MIFlu: Large Language Model-Based Multimodal Influenza Forecasting Scheme.

Unleashing the potential of large language models (LLMs) in healthcare is a hot topic of research amongst computational scientists and bioinformatics researchers worldwide. Large language models have the potential to assist in various aspects of medicine given their capability to process complex concepts. The constant development of new treatment regimens is daunting for healthcare professionals and the introduction of these models allows for professionals to be updated with the latest advancements which improves patient care. CancerGPT [1] , an LLM, employs zero-shot learning to extract knowledge from medical literature and then use it to infer biological tasks. With approximately 124 million parameters, it matches the scale of the larger, fine-tuned GPT-3 model with around 175 million parameters. The observations from the study demonstrated that CancerGPT provided accurate responses for seven different biological tasks. However, the study also highlights the vulnerability of LLMs to hallucination. To address hallucination, which is incorrect and nonsensical content of large language models , our research seeks to address this limitation of LLM by seeking answers to the following questions. Can these models be trusted to responsibly recommend personalized treatment plans for cancer patients thereby improving their health outcomes? Can they serve as reliable clinical predictor tools for tailoring treatment strategies? To answer these questions, we curate a dataset consisting of treatment outcomes for monoclonal antibody treatment regimens in head and neck squamous cell carcinoma (HNSCC), which is the seventh most common cancer worldwide. The treatment regimens [2] include Cetuximab, Pembrolizumab, Durvalumab, and Nivolumab prescribed from the National Cancer Control Programme (NCCP) and The Irish Society of Medical Oncology (ISMO) which has a total of thirteen different drug regimens for HNSCC. The treatment outcomes are based on RECIST [3] score for each of the four drug regimens. Ten clinical trials for each of the individual treatment regimens were incorporated in our study. We augment prompt engineering process by providing contextual information such as drug regimens for targeted therapy, regimen doses, targeted oncogenes, and clinical outcomes from the clinical trials used in our study. We use one shot learning during prompt engineering phase. We test the LLM responses by providing contextual and non-contextual information. The LLM responses are evaluated using true positives and true negative scores on ten patient profiles of similar and different cancer type on the same drug regimen. Preliminary results indicate that contextual information significantly reduces LLM’s from hallucinating. To make the models reliable and trustworthy our study encompasses age, gender and genomic profile of patients. By overcoming hallucination and incorporating multimodal features, we demonstrate that LLMs holds the potential to recommend personalised treatments for HNSCC in Irish patients.

Time series (TS) data are ubiquitous across various application areas, rendering time series forecasting (TSF) a fundamental task. With the astounding advances in large language models (LLMs), a variety of methods have been developed to adapt LLMs for time series forecasting. Despite unlocking the potential of LLMs in comprehending TS data, existing methods are inherently constrained by their shallow integration of TS information, wherein LLMs typically access TS representations at shallow layers, primarily at the input layer. This causes the influence of TS representations to progressively fade in deeper layers and eventually leads to ineffective adaptation between textual embeddings and TS representations. In this paper, we propose the Multi-layer Steerable Embedding Fusion (MSEF), a novel framework that enables LLMs to directly access time series patterns at all depths, thereby mitigating the progressive loss of TS information in deeper layers. Specifically, MSEF leverages off-the-shelf time series foundation models to extract semantically rich embeddings, which are fused with intermediate text representations across LLM layers via layer-specific steering vectors. These steering vectors are designed to continuously optimize the alignment between time series and textual modalities and facilitate a layer-specific adaptation mechanism that ensures efficient few-shot learning capabilities. Experimental results on seven benchmarks demonstrate significant performance improvements by MSEF compared with baselines, with an average reduction of 31.8% in terms of MSE. The code is available at https://github.com/One1sAll/MSEF.

