Introduction

Large language models (LLMs) have emerged as a powerful tool for analyzing and interpreting enormous amounts of data. Adding to the fervor is the capacity of LLMs as a form of generative artificial intelligence (AI) able to construct meaningful and contextually appropriate text based on a given prompt, emulating human-like creativity, and reasoning. The excitement and speculation generated by recent reports and media coverage on the potential of LLMs has led to additional questions posed in the public sphere surrounding their appropriate use.1,-,3 One of the primary reasons for the surging public interest is the ability of LLMs to generate text that is on-par, if not better than humans, when prompted with questions. For example, GPT-4, OpenAI's latest LLM, has scored high enough to pass all 3 parts of the US Medical Licensing Examination.4 Such models provide an opportunity to help address existing scientific, clinical, and social needs.

LLMs are deep learning frameworks designed to process natural language text (Table 1 and Table 2 for a glossary of technical terms).5,6 Unlike more traditional machine learning models, such as naïve Bayes classifiers, which rely on explicit labels (“happy” or “sad”), features (for example, full words or phrases), and rules to identify patterns, LLMs learn to recognize patterns and fill gaps or generate text using deep learning with vast amounts of data. LLMs are typically trained using large text corpora, such as text on the Internet, Wikipedia, books, newspaper articles, and other documents. Once an LLM has been trained, it can be used to perform a variety of tasks, such as language translation, text summarization, and generation of human-like text.

Table 1

Types of Large Language Models

Table 2

Glossary of Technical Terms

In the context of research related to neurologic disorders, LLMs could be trained de novo or “fine-tuned” on large clinical data sets to identify patterns and relationships that would be difficult for humans to detect manually.7,-,9 This article aims to provide researchers and clinicians with a discussion of this emerging technology and highlighting its potential for study, treatment, and care.

How Large Language Models Work

LLMs are built using neural network architectures that allow them to recognize complex patterns in natural language data and generate realistic text. Although there are several language models published in the literature,7,10,11 most LLMs use a specific type of neural network called a transformer, which is designed to handle long sequences of text. These neural networks are organized into multiple layers, with each layer consisting of multiple attention heads and feedforward neural networks (see Table 2 for a glossary of these terms). Attention heads are components that decide which parts of the input text the model should focus on when it generates output, helping to understand context and relationships between words. These attention heads consist of matrix multiplications and other mathematical operations. Feedforward neural networks then process the output of the attention heads and compute higher-order relationships between these dependencies and contextual relationships.

To train an LLM (Figure 1), large natural language data sets are used to determine the weights of the neural network. During training, the LLM is presented with input sequences and asked to predict the next word, fill “masked” words, or generate a new sentence based on the input. By minimizing the difference between the predicted output and the actual output, the LLM gradually learns to recognize patterns in the data and generate text that is consistent with the input.

Figure 1 Schematic of the General Process for Training and Testing a Large Language Model

(1) Data preparation: This step involves collection and preprocessing a large corpora of text data to be used for model training. (2) Pretraining: This step involves training the model on the large corpus of text data using unsupervised learning frameworks. The goal is to predict missing words or predict the next word in a sentence, given the previous words. (3) Training: Here, the model is further trained using a more specific data set, often involving human supervision. This stage includes the process of “reward modeling” where the model generates a set of potential responses, and human reviewers rank them. The model uses this ranking to generate responses in the future. This stage is crucial to ensuring the alignment of the model's outputs with human values and instructions. (4) Fine-tuning: This step involves improving the trained model on a smaller, labeled data set for a specific task. (5) Testing and evaluation: This key step involves validation of the fine-tuned model on an external (i.e., separate) data set and evaluate the model's performance on a specific task. Metrics, such as accuracy, precision, recall, and F1-score, can be used for model evaluation. (6) Deployment: If the validated model meets the desired performance criteria, then it can be used to perform the specific task in a real-world setting.

Language models can be trained using different methods,12,-,16 such as supervised learning, unsupervised learning, or self-supervised learning. Supervised learning is the most common approach to training language models, where a model is trained on a data set of labeled data and learns to predict the correct label for a given input. When there are no labeled data, then unsupervised learning can be used. An unsupervised learning model could be trained to generate new text by being given a large corpus of text in which it can learn patterns. Self-supervised learning is a newer approach to training language models that has been shown to be effective in learning complex relationships between words and phrases.17 In this case, the model is trained on a data set that has been artificially labeled, and the model learns to predict a missing label for a given input. For example, a self-supervised learning model could be trained to answer questions by being given a data set of questions and answers. The model would learn to predict the answer to a given question by finding patterns in the data. For LLMs, such as GPT-4, designed to respond well to questions and prompts, the system's output is improved by letting it interact with human testers and applying reinforcement learning techniques.

