Marta Fernandes

Objective. Periodic discharges (PDs) and rhythmic delta activity (RDA) are common electroencephalographic patterns in critically ill patients that require detailed characterization—including lateralization, spatial localization, and frequency estimation—according to the American Clinical Neurophysiology Society 2021 standardized terminology. Manual characterization is subjective, time-consuming, and exhibits substantial inter-rater variability. This study presents a comprehensive automated system for characterizing all four major subtypes: lateralized PDs (LPDs), generalized PDs (GPD), lateralized RDA (LRDA), and generalized RDA (GRDA). Approach. Two complementary pipelines were developed: the PD-Profiler and the RDA-Profiler. The PD-Profiler combines a per-channel convolutional neural network (Channelpd-Net) with hemisphere-specific learned evidence traces (HemiCET-UNet), dynamic programming with an approximately-periodic prior, and discharge-locked topographic localization. The RDA-Profiler uses iterative narrowband Hilbert refinement for frequency and lateralization, with phase-locking value analysis for spatial extent. Both pipelines were trained and evaluated on 12 425 EEG segments from 11 729 unique patients. Main results. Four electroencephalographers—three (SZ, TZ, AS) not involved in algorithm development, plus the primary annotator (MW) — independently scored 200 stratified segments per pattern subtype, with analysis restricted to segments accepted by a majority of the four raters. The algorithm exceeded expert–expert (EE) inter-rater agreement on LPD frequency (mean expert–algorithm (EA) intraclass correlation coefficient (ICC) 0.931 vs. EE 0.897; paired segment-bootstrap Δ = + 0.034 , 95% confidence interval [ + 0.016 , + 0.055 ] , p < 0.001 ), and was statistically indistinguishable from EE on every other attribute tested: GPD frequency (EA ICC 0.980 vs. EE 0.973; Δ = + 0.007 , [ − 0.000 , + 0.017 ] , p = 0.061 ), GRDA frequency ( Δ = − 0.020 , p = 0.056 ), LRDA frequency ( Δ = − 0.055 , p = 0.435 ), LPD laterality (EA κ 0.945 vs. EE 0.974; Δ = − 0.030 , p = 0.079 ) and LRDA laterality (EA κ 0.961 vs. EE 0.993; Δ = − 0.032 , [ − 0.087 , + 0.004 ] , p = 0.269 ). No attribute fell significantly below the EE ceiling. PD discharge timing achieved an F 1 of 0.817 at single-sample precision (mean absolute timing error 3.1 ms at 200 Hz; median error 0 samples). Pattern lateralization area under the receiver operating characteristic curve was 0.989 for PD hemisphere ( n = 1276 ), 0.909 for LPD vs. GPD ( n = 7037 ), and 0.837 for RDA ( n = 4253 ). PD spatial localization reached 97.3% of expert inter-rater Jaccard agreement; RDA spatial extent matched EE ICC (0.371 vs. 0.373), at a ceiling set by the inter-rater noise of verbal region-counting labels. Significance. This work presents the first system to reach expert-level inter-rater reliability across lateralization, spatial localization, discharge timing and frequency for both periodic and rhythmic EEG patterns: it exceeded the EE ceiling on LPD frequency and was statistically indistinguishable from that ceiling on every other attribute tested, with none falling significantly below it. Agreement remains lowest, though not significantly below the ceiling, on LRDA laterality and LRDA frequency, which we identify as the priority targets for further work. In post-hoc review of discordant cases, experts judged the algorithm’s frequency estimates more accurate than the original expert labels in 94% of cases. Automated characterization can now both substitute for and improve manual annotation in critical-care EEG.

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.

Background. Epileptiform activity (EA), including seizures and rhythmic and periodic patterns (RPPs), such as generalized periodic discharges (GPD), lateralized periodic discharges (LPD), and lateralized rhythmic delta activity (LRDA), are frequently observed in critically ill patients and associated with poor outcomes. However, the roles of RPP burden versus frequency, and differences across pattern subtypes, remain uncertain. Methods. We performed a retrospective multicenter study of adults (>18 years) admitted to three Boston hospitals (2009-2024) who underwent ³2h of EEG monitoring. Automated validated algorithms segmented EEG data, classified patterns, and quantified frequency characteristics, generating 906 variables reflecting burden, frequency, and combined metrics across three time windows (first 24h, first 72h, and peak 6h burden). The primary outcome was poor discharge modified Rankin Scale (mRS) score (3-6). Secondary outcomes included mortality and favorable 90-day outcome (mRS 0-2). Associations were evaluated using regularized regression, Cox models, and Fine-Gray competing risk analysis with adjustment for clinical covariates. Findings. Of 34773 screened individuals, 20675 met inclusion criteria (median age 64 years [IQR 50-74], 46.7% female). Greater RPP burden was consistently associated with worse discharge mRS, lower odds of favorable outcome, and increased mortality. Each 25% increment in peak 6h burden corresponded to ~20% higher risk of poor outcome, particularly for RPPs overall and GPDs. GPD and LPD demonstrated the strongest associations with adverse outcomes, whereas LRDA showed no or inverse association. Frequency-only measures had limited predictive value, and incorporating frequency into burden-based models did not improve discrimination (AUC 0.639). Results were consistent in recordings >24h. Interpretation. RPP burden has stronger association with outcomes compared with frequency. GPD and LPD burden are strongly associated with poor prognosis, while LRDA appears less harmful. These findings support a pattern-specific, burden-focused clinical management and highlight the need for prospective trials testing burden-based treatment strategies.

