Application of machine learning models to identify predictors of good outcome after laparoscopic fundoplication

BACKGROUND: Laparoscopic fundoplication remains the gold standard treatment for gastroesophageal reflux disease. However, 10% to 20% of patients experience new, persistent, or recurrent symptoms warranting further treatment. Potential predictors for the best outcome after laparoscopic fundoplication were tested using a mature prospectively maintained database. METHODS: Data from 894 consecutive patients who underwent primary laparoscopic fundoplication from 1998 to 2015 were examined using regression and machine learning (ML) models. Preoperative factors were assessed for influence on postoperative outcomes: heartburn, dysphagia, and satisfaction scores at a median follow-up of 5 years. RESULTS: The accuracy in predicting heartburn score (range, 0-10) assessed using the root mean squared error (RMSE) was similar to a negative binomial regression model (RMSE = 2.39) and the least absolute shrinkage support operator ML model (RMSE = 2.34). The multivariate analysis using only patients with complete data (n = 221) generated a lower error than using mean imputation for patients with missing values. The most predictive variables were male sex for heartburn (β = -1.48 95% CI, -2.37 to -0.6; P =.001) and dysphagia (β = -4.70 [95% CI, -8.02 to -1.39; P =.006) and percentage of esophageal peristalsis for satisfaction (β = 0.63 [95% CI, 0.16-1.10; P =.009) and dysphagia (β = -1.85 95% CI, -3.43 to -0.27; P =.02). CONCLUSION: Although male sex and degree of intact peristalsis are significant predictors for outcomes after laparoscopic fundoplication, prediction of individual patient outcome was relatively poor, and ML prediction models provided only marginal improvement in accuracy. Clinical acumen and a discussion with patients to set realistic postoperative expectations cannot be replaced by regression models or standard ML prediction algorithms at the present time.