Machine Learning Prediction of Spinal Anaesthesia Block Duration

Background. Spinal anaesthesia produces block of unpredictable duration. Cerebrospinal fluid volume in the lumbosacral region is the dominant determinant of block spread and regression, and correlates inversely with abdominal circumference and directly with vertebral column length. Existing single-variable rules perform poorly, and no validated multivariable model is in routine clinical use. Objectives. The primary objective is to develop and internally validate models predicting the duration of sensory block and the duration of motor block after single-shot spinal anaesthesia. The secondary objective is to compare the predictive performance of machine learning approaches with conventional multivariable linear regression. Design. Single-centre prospective observational cohort study. No study-directed intervention is applied; anaesthetic technique, drug, and dose are determined by the attending anaesthetist according to routine practice and are recorded as candidate predictors. Assessments. Sensory block level is assessed bilaterally by pinprick with a 25-gauge needle along the mid-clavicular line, over dermatomes T4 to S1. Motor block is assessed using the modified Bromage scale (0-3). Assessments are performed at fixed intervals from intrathecal injection until complete regression of sensory and motor block. Candidate predictors. Age, sex, height, weight, body mass index, abdominal circumference, vertebral column length, ASA physical status, local anaesthetic dose and baricity, and adjuvant use. Sample size. A target of 255 participants was derived using the Riley criteria for continuous-outcome prediction models (pmsampsize), assuming 9 candidate predictors, an outcome standard deviation of 40 minutes, anticipated R-squared of 0.35, shrinkage of at least 0.90, and a multiplicative margin of error of 1.10. The four criteria yielded 183, 183, 234, and 246; the largest requirement of 246 was inflated by approximately 4 per cent for attrition. Analysis. Model development will use nested repeated 10-fold cross-validation with bootstrap optimism correction. Performance will be reported as R-squared, root mean squared error, and mean absolute error, with calibration plots. SHAP values will be used to describe predictor contributions. Reporting will follow TRIPOD+AI.