Using Narrow AI to
Predict Complications in Urologic Oncology Patients
AI Model to Predict Complications During And After Surgery for Patients Undergoing Robotic Partial Nephrectomy
The Challenge
Accurate machine learning (ML) models, when deployed in clinical practice, could play a significant role in preempting adverse events with timely interventions and improving patient outcomes.
We collaborated with the Vattikuti Foundation to develop models to predict Intraoperative Complications (IOC) and 30-day Morbidity (M). The models were constructed using a multi-institutional dataset focused on a specific cohort of kidney cancer patients who underwent Robotic Partial Nephrectomy.
Approach
Demographics, Pre-operative, and Intra- operative data was extracted and pre-processed from 18 centers around the world to develop the models to predict Intraoperative complications and 30-day postoperative morbidity.
Three different ML models were developed for each outcome using Logistic Regression, Random Forest classifier, and Neural Network to determine the best model for each outcome.
All data analysis and model development work was performed on an on-premises server, however, this work can also be modified to train models using AutoML on cloud platforms.
For model deployment and hosting Cloud Functions (autoscaling serverless compute) were utilized from Google Cloud Platform with the ability to migrate to AWS or Azure.
An end to end framework for clinical adoption and utility was defined as a next step to monitor and improve the performance of the models.
Results
%
AUC-ROC
%
PR-AUC
%
Balanced Accuracy
%
AUC-ROC
%
PR-AUC
%
Balanced Accuracy
Conclusion
The surgeon can leverage this window of opportunity to manage the risk by developing and applying situation-specific intervention protocols. The methodology used to construct and train ML models can be applied to develop models that can predict outcomes for patients undergoing surgical procedures in other specialties.