NLP

NLP predicts likelihood of non‑home discharge after craniotomy for meningioma resection by applying machine learning to unstructured preoperative notes and radiology reports.

Key Features:

• Data source: Preoperative notes and radiology reports from 595 adults undergoing meningioma resection at a single academic center (1995–2015).
• Multi‑institutional extension: A multi‑institutional model incorporates an additional center with 693 patients.
• Algorithm ensemble: Thirty‑two machine learning algorithms were trained and the top three performing models were combined into an ensemble.
• Input type: Analysis of unstructured clinical text using natural language processing methods.
• Performance metrics: Single‑institution model AUC 0.80 (internal) and 0.76 (holdout); multi‑institutional model AUC 0.78 (internal) and 0.76 (holdout).
• Comparison: The NLP approach outperformed a previously published model that relied on 52 neurosurgeon‑selected variables.
• Feature analysis: Permutation importance identified preoperative notes as a significant predictor; word clouds and non‑negative matrix factorization were used to analyze predictive text features.

Scientific Applications:

• Discharge disposition prediction: Predicting postoperative non‑home discharge to inform patient management and resource allocation in neurosurgery.

Methodology:

NLP extraction from unstructured clinical notes and radiology reports; training of 32 algorithms with an ensemble composed of the top three models; permutation importance, word clouds, and non‑negative matrix factorization for feature analysis.