic accuracy statistics including overall accuracy, area-under-the-curve (AUC), sensitivity and specificity.

Results and Limitations

4412 patients were included, with 3088 in the training set and 1324 in the test set. The models predicted need for transfusion and post-operative infection with a very high degree of accuracy (99%) and high AUC (0.99-1.00). Unfortunately, the remainder of the outcomes did not achieve the same high levels. These two outcomes were therefore included in the provisional web-based application: https://endourology.shinyapps.io/PCNL_Prediction_tool/

Conclusions

This is the largest machine learning study on post-PCNL outcomes to date. These models can predict the need for post-PCNL transfusion and post-PCNL infection at an individual level with excellent accuracy. Further work will be done on model tuning and external validation.

Patient Summary

We used a national database of people having a major kidney stone operation (PCNL). Using this data, we built and tested 8 machine learning models for 11 different outcomes from the operation. Using this method, we can give individual predictions for the likely need for a blood transfusion and development of an infection. We have developed an app to allow surgeons to calculate an individual patient’s risk prior to surgery.