ObjectiveThrough this scoping review, we aim to explore and synthesize existing knowledge and evidence on the learning approaches for incorporating LLMs into healthcare education and public health research and learning spaces. Specifically, we will attempt to investigate methods for auditing prompts for accuracy, fairness, and effectiveness; tailoring prompts to improve task-specific accuracy and utility; and exploring how end-user feedback is used to refine and optimize LLM prompts over time. This review will provide a comprehensive understanding of how LLMs are being tailored and improved in these fields, contributing to the development of evidence-based strategies for their implementation. It will also identify areas for future research and innovation.IntroductionThe increasing integration of large language models (LLMs) into healthcare education and public health research and learning spaces, highlights their potential to revolutionize service delivery, decision-making, and ultimately patient care and outcomes. Despite these advancements, understanding how LLMs can be effectively tailored, audited, and refined for learning remains a critical area of inquiry. Key issues include, the accuracy of generated information, and their relevance to the medical and public health fields.Inclusion criteriaOur focus will be on studies addressing LLM applications in healthcare education and public health research and learning spaces, prompt engineering techniques, prompt auditing methods, and processes geared towards integrating user feedback. Articles that do not focus on healthcare or public health contexts and lack relevance to LLM learning approaches will be excluded.MethodsThe review is guided by the JBI methodology for scoping reviews complemented by updates from Levac et al. Databases including PubMed, Scopus, IEEE Xplore, and Web of Science will be searched for peer-reviewed articles, conference proceedings, and grey literature published in English and French from 2015 to 2025. Data extraction will include information on study characteristics, LLM models, prompt engineering strategies, auditing methodologies, and user feedback mechanisms. We will synthesize to identify trends, gaps, and best practices in leveraging LLMs to generate baseline data for auditing prompts that optimize AI learning and education needs in the healthcare and public health sector.

BackgroundRecognizing patient symptoms is fundamental to medicine, research, and public health. However, symptoms are often underreported in coded formats even though they are routinely documented in physician notes. Large language models (LLMs), noted for their generalizability, could help bridge this gap by mimicking the role of human expert chart reviewers for symptom identification.ObjectiveThe primary objective of this multisite study was to measure the accurate identification of infectious respiratory disease symptoms using LLMs instructed to follow chart review guidelines. The secondary objective was to evaluate LLM generalizability in multisite settings without the need for site-specific training, fine-tuning, or customization.MethodsFour LLMs were evaluated: GPT-4, GPT-3.5, Llama2 70B, and Mixtral 8×7B. LLM prompts were instructed to take on the role of chart reviewers and follow symptom annotation guidelines when assessing physician notes. Ground truth labels for each note were annotated by subject matter experts. Optimal LLM prompting strategies were selected using a development corpus of 103 notes from the emergency department at Boston Children’s Hospital. The performance of each LLM was measured using a test corpus with 202 notes from Boston Children’s Hospital. The performance of an International Classification of Diseases, Tenth Revision (ICD-10)–based method was also measured as a baseline. Generalizability of the most performant LLM was then measured in a validation corpus of 308 notes from 21 emergency departments in the Indiana Health Information Exchange.ResultsSymptom identification accuracy was superior for every LLM tested for each infectious disease symptom compared to an ICD-10–based method (F1-score=45.1%). GPT-4 was the highest scoring (F1-score=91.4%; P<.001) and was significantly better than the ICD-10–based method, followed by GPT-3.5 (F1-score=90.0%; P<.001), Llama2 (F1-score=81.7%; P<.001), and Mixtral (F1-score=83.5%; P<.001). For the validation corpus, performance of the ICD-10–based method decreased (F1-score=26.9%), while GPT-4 increased (F1-score=94.0%), demonstrating better generalizability using GPT-4 (P<.001).ConclusionsLLMs significantly outperformed an ICD-10–based method for respiratory symptom identification in emergency department electronic health records. GPT-4 demonstrated the highest accuracy and generalizability, suggesting that LLMs may augment or replace traditional approaches. LLMs can be instructed to mimic human chart reviewers with high accuracy. Future work should assess broader symptom types and health care settings.

