Objective: This article presents the Harvard Electroencephalography Database (HEEDB), a large‐scale, deidentified, and standardized electroencephalographic (EEG) resource supporting artificial intelligence‐driven and reproducible research in epilepsy and broader clinical neuroscience.Methods: HEEDB aggregates more than 280 000 EEG recordings from more than 108 000 patients across four Harvard‐affiliated hospitals. Data are harmonized using the Brain Imaging Data Structure and hosted on the Brain Data Science Platform. EEG data are linked with clinical notes, International Classification of Diseases, 10th Revision codes, medications, and EEG reports. Deidentification follows Health Insurance Portability and Accountability Act Safe Harbor standards.Results: The database includes routine, epilepsy monitoring unit, and intensive care unit EEGs across all age groups, with 73% linked to deidentified clinical reports and 96% of those matched to recordings. Findings are extracted using expert curation, regular expressions, and medical natural language processing models. Auxiliary data include diagnoses, medications, and hospital course, supporting multimodal analysis.Significance: HEEDB fills a critical gap in EEG data availability for epilepsy research. By enabling large‐scale, privacy‐compliant, and clinically relevant analysis, it accelerates the development of diagnostic tools, improves training datasets for machine learning, and promotes data‐sharing in alignment with FAIR (Findable, Accessible, Interoperable, Reusable) and National Institutes of Health data policies.

Epileptiform discharges, or spikes, within electroencephalogram (EEG) recordings are essential for diagnosing epilepsy and localizing seizure origins. Artificial intelligence (AI) offers a promising approach to automating detection, but current models are often hindered by artifact-related false positives and often target either event- or EEG-level classification, thus limiting clinical utility. We developed SpikeNet2, a deep-learning model based on a residual network architecture, and enhanced it with hard-negative mining to reduce false positives. Our study analyzed 17,812 EEG recordings from 13,523 patients across multiple institutions, including Massachusetts General Brigham (MGB) hospitals. Data from the Human Epilepsy Project (HEP) and SCORE-AI (SAI) were also included. A total of 32,433 event-level samples, labeled by experts, were used for training and evaluation. Performance was assessed using the area under the receiver operating characteristic curve (AUROC), the area under the precision–recall curve (AUPRC), calibration error, and a modified area under the curve (mAUC) metric. The model’s generalizability was evaluated using external datasets. SpikeNet2 demonstrated strong performance in event-level spike detection, achieving an AUROC of 0.973 and an AUPRC of 0.995, with 44% of experts surpassing the model on the MGB dataset. In external validation, the model achieved an AUROC of 0.942 and an AUPRC of 0.948 on the HEP dataset. For EEG-level classification, SpikeNet2 recorded an AUROC of 0.958 and an AUPRC of 0.959 on the MGB dataset, an AUROC of 0.888 and an AUPRC of 0.823 on the HEP dataset, and an AUROC of 0.995 and an AUPRC of 0.991 on the SAI dataset, with 32% of experts outperforming the model. The false-positive rate was reduced to an average of nine spikes per hour. SpikeNet2 offers expert-level accuracy in both event-level spike detection and EEG-level classification, while significantly reducing false positives. Its dual functionality and robust performance across diverse datasets make it a promising tool for clinical and telemedicine applications, particularly in resource-limited settings. (Funded by the National Institutes of Health and others.)

