Primer on large language models: an educational overview for intensivists (2025)

In 2022, the US Food and Drug Administration (FDA) approved 139 medical devices incorporating artificial intelligence (AI) [1]. By August 2024, this number had surged to 950 AI or machine learning-enabled medical devices [2], highlighting the exponential growth of AI integration into medicine in general and clinical practice in particular [3].

Artificial intelligence (AI) is broadly defined as the seeming ability of a device to perform tasks that typically require human cognitive skills, such as reasoning, learning, and decision-making. At its core, AI is intended to embody the capacity to “do the right thing” in a given context by employing logic-based methods efficiently and safely.

Early AI systems were designed to achieve specific outcomes for well-defined tasks and are therefore often referred to as task-specific AIs. Typical examples are the early customer service chatbots, which relied on rigid rules to recognize specific keywords and generate pre-programmed responses to these words. These simple systems relied on algorithms - mathematical constructs first formalized in the ninth century by the Persian mathematician Muhammad ibn Musa al-Khwarizmi. Algorithms define specific rules for specific outcomes and have been the foundation of decision-making models for centuries [4].

More recently, AI has gained an additional layer of sophistication - the ability to learn from vast datasets (i.e., big data). This progress has given rise to neural networks, which mimic the interconnected pathways of the brain to model complex relationships. Neural networks evaluate multiple possible pathways and select the most probable outcome (probabilistic reasoning). Bayesian statistics comprise a critical component of this process. Bayesian methods may be applied to all types of data, and they are typically used to calculate outcome likelihoods based on prior probabilities and new evidence. The resultant outputs of these methods are ranges of confidence related to the likelihood of the outcome (i.e., confidence intervals) rather than a definitive answer (i.e., cutoff levels). While highly effective, Bayesian approaches to data highlight a fundamental AI challenge - the need to balance statistical precision with uncertainty. Or, in other words, creative problem-solving.

Here enters generative AI, whose emergence marked another milestone in AI development. Generative AI models differ from earlier models in the production of outputs for new inputs, based on learned patterns rather than solely following predefined pathways. Generative AI models introduce an element of creativity, as they generate predictions (often also called insights) by synthesizing prior information that is frequently not overt. For example, AI systems trained on extensive imaging datasets can identify new patterns and diagnose novel cases [5]. Such advancements fuel ongoing debates regarding the limits of AI and the delicate balance between creativity, transparency, and replicability.

Of particular relevance to critical care is the AI field of machine learning (ML) that leverages vast datasets to train algorithms that infer logic and adapt to new scenarios. Using electronic health records (EHRs) as the data source, granular and diverse clinical data are aggregated to enable predictive modeling using ML. As discussed below, these datasets can be leveraged to empower AI systems to anticipate clinical decisions, optimize patient care pathways, and improve medical decision-making. This educational primer aims to teach intensivists the foundational concepts of large language models and how to approach emerging literature in this area.

Large language models (LLMs) are a type of machine learning specifically designed to understand the relationships between words and phrases [6]. LLMs learn the grammar, semantics, and contextual use of human language by processing vast amounts of data from diverse sources. A key component of this learning involves creating embeddings, mathematical representations of words. These allow the model to group words with similar meaning close together in a kind of language map [7]. For example, the words “heart rate”, “blood pressure”, “temperature”, and ”respiratory rate” would be grouped in the map as “vital signs”. These relationships help the model understand how words are used in different contexts.

These models originate from natural language processing (NLP), a field that employs computational techniques to represent text using algorithmic structures based on word co-occurrence frequencies within a given dataset [8]. Unlike traditional NLP models, which rely strictly on the data provided within a specific dataset, LLMs can incorporate contextual information and generate responses based on learned patterns. Table1 presents key terms related to augmented intelligence and large language models.

A key distinction for understanding how LLM algorithms function is the model distinction between words and tokens. Words are linguistic units that convey meaning, while tokens are discrete symbols that comprise or represent these units (e.g., sub-words or even punctuation marks). For instance, a simple word like “note” may be represented by a single token, while a longer or compound word such as “hospitalization” might be broken into two tokens (e.g., “hospital”, “ization”). Each token is assigned a probability of appearing next in a sequence based on the tokens that preceded it. The model employs a mechanism called self-attention to evaluate all parts of the input simultaneously, allowing it to determine contextual relationships and generate coherent output. In other words, this mechanism is designed to evaluate the relevance of different parts of the input when generating a response. For example, when a physician prompts an LLM to generate an admission or discharge summary, the model may refer to the patient using both “the patient” and the pronoun “his/her.” In the sentence, “The patient was given his medications,” the model is able to associate “his” with “the patient” because it considers the entire sentence holistically, not just word by word. This process ultimately results in an output reflecting the most likely and contextually appropriate sequence of tokens [9] - a “predictive language assembly” that enables the model to form medically coherent and grammatically accurate responses (Fig.1).

