Else Kröner Fresenius Center for Digital Health

Large language models (LLMs) are artificial intelligence (AI) tools specifically trained to process and generate text. LLMs attracted substantial public attention after OpenAI's ChatGPT was made publicly available in November 2022. LLMs can often answer questions, summarize, paraphrase and translate text on a level that is nearly indistinguishable from human capabilities. The possibility to actively interact with models like ChatGPT makes LLMs attractive tools in various fields, including medicine. While these models have the potential to democratize medical knowledge and facilitate access to healthcare, they could equally distribute misinformation and exacerbate scientific misconduct due to a lack of accountability and transparency. In this article, we provide a systematic and comprehensive overview of the potentials and limitations of LLMs in clinical practice, medical research and medical education.

Recent advances in computer vision have shown promising results in image generation. Diffusion probabilistic models have generated realistic images from textual input, as demonstrated by DALL-E 2, Imagen, and Stable Diffusion. However, their use in medicine, where imaging data typically comprises three-dimensional volumes, has not been systematically evaluated. Synthetic images may play a crucial role in privacy-preserving artificial intelligence and can also be used to augment small datasets. We show that diffusion probabilistic models can synthesize high-quality medical data for magnetic resonance imaging (MRI) and computed tomography (CT). For quantitative evaluation, two radiologists rated the quality of the synthesized images regarding "realistic image appearance", "anatomical correctness", and "consistency between slices". Furthermore, we demonstrate that synthetic images can be used in self-supervised pre-training and improve the performance of breast segmentation models when data is scarce (Dice scores, 0.91 [without synthetic data], 0.95 [with synthetic data]).

BACKGROUND: The introduction of large language models (LLMs) into clinical practice promises to improve patient education and empowerment, thereby personalizing medical care and broadening access to medical knowledge. Despite the popularity of LLMs, there is a significant gap in systematized information on their use in patient care. Therefore, this systematic review aims to synthesize current applications and limitations of LLMs in patient care. METHODS: We systematically searched 5 databases for qualitative, quantitative, and mixed methods articles on LLMs in patient care published between 2022 and 2023. From 4349 initial records, 89 studies across 29 medical specialties were included. Quality assessment was performed using the Mixed Methods Appraisal Tool 2018. A data-driven convergent synthesis approach was applied for thematic syntheses of LLM applications and limitations using free line-by-line coding in Dedoose. RESULTS: We show that most studies investigate Generative Pre-trained Transformers (GPT)-3.5 (53.2%, n = 66 of 124 different LLMs examined) and GPT-4 (26.6%, n = 33/124) in answering medical questions, followed by patient information generation, including medical text summarization or translation, and clinical documentation. Our analysis delineates two primary domains of LLM limitations: design and output. Design limitations include 6 second-order and 12 third-order codes, such as lack of medical domain optimization, data transparency, and accessibility issues, while output limitations include 9 second-order and 32 third-order codes, for example, non-reproducibility, non-comprehensiveness, incorrectness, unsafety, and bias. CONCLUSIONS: This review systematically maps LLM applications and limitations in patient care, providing a foundational framework and taxonomy for their implementation and evaluation in healthcare settings.

Deep learning (DL) can accelerate the prediction of prognostic biomarkers from routine pathology slides in colorectal cancer (CRC). However, current approaches rely on convolutional neural networks (CNNs) and have mostly been validated on small patient cohorts. Here, we develop a new transformer-based pipeline for end-to-end biomarker prediction from pathology slides by combining a pre-trained transformer encoder with a transformer network for patch aggregation. Our transformer-based approach substantially improves the performance, generalizability, data efficiency, and interpretability as compared with current state-of-the-art algorithms. After training and evaluating on a large multicenter cohort of over 13,000 patients from 16 colorectal cancer cohorts, we achieve a sensitivity of 0.99 with a negative predictive value of over 0.99 for prediction of microsatellite instability (MSI) on surgical resection specimens. We demonstrate that resection specimen-only training reaches clinical-grade performance on endoscopic biopsy tissue, solving a long-standing diagnostic problem.