Sleep underpins cognition, disease prevention, and overall brain health, yet objective, integrative biomarkers of brain health remain lacking. We hypothesized that overnight sleep electroencephalography (EEG) could provide a substrate for such a biomarker. We asked whether a newly developed, end-to-end, data-driven deep learning framework for sleep EEG can learn a latent representation of brain health and distill it into a single score relevant to cognition, disease status, and mortality. We analyzed 36,000 polysomnography recordings from 27,000 subjects from six cohorts. EEG data were represented as one-dimensional time series or a two-dimensional time-frequency spectrogram. A multitask deep neural network, trained end-to-end without expert-defined features, learned a 1024-dimensional brain health latent space and jointly predicted cognitive performance, disease status, and sleep metrics. The latent representation was additionally distilled into a single brain health score. We compared performance with demographic baselines, conventional EEG metrics (e.g., rapid eye movement fraction, spindle density), and classic multivariate machine learning approaches. The deep learning–derived brain health scores consistently surpassed demographic and expert-defined EEG feature models. For cognitive outcomes, correlations (r) rose from small (demographic-only) to moderate (up to r=0.40), while disease classification areas under the receiver operator curve improved from 0.50–0.55 at baseline to 0.65–0.75. In age-adjusted Cox models, a one-standard-deviation increase in the brain health score was associated with a 31%–35% reduced risk of mortality (hazard ratio 0.65 to 0.69; P<0.0001), topping conventional EEG metrics. Gains over classic machine learning, plus latent space visualization, indicated that both established physiological markers and novel EEG features drove enhanced performance. A multitask, end-to-end deep learning approach generated an interpretable, sleep-derived brain health biomarker. By modeling cognition, disease, and mortality, this framework provides a robust index of brain health and may be extended to additional modalities, further enhancing its clinical utility. (Funded by the National Institutes of Health and others.)

Purpose: Population level tracking of post-stroke functional outcomes is critical to guide interventions that reduce the burden of stroke-related disability. However, functional outcomes are often missing or documented in unstructured notes. We developed a natural language processing (NLP) model that reads electronic health records (EHR) notes to automatically determine the modified Rankin Scale (mRS). Method: We included consecutive patients (⩾18 years) with acute stroke admitted to our center (2015–2024). mRS scores were obtained from the Get With the Guidelines registry and clinical notes (if documented), and used as the gold standard to compare against NLP-generated scores. We used text-based features from notes, along with age, sex, discharge status, and outpatient follow-up to train a logistic regression for prediction of good (0–2) versus poor (3–6) mRS, and a linear regression for the full range of mRS scores. The models were trained for prediction of mRS at hospital discharge and post-discharge. The models were externally validated in a dataset of patients with brain injuries from a different healthcare center. Findings: We included 5307 patients, 5006 in train and test and 301 in validation; average age was 69 (SD 15) and 65 (SD 17) years, respectively; 47% female. The logistic regression achieved an area under the receiver operating curve (AUROC) of 0.94 [CI 0.93–0.95] (test) and 0.94 [0.91–0.96] (validation), and the linear model a root mean squared error (RMSE) of 0.91 [0.87–0.94] (test) and 1.17 [1.06–1.28] (validation). Discussion and Conclusion: The NLP-based model is suitable for use in large-scale phenotyping of stroke functional outcomes and population health research.