Translating words into tokens is called encoding, and translating these tokens back into text is called decoding. Embeddings are a critical component of LLM formulation of the context, nuance, and subtle meanings of words and phrases that undergo decoding and encoding. These embeddings represent each token’s place in the map of language space. LLM inputs are first transformed into tokens and embeddings, and after being processed as a series of layers in the model architecture, the output is created using the reverse process. This multi-layer architecture, known as a transformer encoder-decoder, allows the representations of tokens to be progressively refined. Each successive layer helps the model learn increasingly complex patterns, from basic grammar to more abstract semantic relationships. Statistical strategies are applied in the decoding process when the model selects one token over another based on learned probabilities - the likelihood of each token appearing in a given context based on training data. As a result of this selection process, the model generates text. Figure2 presents a simplified large language model processing algorithm.

Large language models in critical care

LLMs are still in their infancy concerning use in critical care, with very few clinically validated applications and a glaring lack of scientific consensus on their actual use. Today’s truths may be rapidly swept away by exponentially improving technology. Yet, considering how these tools might be integrated into critical care could help frame the essential discussion on their future role in clinical decision-making. LLMs will likely be used across the critical care continuum to assist in triage and documentation, diagnosis and prediction of patient deterioration, and patient management. LLMs have demonstrated a high median score for appropriateness in addressing clinical questions related to critical care [10] and have also been proposed to support patient rehabilitation after discharge from the ICU [11, 12].

As of this writing, at least six additional papers are under review in prepublication databases that propose using LLMs in critical care for treatment planning, patient care management, and prediction of deterioration and mortality.

Early identification of patients at risk of rapid clinical deterioration can improve triage and enable early response, including ICU admission and implementation of care interventions, which machine learning models aim to predict. Efforts are being made to improve the accuracy of ICU admission predictions, including integrating NLP technology to enhance the quality of the data used for model development [13].

Retrieval-augmented generation (RAG) is a framework that augments a Large Language Model (LLM) with updated data to generate more informed responses. The retrieval model accesses, selects, and prioritizes the most relevant documents and data from appropriate sources, transforms these into an enriched contextual prompt, and invokes the LLM through an application programming interface (API, see Table1) to generate the response. This type of machine learning is comparable to stock traders’ use of publicly available historical financial information and live market data feeds to make decisions. Numerous approaches to integrating LLMs for patient triage into clinical practice are being explored, including a RAG approach. Yazaki et al. used Chat-GPT3.5 with RAG to enhance contextual understanding and achieved a 70% accuracy rate in triaging emergency cases from the Japanese National Examination for Emergency Medical Technicians [14]. A retrospective study used real-world data from seven hospitals to evaluate ChatGPT-4’s accuracy in predicting hospital admissions after Department of Emergency Medicine visits. When RAG was incorporated, the prediction accuracy was 81.3% [15].

Documentation is essential to any clinical care process, starting from patient intake. LLMs have been proposed for creating clinical notes based on the assumption that such use may reduce physician burnout [16]. In the demanding field of critical care, where clinicians face significant workload pressures, yet daily notetaking must cover most physiological systems, such use may be particularly beneficial.

Madden et al. highlighted the transformative potential of ChatGPT in processing and synthesizing real-time summaries from the daily free-text entries from ICU electronic health records [17]. These entries, authored by doctors, nurses, and allied health professionals, often contain critical information but are typically informal, abbreviated, and poorly structured. In their study, ChatGPT-4 generated concise, actionable summaries and responded to queries (e.g., provided timelines of administered medications). A pilot feasibility study also demonstrated the ability of LLMs to generate concise summaries of ICU admissions for discharge documentation [18]. Another single-blind trial found that the quality of discharge letters generated by Chat-GPT4 was comparable to those written by junior clinicians [19].