Accurate air quality prediction is crucial for environmental monitoring and public health. This study explores a novel approach using machine learning algorithms and large language models to predict the Air Quality Index (AQI). While traditional AQI prediction relies on complex models and extensive data on pollutants and meteorological factors, this research investigates the use of readily available fuel consumption data as an alternative predictor, given its close link to emissions. The study employs supervised machine learning algorithms, including Random Forest and Gradient Boosting, utilizing fuel consumption data as input features to build predictive models. Additionally, a state-of-the-art large language model, GPT-3.5-turbo-instruct, is fine-tuned on historical AQI data and evaluated for its predictive capabilities. The performance of both machine learning models and the language model is compared using various metrics, and the results demonstrate that both approaches achieve high AQI prediction accuracy, outperforming traditional methods based on pollutant concentration data. Notably, the fine-tuned language model exhibits superior performance, potentially due to its ability to capture complex dependencies and contextual information from the training data. This work highlights the potential of leveraging readily available fuel consumption data and advanced language models for accurate and cost-effective AQI prediction. The findings have significant implications for developing scalable air quality monitoring systems, enabling timely interventions and informed decision-making to mitigate the adverse effects of air pollution.

This paper investigates the impact of knowledge extraction methods on the effectiveness of RAG (Retrieval-Augmented Generation) systems that utilize knowledge graphs. It highlights that the quality of the knowledge graph, which is formed using various knowledge extraction methods, is crucial for overcoming limitations of large language models (LLMs), such as “hallucinations”. The paper analyzes the architectures of LightRAG and GraphRAG, emphasizing that the selection of an optimal knowledge extraction strategy depends on specific tasks and the subject area. LLMs have advanced significantly, but they have limitations, including generating factually incorrect information (“hallucinations”) and possessing “outdated knowledge”. RAG systems were proposed to address these issues by combining LLMs with external knowledge bases. This approach reduces hallucinations, ensures factual accuracy, solves the problem of outdated knowledge, and increases transparency. Knowledge graphs are powerful tools for structuring information, consisting of entities (nodes) and relations (edges). They enhance RAG systems by enabling more precise and contextually grounded retrieval compared to keyword-based searches. The quality of a knowledge graph depends on the knowledge extraction methods used, which include named entity recognition (NER), relation extraction (RE), entity linking (EL), and event extraction. Different methods, such as rule-based, classical machine learning, and deep learning approaches, have varying trade-offs in terms of accuracy, completeness, and scalability. Entity linking and knowledge graph completion are also crucial for accuracy and richness. LightRAG and GraphRAG are two main graph-based RAG systems. LightRAG uses the knowledge graph as a quick reference, requiring high precision in knowledge extraction to avoid noise degradation. GraphRAG uses the knowledge graph as a domain model, where completeness of extracted knowledge is more critical, though systematic errors are still harmful. Both systems rely on LLMs for knowledge extraction, which makes them dependent on the LLM’s quality and the size of document fragments processed. The theoretical analysis confirms that the effectiveness of RAG systems is critically dependent on knowledge extraction methods. The quality, completeness, and accuracy of the knowledge graph directly influence the RAG system’s ability to provide relevant, accurate, and truthful answers. Different RAG system architectures like LightRAG and GraphRAG have distinct requirements for knowledge graph characteristics, prioritizing either accuracy or completeness in knowledge extraction.

The rapid advancement of artificial intelligence (AI) has created new opportunities in educational research, particularly in the efficient analysis of complex social interactions within classrooms. One promising area involves the classification of teachers' motivational messages. Traditionally, such assessments have relied on self-reports and observer evaluations, which require a lot of staff and time resources. Recently, large language models (LLMs) have been employed to classify teachers' motivational messages, offering novel, less labour-intensive approaches for classification. Building on these recent developments, this work presents a comprehensive literature overview exploring the applications, potential, and limitations of using LLMs to classify teachers' motivational messages. The present comprehensive literature overview indicates that the use of LLMs for classifying teachers' motivational messages is a promising yet still emerging field of research. Recent studies have applied LLMs in innovative ways, drawing on established motivational theories and employing novel classification techniques, such as zero-shot and few-shot prompting or fine-tuning, to classify motivational messages. Open questions remain, particularly concerning the structure, quantity, and quality of annotated material. Whereas recent studies have demonstrated the potential of LLMs to offer scalable and time-efficient alternatives for classifying motivational messages in the classroom, several challenges persist. These include concerns related to the quality and quantity of training data, model generalisability, the ability to capture the complexity of classroom interactions, and biases involved in integrating LLMs as a classification method. This comprehensive literature overview provides practical recommendations for the responsible use of LLMs in educational research and school practice.