In addition to the stages presented in Figure 1, LLM development should consider how biases are addressed during the training process. This can be accomplished by, for example, ensuring that the training data are representative of the relevant population. For population-level queries, it is crucial to integrate fairness metrics during the fine-tuning process to assess model performance across different subgroups. Explicit instructions can be given to the model during this phase to avoid bias, thus further promoting unbiased outputs. Improving alignment between model output and the relevant task by iterating on the model's behavior should also be performed. Improving the clarity of guidelines given to human reviewers and developing upgrades to allow users to customize the model's behavior within broad bounds may aid in achieving this alignment. These measures ensure that the model's decisions are interpretable and explainable, allowing us to better understand any underlying bias. Regular audits should also be performed to monitor ongoing model outputs. Overall, these proactive steps need to be integrated to ensure that the LLM development process is mindful of biases and maintains a consistent alignment with human values.

Owing to the ability of these models to process enormous amounts of data, including medical records and patient interviews, and generate high-quality text that accurately reflects the complex symptoms and experiences of patients, LLMs constitute suitable tools for neurologic research and practice. LLM development has evolved over the past decade (Table 3), and the current state-of-the-art models can perform many tasks, including language modeling, text classification, and sentiment analysis.

Table 3

Timeline of the Development of LLMs

Language modeling is a fundamental task in natural language processing (NLP) that involves training LLMs to predict the next word in a sentence based on the context of the previous words. This task is often referred to as autoregression and can be performed in 2 ways: left to right or right to left. In left-to-right language modeling, the LLM predicts the next word in a sentence based on the context of the words to its left. By contrast, right-to-left language modeling involves predicting the next word based on the context of the words to its right. This function of LLMs can be bidirectional,5 predicting the next word based on context either to the left or to the right. This flexibility is important when dealing with conversational or narrative data where the meaningful context can come from either direction.

Having learned to predict and fill in “masked” words, LLMs could theoretically aid in communication for those with language impairment due to dementia or a traumatic brain injury. For these persons, this could mean using context to fill in gaps in a patient's narrative caused by memory loss, expressive aphasia, or poor engagement in conversation. This could potentially facilitate better communication between patients and their loved ones or caregivers.

LLMs For Cognitive Assessment and Rehabilitation

LLMs could also be applied to analyze language patterns in patients' spoken or written communication, potentially revealing cognitive shifts or deficits. For instance, by training LLMs on language data collected from patients with Parkinson's disease, at high risk for Huntington Disease, or at high risk for Alzheimer disease, it may become possible to detect subtle variations that evolve gradually, which human observers might overlook. These variations could encompass alterations in word production, vocabulary choice, sentence structure, or the sophistication of concepts over time. Detecting such audio and linguistic changes early on could enable timely intervention and tailored rehabilitation strategies. In a similar vein, LLMs could provide useful tools to augment clinician expertise in identifying language deficits in persons who experienced traumatic brain injury or undergone tumor resection, possibly facilitating more targeted cognitive rehabilitation.

As proof of concept, data sets from 2 recent challenges (ADReSS and ADReSSo) inspired the research community to develop automated methods to analyze speech, acoustic, and linguistic patterns in individuals to detect cognitive changes.18,19 Valsaraj et al.20 leveraged pretrained BERT to extract features on the autogenerated transcripts and assess cognitive function. Similar work was performed by Vats et al.,21 where they used BERT to perform dementia classification. Our own group previously showed that frameworks such as BERT and neural network–based sentence encoding can be used to automatically transcribe digital voice recordings and differentiate cognitively impaired persons from those with normal cognition.22 Agbavor and Liang23 similarly leveraged GPT-3 to develop a model to predict dementia in persons using their spontaneous speech.

Cognitive rehabilitation itself could also benefit from language modeling. For example, based on data about a patient's linguistic abilities, an LLM could generate word games or storytelling activities that match the patient's current cognitive level. By tracking the patient's performance over time, the model could adjust the difficulty of the tasks, providing a form of dynamic cognitive stimulation and training. Tasks assessing semantic (category) and phonemic (letter) fluency are commonly used in neuropsychological evaluations for cognitive impairment.24 In these tasks, individuals are asked to generate as many words as possible from a specific category or starting with a specific letter within a given time. LLMs could be used to automate the analysis of these tasks, providing scores based on not just the number and correctness of words generated but also the uniqueness of the temporal speed variations. This could lead to more objective, reliable, and efficient scoring of these assessments.

