P68. Machine learning algorithms for predicting patient-controlled analgesia use after minimally invasive spine surgery

Abstract

BACKGROUND CONTEXT Machine learning models have accurately predicted outcomes after various types of orthopedic surgery. These models have potential to enhance patient selection, risk stratification and preoperative planning. PURPOSE The aim of this study was to assess the efficacy of machine learning algorithms for prediction of patient-controlled analgesia use after minimally invasive spine surgery. STUDY DESIGN/SETTING Retrospective review. PATIENT SAMPLE Consecutive patients who underwent minimally invasive spine surgery between 2017 and 2021. OUTCOME MEASURES Whether or not patient-controlled analgesia (PCA) was used in the first 24 hours after surgery. METHODS From a single-institution, multi-surgeon, prospectively maintained database, patients undergoing minimally invasive spine surgery were included. Cases were excluded for revision, cancer, nonelective surgery, infection and missing categorical data. Cases were randomly partitioned into training (75%) and testing (25%) data sets. K-nearest neighbor imputation was used for missing continuous variables. A random forest classifier was used to select the top 10 most important features. Eighteen different models were trained to predict whether PCA was used in the immediate post-operative period. When appropriate, class weights were adjusted to address class imbalance and models were calibrated using Platt's (sigmoid) method. Using the testing set and the top 6 classifiers, model performance was evaluated using overall accuracy and area under the receiver operating characteristic curve (AUC). RESULTS A total of 848 cases were included, where 636 (75%) were used for training and 212 (25%) for testing. Overall incidence of PCA use was 13.9%. The 10 most important predictors were BMI, age, number of operative table positions, fusion vs decompression, age-adjusted Charlson comorbidity index, number of operative levels, gender, primary vs revision, pre-operative narcotic use and type of insurance. The 6 best performing models were naive bayes classifier (AUC= 0.89; Accuracy= 86.8%), ridge classifier (AUC= 0.885; Accuracy= 89.2%), logistic regression (AUC= 0.884; Accuracy= 89.2%), linear support vector classifier (AUC= 0.882; Accuracy= 89.2%), perceptron (AUC= 0.873; Accuracy= 88.2%), and random forest classifier (AUC= 0.859; Accuracy= 87.7%). CONCLUSIONS Machine learning algorithms were found to reliably predict postoperative PCA use after minimally invasive spine surgery. These results can help guide important clinical decisions, reduce costs, and improve expectation management. FDA DEVICE/DRUG STATUS This abstract does not discuss or include any applicable devices or drugs. Machine learning models have accurately predicted outcomes after various types of orthopedic surgery. These models have potential to enhance patient selection, risk stratification and preoperative planning. The aim of this study was to assess the efficacy of machine learning algorithms for prediction of patient-controlled analgesia use after minimally invasive spine surgery. Retrospective review. Consecutive patients who underwent minimally invasive spine surgery between 2017 and 2021. Whether or not patient-controlled analgesia (PCA) was used in the first 24 hours after surgery. From a single-institution, multi-surgeon, prospectively maintained database, patients undergoing minimally invasive spine surgery were included. Cases were excluded for revision, cancer, nonelective surgery, infection and missing categorical data. Cases were randomly partitioned into training (75%) and testing (25%) data sets. K-nearest neighbor imputation was used for missing continuous variables. A random forest classifier was used to select the top 10 most important features. Eighteen different models were trained to predict whether PCA was used in the immediate post-operative period. When appropriate, class weights were adjusted to address class imbalance and models were calibrated using Platt's (sigmoid) method. Using the testing set and the top 6 classifiers, model performance was evaluated using overall accuracy and area under the receiver operating characteristic curve (AUC). A total of 848 cases were included, where 636 (75%) were used for training and 212 (25%) for testing. Overall incidence of PCA use was 13.9%. The 10 most important predictors were BMI, age, number of operative table positions, fusion vs decompression, age-adjusted Charlson comorbidity index, number of operative levels, gender, primary vs revision, pre-operative narcotic use and type of insurance. The 6 best performing models were naive bayes classifier (AUC= 0.89; Accuracy= 86.8%), ridge classifier (AUC= 0.885; Accuracy= 89.2%), logistic regression (AUC= 0.884; Accuracy= 89.2%), linear support vector classifier (AUC= 0.882; Accuracy= 89.2%), perceptron (AUC= 0.873; Accuracy= 88.2%), and random forest classifier (AUC= 0.859; Accuracy= 87.7%). Machine learning algorithms were found to reliably predict postoperative PCA use after minimally invasive spine surgery. These results can help guide important clinical decisions, reduce costs, and improve expectation management.