Predicting intraoperative complications and 30-days morbidity using machine learning techniques for patients undergoing robotic partial nephrectomy (RPN)
Bhandari M, Nallabasannagari AR, Reddiboina M, Porter J, Jeong W, Mottrie A, Dasgupta P, Challacombe B, Abaza R, Rha KH, Parekh D, Ahlawat R, Capitanio U, Yuvaraja T, Rawal S, Moon D, Buffi's N, Sivaraman A, Maes K, Porpiglia F, Gautam G, Turkeri L, Preethi P, Menon M, and Rogers C. Predicting intraoperative complications and 30-days morbidity using machine learning techniques for patients undergoing robotic partial nephrectomy (RPN). European Urology Open Science 2020; 19:e1969-e1970.
European Urology Open Science
Introduction & Objectives: Personalization of a patient risk profile using instruments such as RENAL, PADUA, and MAP scores have limited clinical value. High morbidity of the RPN is attributed to the combination of tumor complexity, patient-related comorbidities, tumor surroundings, and surgeon experience. Predictive models built using Machine Learning (ML) could play a significant role in preempting events, timely intervention and improving patient outcomes. Our objective was to predict Intraoperative Complications (IOC) and 30-day Morbidity (M) as a prelude to the prospective deployment of models in a clinical setting at VCQI collaborating institutions to evolve personalized management strategies.
Materials & Methods: Predictive models were developed using Logistic Regression, Random Forest, and Neural Networks. Models to predict IOC were trained using patient demographics and preoperative data. In addition to the above data, intraoperative data was used to build models to predict M. Model performance on the test dataset was assessed using Area Under Receiver Operating Curve (AUC-ROC), and Area Under Precision-Recall Curve (PR-AUC). We used bootstrapping to generate confidence intervals for the scores and performed permutation test to assess if the observed difference in AUC-ROC and PR-AUC was significant.
Results: Models for predicting IOC were constructed using data from 1690 patients and 38 variables; the best model had AUC-ROC of 0.825 (95% CI 0.717,0.919), and PR-AUC of 0.585 (95% CI 0.394,0.756). Models for predicting M were trained using data from 1455 patients and 59 variables; the best model had AUC-ROC of 0.868 (95% CI 0.827,0.906), and PR-AUC 0.697 (95% CI 0.603,0.781). Our dataset with pre-defined variables does not account for the temporal shift in patient characteristics, a key limitation of this study.
Conclusions: ML model performance during this study is encouraging. These models can be used to predict complications during, and after surgery with good accuracy; paving the way for application in clinical practice to predict, intervene at an opportune time, avert complications and improve patient outcomes. Further validation in a clinical setting would be necessary to establish their clinical value. We propose deploying the models at participating centers contributing to the database (Figure 1).