Related Experiment Video
Updated: Nov 20, 2025
03:14
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
845
Predicting Discharge Disposition Following Meningioma Resection Using a Multi-Institutional Natural Language
Whitney E Muhlestein1, Meredith A Monsour2, Gabriel N Friedman3
1Department of Neurosurgery, University of Michigan Medical Center, Ann Arbor, Michigan.
Neurosurgery
|January 23, 2021
Summary
Natural language processing (NLP) models can predict non-home discharge after meningioma resection, outperforming traditional methods. This approach effectively utilizes unstructured clinical notes for improved patient outcome prediction in neurosurgery.
Area of Science:
- Neurosurgery
- Artificial Intelligence
- Clinical Informatics
Background:
- Machine learning (ML) models in neurosurgery often require extensive discrete data.
- Natural language processing (NLP) offers a method to extract valuable information from unstructured clinical text.
- There is an underutilization of ML and NLP in neurosurgical applications.
Purpose of the Study:
- To develop and present an NLP model for predicting non-home discharge after meningioma resection.
- To implement this NLP model at the point-of-care for practical clinical use.
Main Methods:
- Retrospective collection of preoperative notes and radiology reports from 595 patients undergoing meningioma resection.
- Training and ensembling of 32 ML algorithms, with the top 3 forming the final model.
- Comparison of the NLP model's predictive performance (AUC, calibration) against a 52-variable neurosurgeon-selected model.
- Development of a multi-institutional model by incorporating data from an additional 693 patients.
- Analysis of input importance using permutation importance and text analysis techniques (word clouds, non-negative matrix factorization).
Main Results:
- The single-institution NLP model achieved an AUC of 0.80 (internal) and 0.76 (holdout) for predicting non-home discharge.
- This performance surpassed the AUC of 0.77 (internal) and 0.74 (holdout) for the 52-variable ensemble model.
- The multi-institutional NLP model demonstrated comparable performance with AUCs of 0.78 (internal) and 0.76 (holdout).
- Preoperative notes were identified as the most influential input feature for the model's predictions.
- The developed NLP model is accessible at http://nlp-home.insds.org.
Conclusions:
- Machine learning and NLP are powerful, yet underutilized, tools in neurosurgery.
- A multi-institutional NLP model effectively predicts non-home discharge, demonstrating the potential of leveraging unstructured clinical data.
- The study highlights the feasibility of point-of-care implementation for such predictive models.