Accurate identification of acute and subacute subdural hematoma (acute/subacute SDH) is critical for improved patient outcomes. However, large-scale research is hindered by unreliable identification methods in electronic health records (EHRs). Current approaches relying on International Classification of Diseases (ICD) codes lack specificity and cannot distinguish acute, subacute, and chronic cases; manual chart review is too labor-intensive to scale. We developed an automated phenotyping algorithm using structured data and unstructured clinical notes for high-accuracy retrospective identification of acute/subacute SDH. We analyzed 2999 records from two hospitals, including ICD-positive and ICD-negative acute/subacute SDH cases verified by manual chart review. Features for model training included ICD codes, Current Procedural Terminology (CPT) codes, and clinical note keywords. Logistic regression and random forest models were trained using cross-validation and evaluated using AUROC and AUPRC. External validation involved training on one hospital and testing on the other. The random forest keywords-only model performed best, achieving an AUROC of 0.985 (95% CI: 0.980–0.990) and AUPRC of 0.944 (95% CI: 0.923–0.962) on the test set. External validation demonstrated strong AUROCs of 0.965 and 0.971 and AUPRCs of 0.831 and 0.840. The overall error rate was <1%. This model provides a scalable, highly accurate approach to acute/subacute SDH detection in EHR research.

Objective. Rhythmic and periodic patterns (RPPs) are harmful brain activity observed on electroencephalography (EEG) recordings of critically ill patients. This work describes automatic methods for detection of the frequency and spatial extent of specific RPPs: lateralized and generalized rhythmic delta activity (LRDA, GRDA) and lateralized and generalized periodic discharges (LPD, GPD). Approach. The frequency and spatial extent of RPPs is estimated using signal processing and rule-based logic. Three algorithm variants based on fast Fourier transform (FFT) and Hilbert–Huang transforms (HHT) were developed for rhythmic delta activity, and three using derivative and time-based peak detection for periodic discharges. Annotations from three expert neurophysiologists served as the gold standard, and inter-rater reliability (IRR) and mean absolute error (MAE) were used to assess performance. Main results. We evaluated the algorithms on segments with 100% agreement on event classification ( n = 389) and on the full cohort of 1087 segments (including disagreements). For the first subset, top algorithms matched or exceeded expert agreement for RPP frequency/spatial extent. RDA1b-FFT, the best algorithm for rhythmic delta activity, showed an expert-algorithm IRR of good to excellent with an intra-class correlation coefficient (ICC) of 91% and 96% (MAE 0.13 Hz and 0.26 Hz) frequency, and ICCs of 85% and 66% (MAE 0.19 and 0.09) for spatial extent for LRDA and GRDA. For periodic discharges, PD2a, showed and expert-algorithm IRR ICC of 80% and 61% (MAE 0.41 Hz and 0.15 Hz) for frequency, and ICC 77% and 13% (MAE 0.17 and 0.40) for spatial extent of LPD and GPD. For the full cohort, IRR declined, but expert-algorithm IRR remained comparable or superior to experts. Significance. The presence of RPPs at high frequencies and spatial extent are associated with a higher probability of poor outcomes. The proposed algorithms for estimating frequency and spatial extent of RPPs match expert performance and are a viable tool for large-scale EEG analysis.

Acquired epilepsy is a disabling, potentially preventable complication of acute brain injury (ABI). Yet, it remains a leading cause of new onset epilepsy in adults. Acquired epilepsy is a particularly challenging subtype of epilepsy to identify, as the ABI and its acute consequences can confound the later diagnosis of acquired epilepsy. We identified a retrospective cohort of patients with ABI (N=828) and optimized a general epilepsy algorithm to extract relevant keywords. We confirmed that applying a broad epilepsy phenotyping algorithm to a high risk, complex population like acute brain injured patients results in a high number of false positives. We developed multivariate models to identify ABI-acquired epilepsy 1) at the patient level using temporal trends, and 2) the note-level using keywords. Our models achieved high performance in both internal and external validation cohorts. Note-level re-classification also allowed for an estimation of time to epilepsy onset. This work enables large-scale, retrospective studies of ABI-acquired epilepsy across sites. Large-scale implementation may provide insights into acquired epilepsy epidemiology, identification of novel epilepsy risk factors and ultimately new treatments.

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.