Text clustering is an important method for organising the increasing volume of digital content, aiding in the structuring and discovery of hidden patterns in uncategorised data. The effectiveness of text clustering largely depends on the selection of textual embeddings and clustering algorithms. This study argues that recent advancements in large language models (LLMs) have the potential to enhance this task. The research investigates how different textual embeddings, particularly those utilised in LLMs, and various clustering algorithms influence the clustering of text datasets. A series of experiments were conducted to evaluate the impact of embeddings on clustering results, the role of dimensionality reduction through summarisation, and the adjustment of model size. The findings indicate that LLM embeddings are superior at capturing subtleties in structured language. OpenAI’s GPT-3.5 Turbo model yields better results in three out of five clustering metrics across most tested datasets. Most LLM embeddings show improvements in cluster purity and provide a more informative silhouette score, reflecting a refined structural understanding of text data compared to traditional methods. Among the more lightweight models, BERT demonstrates leading performance. Additionally, it was observed that increasing model dimensionality and employing summarisation techniques do not consistently enhance clustering efficiency, suggesting that these strategies require careful consideration for practical application. These results highlight a complex balance between the need for refined text representation and computational feasibility in text clustering applications. This study extends traditional text clustering frameworks by integrating embeddings from LLMs, offering improved methodologies and suggesting new avenues for future research in various types of textual analysis.

Extracting information from unstructured and semi-structured geoscience literature is a crucial step in conducting geological research. The traditional machine learning extraction paradigm requires a substantial amount of high-quality manually annotated data for model training, which is time-consuming, labor-intensive, and not easily transferable to new fields. Recently, large language models (LLMs) (e.g., ChatGPT, GPT-4, and LLaMA), have shown great performance in various natural language processing (NLP) tasks, such as question answering, machine translation, and text generation. A substantial body of work has demonstrated that LLMs possess strong in-context learning (ICL) and even zero-shot learning capabilities to solve downstream tasks without specifically designed supervised fine-tuning. In this paper, we propose utilizing LLMs for geoscience literature information extraction. Specifically, we design a hierarchical PDF parsing pipeline and an automated knowledge extraction process, which can significantly reduce the need for manual data annotation, assisting geoscientists in literature data mining. For the hierarchical PDF parsing pipeline, firstly, a document layout detection model fine-tuned on geoscience literature is employed for layout detection, obtaining layout detection information for the document. Secondly, based on the document layout information, an optical character content parsing model is used for content parsing, obtaining the text structure and plain text corresponding to the content. Finally, the text structure and plain text are combined and reconstructed to ultimately obtain the parsed structured data. For the automated knowledge extraction process, firstly, the parsed long text is segmented into paragraphs to adapt to the input length limit of LLMs. Subsequently, a few-shot prompting method is employed for structured knowledge extraction, encompassing two tasks: attribute value extraction and triplet extraction. In attribute value extraction, prompts are generated automatically by the LLMs based on the subdomain and attribute names, facilitating the location and extraction of values related to subdomain attribute names in the text. For triplet extraction, the LLMs employ a procedural approach to entity extraction, entity type extraction, and relation extraction, following the knowledge graph structure pattern. Finally, the extracted structured knowledge is stored in the form of knowledge graphs, facilitating further analysis and integration of various types of knowledge from the literature. Our proposed approach turns out to be simple, flexible, and highly effective in geoscience literature information extraction. Demonstrations of information extraction in subdomains such as radiolarian fossils and fluvial facies have yielded satisfactory results. The extraction efficiency has significantly improved, and feedback from domain experts indicates a relatively high level of accuracy in the extraction process. The extracted results can be used to construct a foundational knowledge graph for geoscience literature, supporting the comprehensive construction and efficient application of a geoscience knowledge graph.