Z codes (Z55-Z65) are the International Classification of Diseases, Tenth Revision Clinical (ICD-10)Modification diagnosis codes used to document social determinants of health data (e.g., housing, food insecurity, transportation, etc.). Guevara et al. investigated the potential use of LLMs for extracting six social determinants of health categories from narrative text in electronic health records, including employment, housing, transportation, and parental status. The best-performing models accurately identified 95.7% of patients with at least one mention of an SDoH category, compared to just 2.0% identified through structured Z-codes in the electronic health record during the same timeframe [20]. Early identification of poor social determinants of health could be invaluable for the prevention of ICU admissions as well as for planning for post-ICU rehabilitation, particularly in patients with complex medical conditions that are subsequently more likely to require such admission. Another study found that LLMs outperformed human coders in extracting ICD-10 codes from patient notes [21].

Informed consent is another critical domain of documentation. A cross-sectional study of surgical procedures revealed that LLM-based, chatbot-generated presentations of risk, benefit, and possible alternatives to surgery outperformed those presented by surgeons in both composite completeness and accuracy scores, based on expert evaluation and readability assessments, compared to presentations by surgeons. Based on these results, the authors suggested that LLMs be integrated into electronic health records to provide personalized risk and benefit assessments before performing invasive procedures [22].

Medication prescription and clinical documentation share similarities, requiring precise and detailed writing. Given the burden of work imposed on physicians, errors can occur in both processes. The “Healthy Technology Act of 2025,” a bill introduced in the 119 th Congress of the US House of Representatives (H.R.238), proposes permitting the use of AI for medication prescribing (albeit with precautions) [23]. This development introduces a new dimension to the potential role of AI, including LLMs, in this domain.

Finally, LLMs are used to summarize doctor-patient conversations during palliative care teleconsultations performed almost similarly in medical conversation summarization. Chat-GPT4 balanced content understanding and preserved structural similarity to the source somewhat better than other models, suggesting clinicians could use this LLM to generate medical summaries of such meetings. These summaries could then be given to the patient and/or their family, who may need to reflect on the content of the meeting [24].

Diagnostics is another field where LLMs can play a role in critical care. A randomized, double-blind crossover study compared the performance of the LLM tool AMIE (Articulate Medical Intelligence Explorer) to that of twenty primary care physicians during text-based consultations modeled after an Objective Structured Clinical Examination (OSCE). The study included 149 clinical vignettes evaluated by specialist physicians and patient actors. AMIE showed greater diagnostic accuracy and superior performance on 28 of the 32 axes assessed by the specialist physicians and 24 of the 26 axes evaluated by the patient actors [25].

Another retrospective cohort study conducted in a 40-bed PICU demonstrated the capability of domain-specific LLMs, such as those trained on specific medical data, to generate differential diagnoses. While their performance was inferior to that of human clinicians in terms of quality, pediatric critical care specialists gave them high evaluation scores [26].

In the complex diagnostic landscape of the ICU, the perspectives of family members and caregivers are often fraught with misinterpretations and unanswered questions. Scquizzato et al. evaluated the accuracy of ChatGPT in responding to non-professional questions about cardiac arrest. ChatGPT provided highly accurate answers, as assessed by clinicians and researchers specializing in out-of-hospital cardiac arrest, as well as by laypersons. Given the emotionally charged nature of scenarios such as cardiac arrest, which are integral to daily critical care practice, this suggests that leveraging the capabilities of LLMs to help address and clarify clinical situations for families has significant potential [27].

The last decades have seen a surge in the use of diagnostic imaging with a related increase in the need for an efficient image interpretation and reporting process. This rising workload has led to concerns regarding decreased efficacy and a higher likelihood of mistakes due to system overload and radiology staff burnout. Radiologists are expected to handle substantial textual information - from diagnostic request forms, medical charts and summaries, information from prior or other examinations, and the most updated medical literature. LLMs may be used to ameliorate this burden if used wisely. While this use may improve the efficiency of radiology services overall, those most likely to benefit are critically ill patients who often require frequent testing and rapid results. Medical imaging of critically ill patients poses unique challenges, including the need to meet stringent time frames and minimize complications stemming from redundant patient transfers. LLMs may be used to improve radiology service efficacy and effectiveness in ways that may be particularly relevant for critically ill patients.

A study comparing human radiologists to Chat-GPT-4 V and Gemini Pro Vision concluded that human radiologists still outperform these LLMs in diagnostic accuracy across various subspecialties (neuroradiology, gastrointestinal, genitourinary) but concluded that LLMs may potentially be used to support clinical decision-making [28]. Another model, CXR-LLaVa, which integrates an LLM with an image encoder, demonstrated 81% diagnostic accuracy in identifying six common clinical conditions from test sets of X-ray images using the Medical Information Mart for Intensive Care (MIMIC) database [29].