The implementation of LLM-based chatbots represents another transformative aspect with exciting potential.8,25,26 These digital assistants could be programmed to respond to frequently asked questions about their condition, propose various care management options, or even offer emotional support to patients and caregivers.

Electronic Health Record Text Classification

Text classification is a common NLP task where LLMs are trained to classify a given text into categories.27 This task involves providing a model with a set of text examples and corresponding category labels to learn patterns and relationships between the 2, thereby helping to automate the process of assigning categories to new text data. Named entity recognition is a similar task more restricted to assigning categories to individual “things” within a sentence; for example, the word “Boston” or “London” might be assigned the category “city.”

To train an LLM for text classification in a clinical context, text data, such as medical records, patient histories, or clinical trial reports, must be preprocessed. This processing involves tokenizing the text into individual words or subwords, removing irrelevant words, and applying various forms of normalization, such as stemming or lemmatization (Table 2). In brief, stemming and lemmatization are techniques in NLP that reduce words to their root form,28 with stemming chopping off the ends of words while lemmatization uses vocabulary and morphological analysis to find the base or dictionary form of a word, known as the lemma. Once preprocessed, an LLM is trained to predict a category label, such as a specific neurologic disorder, based on the features present in the text. LLMs can accurately perform text classification because of their capacity to learn intricate representations of text and automatically extract relevant features for classification.

As a potential application of text classification, clinical notes during patient encounters or personal narratives provided by patients about their symptoms could be processed and classified by LLMs to identify patterns of text that might correlate with specific neurologic conditions. This could help automate the categorization of new patient information into relevant classes such as ‘migraine’ and ‘Parkinson,' enabling more efficient analysis of patient data and serving as a clinical assistant. For example, Gehrmann et al.29 analyzed discharge summaries using NLP and convolutional neural networks, finding that they were able to categorize respective persons as having “chronic pain,” “advanced cancer,” or “advanced lung disease,” among others. This could be extended to relevant neurologic diagnoses or categories. In the context of neuroimaging, LLMs can identify noteworthy information in radiology reports in emergency department settings.30 In cases of stroke, for instance, where timely intervention is critical and patient communication may be impaired due to aphasia or other neurologic deficits, an LLM could serve as an effective tool to flag essential neuroimaging findings for providers.

When applied to large-scale data sets, such as electronic health records (EHRs) or databases of scientific literature, LLMs could improve classification accuracy and help streamline the process of clinical observational research. For example, Fernandes et al.31 demonstrated that an NLP algorithm was able to assign neurologic disability outcomes after intensive care unit hospitalization, based on free-text clinical notes. In another study, Xie and coauthors used 3 different LLMs (BERT, RoBERTa, and Bio_ClinicalBERT) to comb through clinical notes and determine whether and how frequently patients had seizures.32 Furthermore, LLMs' abilities to learn from multiple languages allows for cross-lingual text classification,33 which could enable the classification of neurologic data regardless of the language, thereby benefitting global neurologic research and patient care.

Text classification could also assist interpretation of neuropsychological tests, brain imaging, neuropathology, and neurophysiology studies, such as electroencephalography and electromyography/nerve conduction study reports. It could help to automatically categorize parts of these reports into clinically relevant predefined classes (e.g., normal/abnormal findings, presence/absence of certain key terms). In addition, it could potentially identify trends across multiple assessments, such as during a complicated or prolonged hospitalization, providing a clearer picture of a patient's clinical trajectory over time. Overall, using LLMs for text classification could enhance the speed and efficiency of processing patient data.

Sentiment Analysis

Sentiment analysis involves training an LLM to identify an underlying sentiment or emotion expressed in text.34 Given the increase in patient provider communication occurring outside the direct face-to-face encounter, such as through patient portal messaging, language processing methods such as sentiment analysis could offer insights into patients' subjective experiences and emotional states, which could be profoundly affected by neurologic impairment and the understanding of which is critical to managing these conditions. The goal of sentiment analysis is to automatically identify the polarity of a text, such as a patient message, voice recording, or video recording, which could be, for example, positive, negative, or neutral. Such analysis could provide crucial insights into patients' psychological well-being, their experiences with various treatments, or the impact of their neurologic condition on their day-to-day life and signal the need for prioritizing dedicated behavioral and mental health resources for the patient.