ObjectiveTo investigate the level of understanding and trust of medical students towards ChatGPT-like large language models, as well as their utilization and attitudes towards these models.MethodsData collection was concentrated from December 2023 to mid-January 2024, utilizing a self-designed questionnaire to assess the use of large language models among undergraduate medical students at Anhui Medical University. The normality of the data was confirmed with Shapiro-Wilk tests. We used Chi-square tests for comparisons of categorical variables, Mann-Whitney U tests for comparisons of ordinal variables and non-normal continuous variables between two groups, Kruskall-Wallis H tests for comparisons of ordinal variables between multiple groups, and Bonferroni tests for post hoc comparisons.ResultsA total of 1774 questionnaires were distributed and 1718 valid questionnaires were collected, with an effective rate of 96.84%. Among these students, 34.5% had heard and used large language models. There were statistically significant differences in the understanding of large language models between genders (p < 0.001), grade levels (junior-level students and senior-level students) (p = 0.03), and major (p < 0.001). Male, junior-level students, and public health management had a higher level of understanding of these models. Genders and majors had statistically significant effects on the degree of trust in large language models (p = 0.004; p = 0.02). Male and nursing students exhibited a higher degree of trust in large language models. As for usage, Male and junior-level students showed a significantly higher proportion of using these models for assisted learning (p < 0.001). Neutral sentiments were held by over two-thirds of the students (66.7%) regarding large language models, with only 51(3.0%) expressing pessimism. There were significant gender-based disparities in attitudes towards large language models, and male exhibited a more optimistic attitude towards these models (p < 0.001). Notably, among students with different levels of knowledge and trust in large language models, statistically significant differences were observed in their perceptions of the shortcomings and benefits of these models.ConclusionOur study identified gender, grade levels, and major as influential factors in students’ understanding and utilization of large language models. This also suggested the feasibility of integrating large language models with traditional medical education to further enhance teaching effectiveness in the future.

The “Impression” section of a radiology report is a critical basis for communication between radiologists and other physicians. Typically written by radiologists, this part is derived from the “Findings” section, which can be laborious and error-prone. Although deep-learning based models, such as BERT, have achieved promising results in Automatic Impression Generation (AIG), such models often require substantial amounts of medical data and have poor generalization performance. Recently, Large Language Models (LLMs) like ChatGPT have shown strong generalization capabilities and performance, but their performance in specific domains, such as radiology, remains under-investigated and potentially limited. To address this limitation, we propose ImpressionGPT, leveraging the contextual learning capabilities of LLMs through our dynamic prompt and iterative optimization algorithm to accomplish the AIG task. ImpressionGPT initially employs a small amount of domain-specific data to create a dynamic prompt, extracting contextual semantic information closely related to the test data. Subsequently, the iterative optimization algorithm automatically evaluates the output of LLMs and provides optimization suggestions, continuously refining the output results. The proposed ImpressionGPT model achieves superior performance of AIG task on both MIMIC-CXR and OpenI datasets without requiring additional training data or fine-tuning the LLMs. This work presents a paradigm for localizing LLMs that can be applied in a wide range of similar application scenarios, bridging the gap between general-purpose LLMs and the specific language processing needs of various domains.