These findings have been supported by an additional study that showed that ClotCatcher, a natural language model with data augmentation, can rapidly and accurately identify venous thromboembolism (VTE) from radiology reports. The authors concluded that the model may improve the efficiency and accuracy of incident VTE adjudication in large databases [30].

Monitoring and early prediction of patient deterioration

Critically ill patients may rapidly deteriorate, a situation that requires early diagnosis and effective treatment decisions in complex clinical situations. Additional difficult decisions that typically need to be addressed in the critical care environment are those relating to patient preferences that must be made in conjunction with the families, often on behalf of patients unable to make decisions themselves. Time constraints, cultural reluctance to address end-of-life issues, and clinician burdens may limit the ability to elucidate individual patients’ value judgments and preferences. A proof-of-concept study explored the potential of LLMs to integrate patient values into critical care decision-making for incapacitated patients. Automated extractions of the treatment in question were accurate in 88% of scenarios. LLM treatment recommendations were rated by adjudicators with an average Likert score of 3.92 out of 5.00 for being medically plausible and reasonable and 3.58 out of 5.00 for reflecting documented patient values [31].

The possible use of LLMs to predict patient deterioration has also been explored in the context of respiratory failure and support. A machine learning model integrated with natural language processing, ARDSFlag, demonstrated an overall accuracy of 89.0% in identifying ARDS cases [32]. Another small, prospective study found that Chat-GPT4 demonstrated an accuracy comparable to that of specialized physicians in predicting the need for endotracheal intubation in patients receiving high-flow nasal cannula therapy for 48 h [33]. A third study used natural language processing to identify under-documentation of ARDS in ICU discharge notes [34]. Such use has more than just research implications - it can also serve as an educational tool.

Sepsis remains a leading cause of ICU mortality [35], yet remains a significant diagnostic challenge in the ICU. The SERA algorithm is an AI-enabled tool that uses natural language processing of physicians’ clinical notes using structured electronic medical records (EMR) data. SERA had a high predictive accuracy for identifying sepsis 12 h before its onset, with an AUC of 94%. Compared to physician predictions, the SERA algorithm increased early detection of sepsis by as much as 32% while reducing false positives by 17% [36].

Acute kidney injury (AKI) affects 30–57% of critically ill patients and is associated with high morbidity and mortality [37]. Among patients discharged from the ICU with normal renal function after AKI, almost one in three will relapse into renal failure within 5 years [38]. One study evaluated the effectiveness of Chat-GPT4 in teaching patients about AKI and continuous renal replacement therapy (CRRT). The model demonstrated a 97–98% overall accuracy, consistent performance across question types, and no significant differences between AKI and CRRT responses [39].

Critical care involves providing a broad spectrum of treatment regimens tailored to diverse clinical scenarios. Integrating LLMs into this process may optimize time and efficiency in care delivery. Howard et al. explored whether ChatGPT (version unspecified) may be used to provide antimicrobial treatment recommendations in eight hypothetical infection scenarios. While limitations were noted in addressing complex cases, ChatGPT demonstrated an overall ability to suggest appropriate antimicrobial spectra and regimens for the diagnoses and recognized the implications of clinical responses [40].

Delirium occurs in approximately 30% of ICU patients, with rates rising to 90% among mechanically ventilated patients [41]. Delirium is also associated with increased mortality after ICU admission. One of the studies still in preprint, suggests that DeLLirium - an LLM-based prediction model - achieved better results than other deep-learning models in predicting delirium from electronic health records [42]. Although this tool is primarily intended for critical care research rather than clinical practice, its potential for detecting delirium through conversations with patients or relatives may enable early identification of at-risk individuals. Alternative models may be developed for predicting post-ICU depression among patients and caregivers.

LLMs may also become an essential educational resource for families and caregivers after discharge from the ICU. For example, non-professional caregivers rarely have the training or preparation required for this challenging role. The quality of post-ICU care and the degree of caregiver strain may both be affected by poor preparation. The CaLM (caregiver large language model) has been proposed as a tool for teaching caregivers. The developers of this model aimed to provide caregivers with at least some of the knowledge they require to undertake this challenging role. They showed that by incorporating retrieval-augmented generation (a method used for improving model performance through connection with external knowledge bases), a valuable support tool tailored to specific caregiver scenarios could be created [43].

