Comparing supervised machine learning algorithms for the prediction of partial arterial pressure of oxygen during craniotomy

• Published in: BMC medical informatics and decision making Vol. 25; no. 1; pp. 326 - 14
• Main Authors: Gutmann, Andrea S., Mandl, Maximilian M., Rieder, Clemens, Hoechter, Dominik J., Dietz, Konstantin, Geisler, Benjamin P., Boulesteix, Anne-Laure, Tomasi, Roland, Hinske, Ludwig C.
• Format: Journal Article
• Language: English
• Published: London BioMed Central 03.09.2025; BioMed Central Ltd; Springer Nature B.V; BMC
• Subjects: Adult; Aged; Algorithms; Analysis; Arterial partial pressure of oxygen; Arterial Pressure - physiology; Blood gas analysis; Blood gases; Blood pressure; Brain research; Business metrics; Clinical medicine; Craniotomy; Creatinine; Feature selection; Female; Gas analysis; General anesthesia; Health Informatics; Humans; Hypoxemia; Information Systems and Communication Service; Learning algorithms; Machine learning; Male; Management of Computing and Information Systems; Measurement; Medicine; Medicine & Public Health; Methods; Middle Aged; Monitoring, Intraoperative - methods; Neural networks; Neurosurgery; Oximetry; Oxygen; Oxygen - blood; Oxygen saturation; Oxygenation; paO2; Partial Pressure; Patient monitoring; Patients; Predictions; Root-mean-square errors; Supervised learning; Supervised Machine Learning; Tissues; Variables; Germany; Blood gas analysis; Craniotomy; Machine learning; Arterial partial pressure of oxygen; paO; paO2
• ISSN: 1472-6947; 1472-6947
• DOI: 10.1186/s12911-025-03148-8

Background and Objectives Brain tissue oxygenation is usually inferred from arterial partial pressure of oxygen (paO 2 ), which is in turn often inferred from pulse oximetry measurements or other non-invasive proxies. Our aim was to evaluate the feasibility of continuous paO 2 prediction in an intraoperative setting among neurosurgical patients undergoing craniotomies with modern machine learning methods. Methods Data from routine clinical care of lung-healthy neurosurgical patients were extracted from databases of the respective clinical systems and normalized. We used recursive feature elimination to identify relevant features for the prediction of paO 2 . Six machine learning regression algorithms (gradient boosting, k-nearest neighbors, random forest, support vector, neural network, linear model with stochastic gradient descent) and a multivariable linear regression were then tuned and fitted to the selected features. A performance matrix consisting of standard deviation of absolute errors ( σ ae ), mean absolute percentage error (MAPE), adjusted R 2 , root mean squared error (RMSE), mean absolute error (MAE) and Spearman’s ρ was finally computed based on the test set, and used to compare and rank each algorithm. Results We analyzed N  = 4,581 patients with n  = 17,821 observations. Between 5 and 22 features were selected from the analysis of the training dataset comprising 3,436 patients with 13,257 observations. The best algorithm, a regularized linear model with stochastic gradient descent, could predict paO 2 values with σ ae  = 86.4 mmHg, MAPE = 16 %, adjusted R 2  = 0.77, RMSE = 44 mmHg and Spearman’s ρ = 0.83. Further improvement was possible by calibrating the algorithm with the first measured paO 2 /FiO 2 (p/F) ratio during surgery. Conclusion PaO 2 can be predicted by perioperative routine data in neurosurgical patients even before blood gas analysis. The prediction improves further when including the first measured p/F ratio, realizing quasi-continuous paO 2 monitoring.