<sec> <title>BACKGROUND</title> Substance use disorder (SUD) affects more than 48 million people in the US, but access to care is often constrained by factors such as stigma, financial barriers, and limited service availability. While digital health interventions have begun to mitigate these challenges, large language models (LLMs) such as GPT introduce new possibilities for scalable SUD recovery support. However, empirical evidence regarding LLM use in SUD contexts re-mains fragmented and unevenly evaluated. A scoping review is needed to synthesize cur-rent approaches, surface methodological and safety gaps, and inform future research. </sec> <sec> <title>OBJECTIVE</title> This scoping review synthesizes current evidence on the use of LLMs in SUD and addiction contexts. It examines how LLM-based tools are being applied across the full spectrum of care, identifies common tasks and data sources, evaluates risks and safety concerns, and maps early evidence of effectiveness and gaps that inform future research. </sec> <sec> <title>METHODS</title> We conducted a scoping review in accordance with PRISMA-ScR guidelines. We searched medical and technical databases for studies that reported empirical findings on LLMs in drug-related SUD contexts. Of 675 records identified, 492 were screened at the title and abstract level, 60 underwent full-text review, and 12 studies met inclusion criteria. </sec> <sec> <title>RESULTS</title> All included studies were published in 2024 or 2025, indicating a very recent and rapidly evolving evidence base. Most studies used public social media or online forums (8 of 12), while 3 studies used clinical data such as EHR, and 1 study used multimodal data that combined sensors, images, and chat logs. GPT-family models were the most frequently used (10 of 12), often evaluated alongside Llama or Mistral models. Across studies, LLM applications clustered into four task categories: text classification and extraction, response generation, conversational decision support, and predictive modeling using LLM-derived features. Classification and extraction tasks were most common and often showed im-proved performance compared with traditional machine learning or rule-based baselines. While general-purpose LLMs were prone to generating stigmatizing, biased, or clinically inappropriate content in SUD-related scenarios, it was also reported that targeted prompt engineering substantially reduced stigmatizing language and improved perceived ade-quacy and appropriateness of responses. However, safety evaluations remained limited in scope and depth. </sec> <sec> <title>CONCLUSIONS</title> LLM-based tools show emerging potential to support SUD-related research and care, particularly scalable classification and information extraction, conversational screening, and feature derivation from unstructured text or speech. Yet current work is predominantly preclinical, relying on public datasets and simulated prompts, with limited evidence on safety, clinical effectiveness, or real-world implementation in recovery settings. Responsible use in SUD contexts will require a safety-first development approach that integrates human-in-the-loop validation, systematic prompt engineering, and design constraints such as curated response libraries. It will also require efforts to strengthen LLM literacy among clinicians, peer specialists, and people in recovery. </sec>

Cross-project defect prediction (CPDP) aims to predict software defects in a target project domain by leveraging information from different source project domains, allowing testers to identify defective modules quickly. However, CPDP models often underperform due to different data distributions between source and target domains, class imbalances, and the presence of noisy and irrelevant instances in both source and target projects. Additionally, standard features often fail to capture sufficient semantic and contextual information from the source project, leading to poor prediction performance in the target project. To address these challenges, this research proposes Smote Correlation and Attention Gated recurrent unit based Long Short-Term Memory optimization (SCAG-LSTM), which first employs a novel hybrid technique that extends the synthetic minority over-sampling technique (SMOTE) with edited nearest neighbors (ENN) to rebalance class distributions and mitigate the issues caused by noisy and irrelevant instances in both source and target domains. Furthermore, correlation-based feature selection (CFS) with best-first search (BFS) is utilized to identify and select the most important features, aiming to reduce the differences in data distribution among projects. Additionally, SCAG-LSTM integrates bidirectional gated recurrent unit (Bi-GRU) and bidirectional long short-term memory (Bi-LSTM) networks to enhance the effectiveness of the long short-term memory (LSTM) model. These components efficiently capture semantic and contextual information as well as dependencies within the data, leading to more accurate predictions. Moreover, an attention mechanism is incorporated into the model to focus on key features, further improving prediction performance. Experiments are conducted on apache_lucene, equinox, eclipse_jdt_core, eclipse_pde_ui, and mylyn (AEEEM) and predictor models in software engineering (PROMISE) datasets and compared with active learning-based method (ALTRA), multi-source-based cross-project defect prediction method (MSCPDP), the two-phase feature importance amplification method (TFIA) on AEEEM and the two-phase transfer learning method (TPTL), domain adaptive kernel twin support vector machines method (DA-KTSVMO), and generative adversarial long-short term memory neural networks method (GB-CPDP) on PROMISE datasets. The results demonstrate that the proposed SCAG-LSTM model enhances the baseline models by 33.03%, 29.15% and 1.48% in terms of F1-measure and by 16.32%, 34.41% and 3.59% in terms of Area Under the Curve (AUC) on the AEEEM dataset, while on the PROMISE dataset it enhances the baseline models’ F1-measure by 42.60%, 32.00% and 25.10% and AUC by 34.90%, 27.80% and 12.96%. These findings suggest that the proposed model exhibits strong predictive performance.

