Machine Learning and Artificial Intelligence to Predict Urinary Tract Infections and Continuous Antibiotic Prophylaxis in Prenatal Hydronephrosis
Yanbo Guo, MD1, Armando J. Lorenzo, MD2, Mandy Rickard, MN NP1, Luis H. Braga, MD1.
1McMaster University, Hamilton, ON, Canada, 2The Hospital for Sick Children, Toronto, ON, Canada.
Introduction: Prenatal hydronephrosis (PNH) affects up to 5% of infants. These children can undergo a battery of serial testing, develop urinary tract infections, and require continuous antibiotic prophylaxis (CAP). Current management strategies are based upon subjective radiographic grading systems. Being able to better identify patients who will need further intervention will allow us to better target our investigations and management. The increasing availability of sophisticated machine learning and artificial intelligence platforms provides a novel opportunity to build complex predictive models that are easy to distribute and to use. We explored the application of this emerging technology by creating a predictive model for patients with PNH who will develop UTIs or require CAP.
Methods: A de-identified prospective PNH database from McMaster University was uploaded to the Microsoft® Azure Machine Learning Studio. The dataset was then split into a boosted decision tree learning model and an artificial neural network model. These were then scored and evaluated to determine the optimal model.
Results: 571 entries were included. The optimized model for UTI prediction achieved an area under the curve of 0.925. With a threshold selected to maximize sensitivity, we achieved a sensitivity of 88% with a specificity of 89% and accuracy of 0.889. The optimized model for CAP prediction achieved an area under the curve of 0.948. With a threshold selected to maximize sensitivity, we achieved a sensitivity of 94% with a specificity of 65% and accuracy of 0.825.
Conclusions: We built two accurate predictive models with a commercially available, easily accessible, cloud-based machine learning platform. These models' strong performance characteristics suggest their affinity to be used as screening tests to identify patients who require further evaluation and specialist consults. Furthermore, we provide evidence that these emerging technologies provide an opportunity to surpass the current standard of predictive analytics and represent the next development in personalized medicine.
Back to 2017 Program