Although generative adversarial networks (GANs) can produce large datasets, their limited diversity and fidelity have been recently addressed by denoising diffusion probabilistic models, which have demonstrated superiority in natural image synthesis. In this study, we introduce Medfusion, a conditional latent DDPM designed for medical image generation, and evaluate its performance against GANs, which currently represent the state-of-the-art. Medfusion was trained and compared with StyleGAN-3 using fundoscopy images from the AIROGS dataset, radiographs from the CheXpert dataset, and histopathology images from the CRCDX dataset. Based on previous studies, Progressively Growing GAN (ProGAN) and Conditional GAN (cGAN) were used as additional baselines on the CheXpert and CRCDX datasets, respectively. Medfusion exceeded GANs in terms of diversity (recall), achieving better scores of 0.40 compared to 0.19 in the AIROGS dataset, 0.41 compared to 0.02 (cGAN) and 0.24 (StyleGAN-3) in the CRMDX dataset, and 0.32 compared to 0.17 (ProGAN) and 0.08 (StyleGAN-3) in the CheXpert dataset. Furthermore, Medfusion exhibited equal or higher fidelity (precision) across all three datasets. Our study shows that Medfusion constitutes a promising alternative to GAN-based models for generating high-quality medical images, leading to improved diversity and less artifacts in the generated images.

The initial vision of cellular communications was to deliver ubiquitous voice communications to anyone anywhere. In a simplified view, 1G delivered voice services for business customers, and only 2G for consumers. Next, this also initiated the appetite for cellular data, for which 3G was designed. However, Blackberry delivered business smartphones, and 4G made smartphones a consumer device. The promise of 5G is to start the Tactile Internet, to control real and virtual objects in real-time via cellular. However, the hype around 5G is, again, focusing on business customers, in particular in the context of campus networks. Consequently, 6G must provide an infrastructure to enable remote-controlled mobile robotic solutions for everyone—the Personal Tactile Internet. Which role can information and communication theory play in this context, and what are the big challenges ahead?

Medical image classification requires labeled, task-specific datasets which are used to train deep learning networks de novo, or to fine-tune foundation models. However, this process is computationally and technically demanding. In language processing, in-context learning provides an alternative, where models learn from within prompts, bypassing the need for parameter updates. Yet, in-context learning remains underexplored in medical image analysis. Here, we systematically evaluate the model Generative Pretrained Transformer 4 with Vision capabilities (GPT-4V) on cancer image processing with in-context learning on three cancer histopathology tasks of high importance: Classification of tissue subtypes in colorectal cancer, colon polyp subtyping and breast tumor detection in lymph node sections. Our results show that in-context learning is sufficient to match or even outperform specialized neural networks trained for particular tasks, while only requiring a minimal number of samples. In summary, this study demonstrates that large vision language models trained on non-domain specific data can be applied out-of-the box to solve medical image-processing tasks in histopathology. This democratizes access of generalist AI models to medical experts without technical background especially for areas where annotated data is scarce.

Among the rapid integration of artificial intelligence in clinical settings, large language models (LLMs), such as Generative Pre-trained Transformer-4, have emerged as multifaceted tools that have potential for health-care delivery, diagnosis, and patient care. However, deployment of LLMs raises substantial regulatory and safety concerns. Due to their high output variability, poor inherent explainability, and the risk of so-called AI hallucinations, LLM-based health-care applications that serve a medical purpose face regulatory challenges for approval as medical devices under US and EU laws, including the recently passed EU Artificial Intelligence Act. Despite unaddressed risks for patients, including misdiagnosis and unverified medical advice, such applications are available on the market. The regulatory ambiguity surrounding these tools creates an urgent need for frameworks that accommodate their unique capabilities and limitations. Alongside the development of these frameworks, existing regulations should be enforced. If regulators fear enforcing the regulations in a market dominated by supply or development by large technology companies, the consequences of layperson harm will force belated action, damaging the potentiality of LLM-based applications for layperson medical advice.