Background: Identifying neuroinfectious disease (NID) cases using International Classification of Diseases billing codes is often imprecise, while manual chart reviews are labor-intensive. Machine learning models can leverage unstructured electronic health records to detect subtle NID indicators, process large data volumes efficiently, and reduce misclassification. While accurate NID classification is needed for research and clinical decision support, using unstructured notes for this purpose remains underexplored. Objective: The objective of this study is to develop and validate a machine learning model to identify NIDs from unstructured patient notes. Methods: Clinical notes from patients who had undergone lumbar puncture were obtained using the electronic health record of an academic hospital network (Mass General Brigham [MGB]), with half associated with NID-related diagnostic codes. Ground truth was established by chart review with 6 NID-expert physicians. NID keywords were generated with regular expressions, and extracted texts were converted into bag-of-words representations using n-grams (n=1, 2, 3). Notes were randomly split into training (80%), 2400 notes out of 3000, and hold-out testing (20%), 600 notes out of 3000, sets. Feature selection was performed using logistic regression with L1 regularization. An extreme gradient boosting (XGBoost) model classified NID cases, and performance was evaluated using the area under the receiver operating curve (AUROC) and the precision-recall curve (AUPRC). The performance of the natural language processing (NLP) model was contrasted with the Llama 3.2 auto-regressive model on the MGB test set. The NLP model was additionally validated on external data from an independent hospital (Beth Israel Deaconess Medical Center [BIDMC]). Results: This study included 3000 patient notes from MGB from January 22, 2010, to September 21, 2023. Of 1284 initial n-gram features, 342 were selected, with the most significant features being “meningitis,” “ventriculitis,” and “meningoencephalitis.” The XGBoost model achieved an AUROC of 0.98 (95% CI 0.96‐0.99) and AUPRC of 0.89 (95% CI 0.83‐0.94) on MGB test data. In comparison, NID identification using International Classification of Diseases billing codes showed high sensitivity (0.97) but poor specificity (0.59), overestimating NID cases. Llama 3.2 improved specificity (0.94) but had low sensitivity (0.64) and an AUROC of 0.80. In contrast, our NLP model balanced specificity (0.96) and sensitivity (0.84), outperforming both methods in accuracy and reliability on MGB data. When tested on external data from BIDMC, the NLP model maintained an AUROC of 0.98 (95% CI 0.96‐0.99), with an AUPRC of 0.78 (95% CI 0.66‐0.89). Conclusions: The NLP model accurately identifies NID cases from clinical notes. Validated across 2 independent hospital datasets, the model demonstrates feasibility for large-scale NID research and cohort generation. With further external validation, our results could be more generalizable to other institutions.

Background: Multicenter electronic health records (EHRs) can support quality improvement and comparative effectiveness research in critical care. However, limitations of EHR-based research include challenges in abstracting key clinical variables, including a patient’s level of consciousness. Objective: This study aimed to develop a natural language processing model to predict Glasgow Coma Scale (GCS) scores from daily EHR notes. Methods: The study included adult patients (aged ≥18 years) admitted to Mass General Brigham (MGB) hospitals (2017‐2024) and patients from the Medical Information Mart for Intensive Care-III (MIMIC-III version 1.4; 2001‐2012) database. A dataset of all patients from both institutions was split into training (70%) or hold-out test (30%) sets. Variables consisted of daily notes, age, sex, and admission type. A pooled ordinal regression model (ordinalNet) with an elastic net penalty was trained to predict the lowest daily level of consciousness across 3 classes of impairment: severe (GCS score 3‐8), moderate (GCS score 9‐12), and mild (GCS score 13‐15), and a pooled linear model was trained to predict continuous GCS scores (3-15). Gold standard GCS was obtained from structured flowsheet data. External generalizability was assessed using a single-institution ordinal model trained on MGB and tested on MIMIC. Following post hoc calibration, the performance of the ordinal and linear models was evaluated on the hold-out test sets using the area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC) for the ordinal models and root mean square error and Pearson correlation coefficient for the linear models. Results: The modeling cohort included 145,897 patients (MGB: n=123,257, MIMIC: n=22,640), with 1,446,965 days of hospitalization between training and testing sets; the average age was 62 (SD 18) years, and the sex distribution was balanced. The pooled ordinalNet achieved an AUROC of 0.96 (95% CI 0.96‐0.96) and an AUPRC of 0.77 (95% CI 0.76‐0.77). The single-institution ordinal model achieved an AUROC of 0.90 (95% CI 0.89‐0.90) and an AUPRC of 0.80 (95% CI 0.79‐0.80). The pooled linear model achieved a root mean square error of 2.30 (95% CI 2.30‐2.30) and a correlation of 0.76 (95% CI 0.76‐0.76). Predictions for severe GCS were driven by terms indicating unresponsiveness and critical interventions, moderate GCS by intermediate alertness descriptors, and mild GCS by mentions of normal or awake behavior. Conclusions: Pooled ordinal and linear models can accurately predict GCS from unstructured data and can support large-scale phenotyping of neurological assessments for future critical care research.