OBJECTIVE: Large-language models (LLMs) can potentially revolutionize health care delivery and research, but risk propagating existing biases or introducing new ones. In epilepsy, social determinants of health are associated with disparities in care access, but their impact on seizure outcomes among those with access remains unclear. Here we (1) evaluated our validated, epilepsy-specific LLM for intrinsic bias, and (2) used LLM-extracted seizure outcomes to determine if different demographic groups have different seizure outcomes. MATERIALS AND METHODS: We tested our LLM for differences and equivalences in prediction accuracy and confidence across demographic groups defined by race, ethnicity, sex, income, and health insurance, using manually annotated notes. Next, we used LLM-classified seizure freedom at each office visit to test for demographic outcome disparities, using univariable and multivariable analyses. RESULTS: We analyzed 84 675 clinic visits from 25 612 unique patients seen at our epilepsy center. We found little evidence of bias in the prediction accuracy or confidence of outcome classifications across demographic groups. Multivariable analysis indicated worse seizure outcomes for female patients (OR 1.33, P ≤ .001), those with public insurance (OR 1.53, P ≤ .001), and those from lower-income zip codes (OR ≥1.22, P ≤ .007). Black patients had worse outcomes than White patients in univariable but not multivariable analysis (OR 1.03, P = .66). CONCLUSION: We found little evidence that our LLM was intrinsically biased against any demographic group. Seizure freedom extracted by LLM revealed disparities in seizure outcomes across several demographic groups. These findings quantify the critical need to reduce disparities in the care of people with epilepsy.

OBJECTIVE: Seizure frequency and seizure freedom are among the most important outcome measures for patients with epilepsy. In this study, we aimed to automatically extract this clinical information from unstructured text in clinical notes. If successful, this could improve clinical decision-making in epilepsy patients and allow for rapid, large-scale retrospective research. MATERIALS AND METHODS: We developed a finetuning pipeline for pretrained neural models to classify patients as being seizure-free and to extract text containing their seizure frequency and date of last seizure from clinical notes. We annotated 1000 notes for use as training and testing data and determined how well 3 pretrained neural models, BERT, RoBERTa, and Bio_ClinicalBERT, could identify and extract the desired information after finetuning. RESULTS: The finetuned models (BERTFT, Bio_ClinicalBERTFT, and RoBERTaFT) achieved near-human performance when classifying patients as seizure free, with BERTFT and Bio_ClinicalBERTFT achieving accuracy scores over 80%. All 3 models also achieved human performance when extracting seizure frequency and date of last seizure, with overall F1 scores over 0.80. The best combination of models was Bio_ClinicalBERTFT for classification, and RoBERTaFT for text extraction. Most of the gains in performance due to finetuning required roughly 70 annotated notes. DISCUSSION AND CONCLUSION: Our novel machine reading approach to extracting important clinical outcomes performed at or near human performance on several tasks. This approach opens new possibilities to support clinical practice and conduct large-scale retrospective clinical research. Future studies can use our finetuning pipeline with minimal training annotations to answer new clinical questions.

Objective: New approaches are needed to interpret large amounts of physiologic data continuously recorded in the ICU. We developed and prospectively validated a versatile platform (IRIS) for real-time ICU physiologic monitoring, clinical decision making, and caretaker notification. Methods: IRIS was implemented in the neurointensive care unit to stream multimodal time series data, including EEG, intracranial pressure (ICP), and brain tissue oxygenation (PbtO2), from ICU monitors to an analysis server. IRIS was applied for 364 patients undergoing continuous EEG, 26 patients undergoing burst suppression monitoring, and four patients undergoing intracranial pressure and brain tissue oxygen monitoring. Custom algorithms were used to identify periods of elevated ICP, compute burst suppression ratios (BSRs), and detect faulty or disconnected EEG electrodes. Hospital staff were notified of clinically relevant events using our secure API to route alerts through a password-protected smartphone application. Results: Sustained increases in ICP and concordant decreases in PbtO2were reliably detected using user-defined thresholds and alert throttling. BSR trends computed by the platform correlated highly with manual neurologist markings (r20.633-0.781; p <; 0.0001). The platform identified EEG electrodes with poor signal quality with 95% positive predictive value, and reduced latency of technician response by 93%. Conclusion: This study validates a flexible real-time platform for monitoring and interpreting ICU data and notifying caretakers of actionable results, with potential to reduce the manual burden of continuous monitoring services on care providers. Significance: This work represents an important step toward facilitating translational medical data analytics to improve patient care and reduce health care costs.