Patients recovering from ICU admission often require a lengthy and multidisciplinary rehabilitation process. A study that evaluated individualized exercise recommendations generated by an AI chatbot found them to be 41.2% comprehensive and 90.7% accurate. The chatbot could not provide complete and precise recommendations. Still, chatbots are early precursors of the LLMs existing at the time of this writing, and this study represents the potential for supporting rehabilitation efforts through such tools [44].

Figure 3 shows a timeline of patient management in the ICU with the potential application of LLMs at each treatment point.

While the examples presented illustrate potential directions for integrating LLMs into critical care daily practice, these remain exploratory. Figure4 offers one possible roadmap from model development to clinical integration.

Since the introduction of ChatGPT by OpenAI in November 2022, the public adoption of virtual assistants powered by large language models (LLMs) has grown rapidly. The interest in their application in healthcare, including critical care settings, highlights their potential, as shown above. This article summarizes a rapidly evolving technology whose clinical impact remains hypothetical. LLMs are still at the stage of isolated experiments in exploratory studies (for assessment of the studies presented above, refer to Table2) with limited incorporation of real-world patient care data; a recent systematic review by Bedi et al. found that only 5% of studies use such data [45]. There is no robust clinical validation, and the widespread use of these tools has also brought attention to their limitations and associated challenges [46,47].

A critical consideration in incorporating LLMs into clinical practice is their susceptibility to various biases, among them sycophancy bias, which may lead to outputs reinforcing clinicians’ preexisting beliefs, potentially increasing errors [48]. Our knowledge of such biases highlights the need to recognize and address how they may influence outputs and the importance of ongoing vigilance when integrating LLM-generated recommendations into clinical decision-making.

Another broader concern regarding the use of AI in general (not limited to LLMs alone) is the phenomenon of overreliance. Clinicians may trust AI-generated diagnoses even when the model produces inaccurate results (Supplementary A) [49]. One study investigated whether providing explanations alongside model-generated diagnoses could help clinicians discern and disregard incorrect outputs. Paradoxically, adding explanations did not improve decision-making accuracy, and reliance on the AI model persisted [50].

Finally, the hurdle of integrating LLMs into clinical workflows remains. These models often function as “black boxes,” with limited transparency regarding their internal decision-making processes - a challenge that extends even to their developers and is particularly pronounced among clinicians without even the basic appropriate training. This lack of clarity, coupled with insufficient familiarity among physicians regarding the known capabilities and limitations of these tools, impairs their ability to engage with LLMs in a safe, informed, and clinically meaningful manner. Physicians in family medicine, internal medicine, and emergency medicine exhibited better diagnostic performance on their own compared to when assisted by an LLM. This was assessed based on the accuracy of differential diagnoses, the relevance of supporting and opposing clinical factors, and the appropriateness of the diagnostic evaluation process. The authors interpreted this finding as highlighting “the need for technology and workforce development to realize the potential of physician-artificial intelligence collaboration in clinical practice“ [51].

AI tools are rapidly transitioning from simple tools to assistants and potentially even collaborative partners in medicine. They should, therefore, be upheld to similarly rigorous quality assurance standards. The Transparent Reporting of a Multivariable Model for Individual Prognosis Or Diagnosis for LLMs (TRIPOD-LLM) framework has recently been proposed for reporting clinical prediction models developed using large language models [52]. Several critical care leaders have also called for action on AI technologies, emphasizing the need to address technical, ethical, social, and practical issues posed by these tools. Their call highlighted the importance of ensuring that AIs, who may someday be viewed as equal partners to physicians, meet the same ethical and professional standards expected of humans. LLMs must uphold integrity, foster trust in clinical environments, and support their physician colleagues while maintaining the highest standards of care [53].

Authors and Affiliations

1. Ben-Gurion Faculty of Health Sciences, Beer-Sheva, Israel

   Daphna Idan

2. Hebrew University Faculty of Medicine and Regional Medical Director at Maccabi Healthcare and Chief Scientist, Medint Medical Intelligence, Hebrew University, Jerusalem, Israel

   Sharon Einav

Contributions

D.I. and S.E. were responsible for the conceptualization of the manuscript idea, drafted the main manuscript text, and prepared Figs. 1, 2 and 3. All authors reviewed and approved the final manuscript.

Corresponding author

Correspondence to Daphna Idan.

Competing interests

Daphna Idan is a medical student and data analyst and has no relevant conflict of interests to disclose. Sharon Einav is a Cochrane Editor and is involved in developing an LLM for use by clinicians.

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions