Using machine learning for personalized prediction of revision paranasal
sinus surgery
Abstract
Background: Uncontrolled chronic rhinosinusitis (CRS) needing
consideration of surgery is a growing health problem yet its risk
factors at individual level are not known. Our aim was to examine risk
factors of revision endoscopic sinus surgery (ESS) at the individual
level by using artificial intelligence. Methods: Demographic
and visit variables were collected from electronic health records (EHR)
of 790 operated CRS patients. The effect of variables on the prediction
accuracy of revision ESS was examined at the individual level via
machine learning models. Results: Revision ESS was performed to
114 (14.7%) CRS patients. The logistic regression, gradient boosting
and random forest classifiers had similar performance (AUC values .746,
.745 and .747, respectively) for predicting revision ESS. The best
performance was yielded by using logistic regression and long predictor
data retrieval time (AUC .809, precision 36%, sensitivity 70%) as
compared with data collection time from baseline visit until 0, 3 and 6
months after the baseline ESS (AUC values .668, .717 and .746,
respectively). The number of visits, number of days from the baseline
visit to the baseline ESS, age, CRS with nasal polyps (CRSwNP), asthma,
NERD and immunodeficiency or its suspicion were associated with revision
ESS. Age and the number of visits before baseline ESS had non-linear
effects for the predictions. Conclusions: Intelligent data
analysis found important predictors of revision ESS at the individual
level, such as visit frequency, age, Type 2 high diseases and
immunodeficiency or its suspicion.