The model uses a retrospective database of consecutive localized RCC patients referred to a single tertiary care center since 2000. Exclusion criteria included < 5 years follow up, bilateral disease, positive lymph nodes, pathology other than clear cell RCC. The database was then anonymized and uploaded to the Microsoft Azure Machine Learning Studio. Variables were categorized and missing data was processed using probabilistic principal component analysis to replace the missing values. The data was divided into a 70/30 split and a multivariate logistical regression model was trained. A two-class neural network model was created using the dataset. The two models were evaluated using area under the curve (AUC) analyses.
They found that the neural network had an AUC of 0.890 and multivariate logistical regression model an AUC of 0.688. The neural network was found to have a sensitivity of 90%, specificity 72%, positive likelihood ratio 3.24, negative likelihood ratio 0.14. The multivariate logistical regression was found to have a sensitivity 79%, specificity 19%, positive likelihood ratio 1.11 and negative likelihood ratio 0.96.
Interestingly, the authors found a recurrence rate of 15% which is quite high compared to existing literature. Thus, some of the discussion questions revolved around the oncologic outcomes in this cohort.
In conclusion, the authors created an accurate model to predict recurrence after RCC resection. They demonstrated feasibility using cloud-based machine learning and found that the neural network model is superior to traditional logistical regression.
Presented by: Yanbo Guo, MD, Division of Urology, McMaster University, Hamilton, Ontario, Canada
Written by: Selma Masic, MD, Urologic Oncology Fellow (SUO), Fox Chase Cancer Center, @selmasic at AUA 2019. at the American Urological Association's 2019 Annual Meeting (AUA 2019), May 3 – 6, 2019 in Chicago, Illinois