Presentation Details
| Machine Learning Based Prediction of Hospitalist Consult Appropriateness in Total Joint Arthroplasty Jeffrey Huang1, Laura Kosseim2, Emmanuel Gibon3, Michele Fang4*, Daniel Montes4*. 1University of Pennsylvania Perelman School of Medicine, Philadelphia, PA, USA.2University of Pennsylvania, Department Of Internal Medicine, Philadelphia, PA, USA.3University of Pennsylvania, Department of Orthopedic Surgery, Philadelphia, PA, USA.4University of Pennsylvania, Department of Medicine, Division of Hospital Medicine* Co-Corresponding Author.These authors contributed equally to this work., Philadelphia, PA, USA |
Abstract
BACKGROUND: Hospitalist co-management benefits selected higher-risk surgical patients, but criteria for identifying those who benefit most are unclear. We developed a machine-learning model to identify elective hip and knee arthroplasty patients most likely to benefit from co-management.
PURPOSE: We randomly selected 210 unique elective hip and knee arthroplasty cases from 2/3/2025–8/3/2025 for chart reviewed. Collected variables included demographics, preoperative clinical data, comorbidities, medications, functional status, surgical risk scores, postoperative complications, AM-PAC mobility scores, and 30-day readmission or mortality. Cases were labeled appropriate for hospitalist consult if a post-operative complication occurred requiring unique hospitalist recommendations, or if 30-day readmission or mortality occurred. Data were split into an 80% training set and 20% test set. Model parameter and hyperparameter selection was performed within the training set using 5-fold cross validation. Model performance was then evaluated on the independent 20% test set. Models included a radial basis function support vector machine (RBF-SVM) as the primary classifier and a Bagging classifier with logistic regression base learners as an interpretable comparison model.
RESULTS: The RFB SVM model achieved a sensitivity of 82%, specificity of 90%, and an area under ROC curve of 0.92 (figure 1A), suggesting excellent discriminatory capacity. The Bagging classifier model achieved a sensitivity of 73%, specificity of 94%, and AUC of 0.92 (figure 1B), and allowed insight into feature importance rankings for model decision making. Medically optimized status, beta-blocker use, history of VTE, AM-PAC scores, PPI use, and comorbidity burden (CAD, obesity, hyperlipidemia, OSA) were among the features that most strongly predicted consult appropriateness.
CONCLUSIONS: RBF-SVM showed similar sensitivity to our manual approach (82% vs. 85%) but much higher specificity (90% vs. 56%), indicating potential to reduce consult volume. An interpretable model identified key preoperative predictors. Future steps include prospective validation and clinical integration.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
PURPOSE: We randomly selected 210 unique elective hip and knee arthroplasty cases from 2/3/2025–8/3/2025 for chart reviewed. Collected variables included demographics, preoperative clinical data, comorbidities, medications, functional status, surgical risk scores, postoperative complications, AM-PAC mobility scores, and 30-day readmission or mortality. Cases were labeled appropriate for hospitalist consult if a post-operative complication occurred requiring unique hospitalist recommendations, or if 30-day readmission or mortality occurred. Data were split into an 80% training set and 20% test set. Model parameter and hyperparameter selection was performed within the training set using 5-fold cross validation. Model performance was then evaluated on the independent 20% test set. Models included a radial basis function support vector machine (RBF-SVM) as the primary classifier and a Bagging classifier with logistic regression base learners as an interpretable comparison model.
RESULTS: The RFB SVM model achieved a sensitivity of 82%, specificity of 90%, and an area under ROC curve of 0.92 (figure 1A), suggesting excellent discriminatory capacity. The Bagging classifier model achieved a sensitivity of 73%, specificity of 94%, and AUC of 0.92 (figure 1B), and allowed insight into feature importance rankings for model decision making. Medically optimized status, beta-blocker use, history of VTE, AM-PAC scores, PPI use, and comorbidity burden (CAD, obesity, hyperlipidemia, OSA) were among the features that most strongly predicted consult appropriateness.
CONCLUSIONS: RBF-SVM showed similar sensitivity to our manual approach (82% vs. 85%) but much higher specificity (90% vs. 56%), indicating potential to reduce consult volume. An interpretable model identified key preoperative predictors. Future steps include prospective validation and clinical integration.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
Content Locked. Log into a registered attendee account to access this presentation.