To train an LLM for sentiment analysis, a labeled data set is required, consisting of text data and its corresponding sentiment labels. Creating an effective data set demands a careful, domain-specific approach. The labeling process in neurology, for example, should ideally involve annotators skilled not only in language but also familiar with the intricacies of neurologic disorders. They would assign sentiment labels to textual data, reflecting a range of emotional responses that are common to patients experiencing neurologic conditions and to their caregivers. Creating a suitable data set for neurology-centric sentiment analysis also calls for balance and representation. As with other common LLM tasks, labelling should cover a variety of neurologic conditions, treatments, and patient-caregiver interactions to avoid model bias and accurately capture the breadth of sentiment in this field. While some generic resources, such as the PhysioBank,35 can provide a base, researchers should also look for neurology-specific data. The LLM would then be trained to identify patterns and relationships between text samples and their respective sentiment labels.

When trained on large-scale data sets such as databases of patient narratives or clinical communication, LLMs can learn complex text representations, capturing subtle nuances and contextual information essential to understanding patients' sentiments. Moreover, LLMs can identify the sentiment of individual words or phrases within a text, providing deeper insights into the emotions and opinions expressed by patients or their caregivers. For instance, this could enable the identification of specific symptoms or side effects that cause substantial distress or therapies that lead to positive emotional responses. Consider a scenario where LLMs are trained on a diverse range of text data from individuals with depression. These models could learn to detect subtle changes in language patterns, sentiment, and expression that correlate with the progression or severity of depression, provide insights into early signs of worsening depression, or even predict episodes of heightened distress that signal to care teams that close follow-up is needed.

Technical and Ethical ChallengesNeed to Ameliorate Bias

The potential for bias in LLMs presents unique technical and ethical challenges. The complex and heterogeneous nature of neurologic disorders necessitates that any model used in this field be trained on diverse, representative, and consistently collected data to keep from propagating health care disparities between different groups of people. Most notably, sampling bias has the potential to propagate preexisting health care disparities given that neurologic conditions manifest differently across a range of demographics, including age, sex, and ethnic groups. For instance, Alzheimer disease presents with varying symptoms and progression rates that differ between individuals and demographic groups. For classification, if an LLM was trained predominantly on data from older adults older than 65 years, its utility in diagnosing early-onset Alzheimer disease could be compromised. Strategies, such as oversampling, underrepresented patient groups, and using debiasing techniques during model training could be used to counter this bias. Training transparency and interpretability techniques could help clinicians understand the reasoning behind LLM recommendations, enhancing trust and clinical adoption.

In the realm of sentiment analysis, sentiment misinterpretation could have significant clinical repercussions such as leading to inappropriate interventions or treatments. To mitigate this, rigorous training of LLMs on diverse language patterns, including nuances in emotional expression across different patient groups, is vital. In addition, integrating feedback loops with health care professionals to validate and fine-tune sentiment predictions could enhance accuracy.

Measurement bias can occur due to the variety of tools and methods used in neurologic assessments, such as cognitive tests, neuroimaging techniques, and neurologic examinations. Data collected from these disparate sources might introduce inconsistency and variability. This bias can be minimized by using standardized protocols for data collection and incorporating a wide range of data sources to train the model. Confirmation and reporting biases pose significant risks in the context of neurology because of the subjectivity involved in assessing symptoms, such as pain, fatigue, or cognitive changes. Overrepresentation or underrepresentation of these symptoms in the training data could result in a skewed model that fails to accurately predict these aspects in patients. Given these biases' potential to affect an LLM's output and thus potentially affect patient care, researchers must generate rigorous clinical evidence through controlled studies assessing the accuracy, benefits, risks, and adverse events of incorporating LLMs in neurology. Furthermore, neurologists must be aware of an LLM's limitations and understand its generalizability across different neurologic conditions and patient demographics. It is crucial for them to approach LLMs as an aid rather than a replacement for their clinical judgment and expertise.

Need for Careful Technical Validation

The inherent complexity of LLMs can pose challenges in neurology. For instance, addressing “hallucinations,"45 where a model might generate significant errors, is critical in neurology where precision in data interpretation is paramount. An example of a “hallucination” from ChatGPT is shown in Figure 2, where it inaccurately assigns a “young age” to a patient based on a clinical note fragment without any age given.

Figure 2 Conversation With GPT

Demonstration of ChatGPT's ability to summarize a clinical note and generate plausible differential diagnoses. Highlighted in green is an example of a “hallucination,” a current limitation of the technology, where in this case, ChatGPT “hallucinates” that the patient has a “young age,” although the age of the patient in the clinical vignette was never specified. The clinical summary is abbreviated from a publicly available resource (mtsamples.com).

To address this challenge, LLMs for neurology ought to undergo careful technical validation to ensure that they are safe and helpful for their intended uses. This validation should include not only generalizable methods such as cross-validation and independent testing on data sets with varied demographics but also tests relevant to specific neurologic disorders. For instance, a model's ability to accurately predict dementia onset from clinical notes or neuroimaging data should be tested using data not involved in the model's training. Furthermore, as detailed above, careful attention should be given to potential biases or limitations in the training data. If the training data overrepresent certain demographics, the model's output may not be accurate or reliable when applied to underrepresented groups. Rigorous methods should be used to mitigate bias during data collection and curation, and the model's performance should be tested across diverse demographics.

Need to Preserve Privacy and Maintain Data Security

Machine learning model development in general, and specifically LLMs, presents significant privacy and data security concerns that must be addressed to protect the rights and confidentiality of study participants and patients.46 In addition to privacy concerns, there are data security concerns associated with the use of LLMs in clinical practice. LLMs are complex models that require significant computational resources to train and run. Researchers must ensure that appropriate data security measures are in place to protect the models and the data used to train them, such as secure cloud-based storage and access controls, and to prevent data breaches, such as regular security audits, data encryption, and secure data transmission. To address these concerns, researchers must ensure that they comply with relevant data protection laws and regulations, such as the General Data Protection Regulation in the European Union and the Health Insurance Portability and Accountability Act in the United States.

Several aspects regarding the usage of LLMs in practice must be carefully considered to ensure that the research is conducted in a responsible and transparent manner, particularly with respect to the principle of autonomy, and the right to decide how one's protected health information (PHI) is used by LLMs. The screening of large EHR databases may require special notification to patients who are vulnerable and may require a waiver of consent granted by institutional review boards to use LLMs to screen EHR data. Inherent to this task is ensuring the privacy and confidentiality of the data being used to train the models. PHI must be carefully protected to avoid any unintended harm or discrimination, particularly against individuals who may have impairment or disability.

Another challenge associated with the use of LLMs in neurology is obtaining proper informed consent from patients or their legally authorized representatives, including in situations when the initial consent to the use of PHI data is given when the individual is cognitively intact, and only later becoming cognitively impaired. It is critical that the evaluation of institutional review boards be included when making determinations about appropriateness of consent, particularly in the context of the evolution of consent as new scientific advances continue to emerge.

Role of Regulation

Federal regulations could serve as a useful adjunct to technical and clinician expertise in addressing the limitations and challenges of LLMs. Many areas of regulation lie outside the scope of this article, although there are several regulatory issues that are particularly important with respect to neurology and more broadly, medicine, that pertain to the technical and ethical challenges we raise.19,47,48

Software as a medical device (SaMD) is defined by the International Medical Device Regulators Forum (IMDRF), as software that is not embedded within hardware and which performs medical tasks. Therefore, many medical LLMs would fall under this umbrella. As Gilbert et al.19 note, even LLM chatbots used for clinical decision support could be considered medical devices. Under the IMDRF framework, adopted as guidance by the FDA, SaMD would need to meet 3 standards during clinical evaluation. These include (1) that there must be an association between SaMD output and the relevant clinical condition; (2) that an input generates “accurate, reliable, and precise” output; and (3) that the output achieves the desired goal in the population of interest. Any regulatory efforts should keep these standards in mind.

Bazoukis et al.49 introduce the idea of incorporating algorithm auditing to augmented intelligence models. Adapted to LLMs specifically, this could involve labelling models with segments of the population on which a particular model may be less effective or even untested. In addition, regulatory bodies could introduce mandated testing of LLMs on a private validation data set with demographics representative of the general population or specific marginalized populations. These steps, mandating testing on standardized data sets and labelling algorithms with expected performance on different population segments, could help to address bias and technological validation during an initial approval process and would help LLMs meet the standards set by the IMDRF framework for clinical evaluation.

Regulation may also be helpful in safeguarding patient data. There need to be specific timelines for removal of patient data from models and data sets and rules regarding the use of generative models pretrained on a patient's data if a respective patient wants their information removed. A balance between feasibility and patient safety must be navigated carefully, and new techniques may need to be developed to hasten this process.