NLP Predicts Glasgow Coma Scale From EHR Notes

Fernandes, Turley, Sun, Mukerji, Moura, Westover, and Zafar published "Automated Prediction of Glasgow Coma Scale Scores From Unstructured Electronic Health Records Using Natural Language Processing: Development and Validation Study" in the Journal of Medical Internet Research (JMIR), with an accompanying dataset release (version 1.0.0, April 17, 2026) on the Brain Data Science Platform (BDSP), a repository managed by Stanford Medicine's Clinical Data Animation Center with Harvard Medical School, Massachusetts General Hospital, and Beth Israel Deaconess Medical Center. The study used daily clinical notes from 145,897 patients across Mass General Brigham hospitals (2017-2024) and the public MIMIC-III critical-care database (2001-2012), totaling 1,446,965 hospitalization-days, split 70/30 into training and hold-out test sets.

The team trained a pooled ordinal regression model (ordinalNet, elastic-net penalty) to classify each patient-day into severe (GCS 3-8), moderate (GCS 9-12), or mild (GCS 13-15) consciousness bands, plus a separate LASSO linear model to predict the continuous GCS score (3-15). The pooled ordinal model reached AUROC 0.96 [95% CI 0.96-0.96] and AUPRC 0.77; a single-institution model trained only on Mass General Brigham data and tested on external MIMIC data generalized reasonably well (AUROC 0.90, AUPRC 0.80). The linear model achieved RMSE 2.30 and Pearson correlation 0.76. Feature analysis showed predictions for severe GCS were driven by terms indicating unresponsiveness and critical interventions, while mild GCS predictions tracked mentions of normal or awake behavior, an interpretable signal that maps onto known clinical language patterns.

The released GitHub repository (bdsp-core/nax-gcs) documents the full pipeline: note cleaning and tokenization, institution-specific data pullers for Mass General Brigham and MIMIC, and separate scripts for pooled versus single-institution ordinal and linear models, in both Python (scikit-learn/statsmodels) and R (ordinalNet). The BDSP data package (about 1.94 GB across 86 files, credentialed access required under a signed Data Use Agreement) includes de-identified sample notes, feature matrices, trained model objects, and predicted probabilities, so credentialed researchers can reproduce results exactly since the authors report fixed random seeds throughout. Teams evaluating similar note-to-score extraction should note the external validation gap: cross-institution AUROC (0.90) is meaningfully lower than pooled in-sample AUROC (0.96), a reminder that clinical NLP models tuned on one health system's documentation style need local recalibration before deployment elsewhere.

Because access to the primary data and trained model weights requires BDSP credentialing, independent replication will be limited to researchers who complete that process; watch for external validation studies at other health systems, and for downstream work that plugs this GCS-extraction pipeline into larger critical-care phenotyping or risk-prediction efforts, which is the stated motivation of the NIH-funded project (grant R01NS131347).