BACKGROUND: The field of Artificial Intelligence (AI) holds transformative potential in medicine. However, the lack of universal reporting guidelines poses challenges in ensuring the validity and reproducibility of published research studies in this field. METHODS: Based on a systematic review of academic publications and reporting standards demanded by both international consortia and regulatory stakeholders as well as leading journals in the fields of medicine and medical informatics, 26 reporting guidelines published between 2009 and 2023 were included in this analysis. Guidelines were stratified by breadth (general or specific to medical fields), underlying consensus quality, and target research phase (preclinical, translational, clinical) and subsequently analyzed regarding the overlap and variations in guideline items. RESULTS: AI reporting guidelines for medical research vary with respect to the quality of the underlying consensus process, breadth, and target research phase. Some guideline items such as reporting of study design and model performance recur across guidelines, whereas other items are specific to particular fields and research stages. CONCLUSIONS: Our analysis highlights the importance of reporting guidelines in clinical AI research and underscores the need for common standards that address the identified variations and gaps in current guidelines. Overall, this comprehensive overview could help researchers and public stakeholders reinforce quality standards for increased reliability, reproducibility, clinical validity, and public trust in AI research in healthcare. This could facilitate the safe, effective, and ethical translation of AI methods into clinical applications that will ultimately improve patient outcomes.

It is currently not well known how necroptosis and necroptosis responses manifest in vivo. Here, we uncovered a molecular switch facilitating reprogramming between two alternative modes of necroptosis signaling in hepatocytes, fundamentally affecting immune responses and hepatocarcinogenesis. Concomitant necrosome and NF-κB activation in hepatocytes, which physiologically express low concentrations of receptor-interacting kinase 3 (RIPK3), did not lead to immediate cell death but forced them into a prolonged "sublethal" state with leaky membranes, functioning as secretory cells that released specific chemokines including CCL20 and MCP-1. This triggered hepatic cell proliferation as well as activation of procarcinogenic monocyte-derived macrophage cell clusters, contributing to hepatocarcinogenesis. In contrast, necrosome activation in hepatocytes with inactive NF-κB-signaling caused an accelerated execution of necroptosis, limiting alarmin release, and thereby preventing inflammation and hepatocarcinogenesis. Consistently, intratumoral NF-κB-necroptosis signatures were associated with poor prognosis in human hepatocarcinogenesis. Therefore, pharmacological reprogramming between these distinct forms of necroptosis may represent a promising strategy against hepatocellular carcinoma.

Clinical decision-making in oncology is complex, requiring the integration of multimodal data and multidomain expertise. We developed and evaluated an autonomous clinical artificial intelligence (AI) agent leveraging GPT-4 with multimodal precision oncology tools to support personalized clinical decision-making. The system incorporates vision transformers for detecting microsatellite instability and KRAS and BRAF mutations from histopathology slides, MedSAM for radiological image segmentation and web-based search tools such as OncoKB, PubMed and Google. Evaluated on 20 realistic multimodal patient cases, the AI agent autonomously used appropriate tools with 87.5% accuracy, reached correct clinical conclusions in 91.0% of cases and accurately cited relevant oncology guidelines 75.5% of the time. Compared to GPT-4 alone, the integrated AI agent drastically improved decision-making accuracy from 30.3% to 87.2%. These findings demonstrate that integrating language models with precision oncology and search tools substantially enhances clinical accuracy, establishing a robust foundation for deploying AI-driven personalized oncology support systems.

Background Deep learning (DL) models can potentially improve prognostication of rectal cancer but have not been systematically assessed. Purpose To develop and validate an MRI DL model for predicting survival in patients with rectal cancer based on segmented tumor volumes from pretreatment T2-weighted MRI scans. Materials and Methods DL models were trained and validated on retrospectively collected MRI scans of patients with rectal cancer diagnosed between August 2003 and April 2021 at two centers. Patients were excluded from the study if there were concurrent malignant neoplasms, prior anticancer treatment, incomplete course of neoadjuvant therapy, or no radical surgery performed. The Harrell C-index was used to determine the best model, which was applied to internal and external test sets. Patients were stratified into high- and low-risk groups based on a fixed cutoff calculated in the training set. A multimodal model was also assessed, which used DL model–computed risk score and pretreatment carcinoembryonic antigen level as input. Results The training set included 507 patients (median age, 56 years [IQR, 46–64 years]; 355 men). In the validation set (n = 218; median age, 55 years [IQR, 47–63 years]; 144 men), the best algorithm reached a C-index of 0.82 for overall survival. The best model reached hazard ratios of 3.0 (95% CI: 1.0, 9.0) in the high-risk group in the internal test set (n = 112; median age, 60 years [IQR, 52–70 years]; 76 men) and 2.3 (95% CI: 1.0, 5.4) in the external test set (n = 58; median age, 57 years [IQR, 50–67 years]; 38 men). The multimodal model further improved the performance, with a C-index of 0.86 and 0.67 for the validation and external test set, respectively. Conclusion A DL model based on preoperative MRI was able to predict survival of patients with rectal cancer. The model could be used as a preoperative risk stratification tool. Published under a CC BY 4.0 license. Supplemental material is available for this article. See also the editorial by Langs in this issue.

