Novel Machine Learning Approaches to Examine Surgical Outcomes in Patients with Craniosynostosis
thesisposted on 01.05.2020, 00:00 by Shayna Azoulay-Avinoam
Objective: Craniosynostosis, defined as the premature fusion of cranial sutures, is one of the most common congenital craniofacial defects occurring 1 in 2000 to 2500 live births. Patients with craniosynostosis undergo multiple surgical interventions the earliest of which is in the first few years of life. Surgical interventions include fronto-orbital advancement, open cranial vault remodeling, extended strip craniectomy, spring-assisted cranial expansion, and cranial vault distraction. The objective of this study is to examine a mix of patient and hospital level factors associated with perioperative outcomes. We hypothesize that a mix of patient factors are associated with outcomes. Methods: The Nationwide Inpatient Sample for the years 2012 to 2014 was used for this study. All patients aged up to 3 years who had a surgical repair of craniosynostosis were selected. Neural Network Models using Multilayer Perception was used. The dataset was partitioned into training (70%) and testing (30%) datasets. Automatic architecture selection was used for specifying the number of Hidden Layers in the model. Outcomes examined included development of infectious complications, length of stay in hospital and hospital charges. Multivariable logistic and linear regression models were used to examine the simultaneous effects of patient and hospital level variables on outcomes. Results: During the study period, across the entire United States a total of 8360 hospitalization had a craniosynostosis correction surgical procedure. 65% were males. The overall infectious complication rate was 3.3%. The mean length of stay in hospital was 4.3 days and the total hospitalization charges was $91,795. In the Neural Network model predicting occurrence of infectious complications, 6 hidden layers were used. In terms of normalized importance, co-morbid burden (100%) and Race (81%) were strongest predictors. The Area Under Curve was 0.83. Overall, 96.2% of patients were classified correctly (infections or no infections). For Length of Stay and Hospital Charges models, comorbid burden (100% normalized importance) was the strongest predictor. 11 hidden layers were used in a sensitivity analysis infection model to identify the best mix of co-morbid conditions associated with high risk of developing infections. Neurological disorders, fluid and electrolyte disorders, and hypothyroidism were the strongest predictors. Conclusions: We developed a novel neural network model to accurately classify those who were highly likely to develop infections, have longer length of stay in hospital, and have high hospitalization charges amongst those undergoing craniosynostosis surgical repairs.