Accurate color alignment in text-to-image (T2I) generation is critical for applications such as fashion, product visualization, and interior design, yet current diffusion models struggle with nuanced and compound color terms (e.g., Tiffany blue, baby pink), often producing images that are misaligned with human intent. Existing approaches rely on cross-attention manipulation, reference images, or fine-tuning but fail to systematically resolve ambiguous color descriptions. To precisely render colors under prompt ambiguity, we propose a training-free framework that enhances color fidelity by leveraging a large language model (LLM) to disambiguate color-related prompts and guiding color blending operations directly in the text embedding space. Our method first employs a large language model (LLM) to resolve ambiguous color terms in the text prompt, and then refines the text embeddings based on the spatial relationships of the resulting color terms in the CIELab color space. Unlike prior methods, our approach improves color accuracy without requiring additional training or external reference images. Experimental results demonstrate that our framework improves color alignment without compromising image quality, bridging the gap between text semantics and visual generation. All supplementary materials are available at https://Sung-Lin.github.io/TintBench/.

Electronic health records (EHRs) with clinical decision support tools are now ubiquitous in healthcare organizations. Clinical foundation models (CFMs) pretrained on large-scale, heterogeneous structured EHR data have emerged as a powerful approach to improve predictive performance and generalizability. Meanwhile, large language models (LLMs) pretrained on broad data sources are being applied to an expanding range of healthcare tasks. However, it remains unclear whether generalist LLMs can match specialized CFMs for disease risk prediction using structured clinical data. We compared CFMs (Med-BERT, CLMBR) against fine-tuned generalist LLMs (Mistral, LLaMA-2/3/3.1), a clinical LLM (Me-LLaMA), and LLM-generated embeddings paired with simple classifiers (using DeepSeek, Qwen3, and GPT-OSS) on two disease risk prediction tasks: heart failure risk among diabetic patients (DHF) and pancreatic cancer diagnosis (PaCa). Evaluations spanned multi-site EHR data, claims data, and an open-source single-institution benchmark (EHRSHOT). Performance was assessed using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). On larger EHR and claims cohorts (>30,000 patients), fine-tuned CFMs outperformed fine-tuned LLMs by a small but statistically significant margin (<1% AUROC). The clinical LLM performed comparably to generalist LLMs despite being smaller. On the open-source PaCa cohort (3,810 patients, 199 cases), LLMs achieved slightly higher AUROCs that were not statistically significant (LLaMA-3.1-70B 86.1% vs. Med-BERT 85.3%, p=0.27), but CFMs achieved significantly higher AUPRC (Med-BERT 55.9% vs. LLaMA-3.1-70B 41.1%, p=0.001). Notably, LLM-generated trajectory embeddings paired with logistic regression or a simple MLP, without any LLM fine-tuning, achieved the best overall performance, with AUROC exceeding 90% (Qwen3) and AUPRC reaching 66% (GPT-OSS 20B). LLM-generated embeddings with lightweight classifiers outperformed both fine-tuned CFMs and fine-tuned LLMs on AUROC and AUPRC. While these results demonstrate the potential of generalist models to match or surpass specialized CFMs, their substantially greater computational cost and variable AUPRC performance in the fine-tuning setting warrant caution. We provide a reproducible evaluation framework and codebase to support continued benchmarking.