The European Union (EU) has taken a critical step towards regulating artificial intelligence (AI) with the adoption of the AI Act by its 27 member states on March 13, 2024 1 . First proposed in April 2021 by the European Commission 2 , the AI Act emerged from the growing recognition of AI’s transformative potential and the need to address associated risks and ethical concerns, building on previous EU initiatives, including the 2018 Coordinated Plan on AI 3 and the 2020 White Paper on AI 4 . The Council issued its position in December 2022 5 , followed by the European Parliament’s adoption of its negotiating position in June 2023 6 . After revising the draft, the final negotiations between the Commission and Council resulted in a provisional agreement in December 2023, which was endorsed by Member States in February 2024 1 .

Most clinical information is encoded as free text, not accessible for quantitative analysis. This study presents an open-source pipeline using the local large language model (LLM) "Llama 2" to extract quantitative information from clinical text and evaluates its performance in identifying features of decompensated liver cirrhosis. The LLM identified five key clinical features in a zero- and one-shot manner from 500 patient medical histories in the MIMIC IV dataset. We compared LLMs of three sizes and various prompt engineering approaches, with predictions compared against ground truth from three blinded medical experts. Our pipeline achieved high accuracy, detecting liver cirrhosis with 100% sensitivity and 96% specificity. High sensitivities and specificities were also yielded for detecting ascites (95%, 95%), confusion (76%, 94%), abdominal pain (84%, 97%), and shortness of breath (87%, 97%) using the 70 billion parameter model, which outperformed smaller versions. Our study successfully demonstrates the capability of locally deployed LLMs to extract clinical information from free text with low hardware requirements.

Deep learning applied to whole-slide histopathology images (WSIs) has the potential to enhance precision oncology and alleviate the workload of experts. However, developing these models necessitates large amounts of data with ground truth labels, which can be both time-consuming and expensive to obtain. Pathology reports are typically unstructured or poorly structured texts, and efforts to implement structured reporting templates have been unsuccessful, as these efforts lead to perceived extra workload. In this study, we hypothesised that large language models (LLMs), such as the generative pre-trained transformer 4 (GPT-4), can extract structured data from unstructured plain language reports using a zero-shot approach without requiring any re-training. We tested this hypothesis by utilising GPT-4 to extract information from histopathological reports, focusing on two extensive sets of pathology reports for colorectal cancer and glioblastoma. We found a high concordance between LLM-generated structured data and human-generated structured data. Consequently, LLMs could potentially be employed routinely to extract ground truth data for machine learning from unstructured pathology reports in the future. © 2023 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Britain and Ireland.

Artificial Intelligence (AI) can support diagnostic workflows in oncology by aiding diagnosis and providing biomarkers directly from routine pathology slides. However, AI applications are vulnerable to adversarial attacks. Hence, it is essential to quantify and mitigate this risk before widespread clinical use. Here, we show that convolutional neural networks (CNNs) are highly susceptible to white- and black-box adversarial attacks in clinically relevant weakly-supervised classification tasks. Adversarially robust training and dual batch normalization (DBN) are possible mitigation strategies but require precise knowledge of the type of attack used in the inference. We demonstrate that vision transformers (ViTs) perform equally well compared to CNNs at baseline, but are orders of magnitude more robust to white- and black-box attacks. At a mechanistic level, we show that this is associated with a more robust latent representation of clinically relevant categories in ViTs compared to CNNs. Our results are in line with previous theoretical studies and provide empirical evidence that ViTs are robust learners in computational pathology. This implies that large-scale rollout of AI models in computational pathology should rely on ViTs rather than CNN-based classifiers to provide inherent protection against perturbation of the input data, especially adversarial attacks.

