Pre-operative prediction of post-operative urinary retention in lumbar surgery: a prospective validation of machine learning model

• Published in: European spine journal 2023-11, Vol.32 (11), p.3868-3874
• Main Authors: Porche, Ken, Maciel, Carolina B., Lucke-Wold, Brandon, Mehkri, Yusuf, Murtaza, Yasmeen, Goutnik, Michael, Robicsek, Steven A., Busl, Katharina M.
• Format: Article
• Language: English
• Subjects: Aged; Complications; Humans; Learning algorithms; Machine Learning; Male; Medicine; Medicine & Public Health; Middle Aged; Neural networks; Neurosurgery; Original Article; Postoperative; Postoperative Complications - diagnosis; Postoperative Complications - etiology; Prediction models; Prospective Studies; Reproducibility of Results; Retrospective Studies; Surgical Orthopedics; Urinary retention; Urinary Retention - diagnosis; Urinary Retention - etiology
• ISSN: 0940-6719; 1432-0932; 1432-0932
• DOI: 10.1007/s00586-023-07954-4

Purpose Predicting urinary retention is difficult. The aim of this study is to prospectively validate a previously developed model using machine learning techniques. Methods Patients were recruited from pre-operative clinic. Prediction of urinary retention was completed pre-operatively by 4 individuals and compared to ground truth POUR outcomes. Inter-rater reliability was calculated with intercorrelation coefficient (2,1). Results 171 patients were included with age 63 ± 14 years, 58.5% (100/171) male, BMI 30.4 ± 5.9 kg/m 2 , American Society of Anesthesiologists class 2.6 ± 0.5, 1.7 ± 1.0 levels, 56% (96/171) fusions. The observed rate of POUR was 25.7%. The model’s performance was found to be 0.663 (0.567–0.759). With a regression model probability cutoff of 0.24 and a neural network cutoff of 0.23, the following predictive power was achieved: specificity 90.6%, sensitivity 22.7%, negative predictive value 77.2%, positive predictive value 45.5%, and accuracy 73.1%. Intercorrelation coefficient for the regression aspect of the model was found to be 0.889 and intercorrelation coefficient for the neural network aspect of the model was found to be 0.874. Conclusions This prospective study confirms performance of the prediction model for POUR developed with retrospective data, showing great correlation. This supports the use of machine learning techniques in the prediction of postoperative complications such as urinary retention.