The epileptic network is characterized by pathologic, seizure-generating 'foci' embedded in a web of structural and functional connections. Clinically, seizure foci are considered optimal targets for surgery. However, poor surgical outcome suggests a complex relationship between foci and the surrounding network that drives seizure dynamics. We developed a novel technique to objectively track seizure states from dynamic functional networks constructed from intracranial recordings. Each dynamical state captures unique patterns of network connections that indicate synchronized and desynchronized hubs of neural populations. Our approach suggests that seizures are generated when synchronous relationships near foci work in tandem with rapidly changing desynchronous relationships from the surrounding epileptic network. As seizures progress, topographical and geometrical changes in network connectivity strengthen and tighten synchronous connectivity near foci-a mechanism that may aid seizure termination. Collectively, our observations implicate distributed cortical structures in seizure generation, propagation and termination, and may have practical significance in determining which circuits to modulate with implantable devices.

Patients with epilepsy can manifest short, sub-clinical epileptic “bursts” in addition to full-blown clinical seizures. We believe the relationship between these two classes of events—something not previously studied quantitatively—could yield important insights into the nature and intrinsic dynamics of seizures. A goal of our work is to parse these complex epileptic events into distinct dynamic regimes. A challenge posed by the intracranial EEG (iEEG) data we study is the fact that the number and placement of electrodes can vary between patients. We develop a Bayesian nonparametric Markov switching process that allows for (i) shared dynamic regimes between a variable numbers of channels, (ii) asynchronous regime-switching, and (iii) an unknown dictionary of dynamic regimes. We encode a sparse and changing set of dependencies between the channels using a Markov-switching Gaussian graphical model for the innovations process driving the channel dynamics. We demonstrate the importance of this model in parsing and out-of-sample predictions of iEEG data. We show that our model produces intuitive state assignments that can help automate clinical analysis of seizures and enable the comparison of sub-clinical bursts and full clinical seizures.

Clinical electroencephalography (EEG) is routinely used to monitor brain function in critically ill patients, and specific EEG waveforms are recognized by clinicians as signatures of abnormal brain. These pathologic EEG waveforms, once detected, often necessitate accute clinincal interventions, but these events are typically rare, highly variable between patients, and often hard to separate from background, making them difficult to reliably detect. We show that Deep Belief Nets (DBNs), a type of multi-layer generative neural network, can be used effectively for such EEG anomaly detection. We compare this technique to the state-of-the-art, a one-class Support Vector Machine (SVM), showing that the DBN outperforms the SVM by the F1measure for our EEG dataset. We also show how the outputs of a DBN-based detector can be used to aid visualization of anomalies in large EEG data sets and propose a method for using DBNs to gain insight into which features of signals are characteristically anomalous. These findings show that Deep Belief Nets can facilitate human review of large amounts of clinical EEG as well as mining new EEG features that may be indicators of unusual activity.

This paper describes an application of one-class support vector machine (SVM) novelty detection for detecting seizures in humans. Our technique maps intracranial electroencephalogram (EEG) time series into corresponding novelty sequences by classifying short-time, energy-based statistics computed from one-second windows of data. We train a classifier on epochs of interictal (normal) EEG. During ictal (seizure) epochs of EEG, seizure activity induces distributional changes in feature space that increase the empirical outlier fraction. A hypothesis test determines when the parameter change differs significantly from its nominal value, signaling a seizure detection event. Outputs are gated in a .one-shot. manner using persistence to reduce the false alarm rate of the system. The detector was validated using leave-one-out cross-validation (LOO-CV) on a sample of 41 interictal and 29 ictal epochs, and achieved 97.1% sensitivity, a mean detection latency of -7.58 seconds, and an asymptotic false positive rate (FPR) of 1.56 false positives per hour (Fp/hr). These results are better than those obtained from a novelty detection technique based on Mahalanobis distance outlier detection, and comparable to the performance of a supervised learning technique used in experimental implantable devices (Echauz et al., 2001). The novelty detection paradigm overcomes three significant limitations of competing methods: the need to collect seizure data, precisely mark seizure onset and offset times, and perform patient-specific parameter tuning for detector training.

We present a quantitative method for identifying the onset of epileptic seizures in the intracranial electroencephalogram (IEEG), a process which is usually done by expert visual inspection, often with variable results. We performed a fractal dimension (FD) analysis on IEEG recordings obtained from implanted depth and strip electrodes in patients with refractory mesial temporal lobe epilepsy (MTLE) during evaluation for epilepsy surgery. Results demonstrate a reproducible and quantifiable pattern that clearly discriminates the ictal (seizure) period from the pre-ictal (pre-seizure) period. This technique provides an efficient method for IEEG complexity characterization, which may be implemented in real time. Additionally, large volumes of IEEG data can be analyzed through compact records of FD values, achieving data compression on the order of one hundred fold. This technique is promising as a computational tool for determination of electrographic seizure onset in clinical applications.