Structured reporting (SR) has long been a goal in radiology to standardize and improve the quality of radiology reports. Despite evidence that SR reduces errors, enhances comprehensiveness, and increases adherence to guidelines, its widespread adoption has been limited. Recently, large language models (LLMs) have emerged as a promising solution to automate and facilitate SR. Therefore, this narrative review aims to provide an overview of LLMs for SR in radiology and beyond. We found that the current literature on LLMs for SR is limited, comprising ten studies on the generative pre-trained transformer (GPT)-3.5 (n = 5) and/or GPT-4 (n = 8), while two studies additionally examined the performance of Perplexity and Bing Chat or IT5. All studies reported promising results and acknowledged the potential of LLMs for SR, with six out of ten studies demonstrating the feasibility of multilingual applications. Building upon these findings, we discuss limitations, regulatory challenges, and further applications of LLMs in radiology report processing, encompassing four main areas: documentation, translation and summarization, clinical evaluation, and data mining. In conclusion, this review underscores the transformative potential of LLMs to improve efficiency and accuracy in SR and radiology report processing. KEY POINTS: Question How can LLMs help make SR in radiology more ubiquitous? Findings Current literature leveraging LLMs for SR is sparse but shows promising results, including the feasibility of multilingual applications. Clinical relevance LLMs have the potential to transform radiology report processing and enable the widespread adoption of SR. However, their future role in clinical practice depends on overcoming current limitations and regulatory challenges, including opaque algorithms and training data.

Histopathology and genomic profiling are cornerstones of precision oncology and are routinely obtained for patients with cancer. Traditionally, histopathology slides are manually reviewed by highly trained pathologists. Genomic data, on the other hand, is evaluated by engineered computational pipelines. In both applications, the advent of modern artificial intelligence methods, specifically machine learning (ML) and deep learning (DL), have opened up a fundamentally new way of extracting actionable insights from raw data, which could augment and potentially replace some aspects of traditional evaluation workflows. In this review, we summarize current and emerging applications of DL in histopathology and genomics, including basic diagnostic as well as advanced prognostic tasks. Based on a growing body of evidence, we suggest that DL could be the groundwork for a new kind of workflow in oncology and cancer research. However, we also point out that DL models can have biases and other flaws that users in healthcare and research need to know about, and we propose ways to address them.

Mismatch Repair Deficiency (dMMR)/Microsatellite Instability (MSI) is a key biomarker in colorectal cancer (CRC). Universal screening of CRC patients for MSI status is now recommended, but contributes to increased workload for pathologists and delayed therapeutic decisions. Deep learning has the potential to ease dMMR/MSI testing and accelerate oncologist decision making in clinical practice, yet no comprehensive validation of a clinically approved tool has been conducted. We developed MSIntuit, a clinically approved artificial intelligence (AI) based pre-screening tool for MSI detection from haematoxylin-eosin (H&E) stained slides. After training on samples from The Cancer Genome Atlas (TCGA), a blind validation is performed on an independent dataset of 600 consecutive CRC patients. Inter-scanner reliability is studied by digitising each slide using two different scanners. MSIntuit yields a sensitivity of 0.96-0.98, a specificity of 0.47-0.46, and an excellent inter-scanner agreement (Cohen's κ: 0.82). By reaching high sensitivity comparable to gold standard methods while ruling out almost half of the non-MSI population, we show that MSIntuit can effectively serve as a pre-screening tool to alleviate MSI testing burden in clinical practice.

A knowledge gap persists between machine learning (ML) developers (e.g., data scientists) and practitioners (e.g., clinicians), hampering the full utilization of ML for clinical data analysis. We investigated the potential of the ChatGPT Advanced Data Analysis (ADA), an extension of GPT-4, to bridge this gap and perform ML analyses efficiently. Real-world clinical datasets and study details from large trials across various medical specialties were presented to ChatGPT ADA without specific guidance. ChatGPT ADA autonomously developed state-of-the-art ML models based on the original study's training data to predict clinical outcomes such as cancer development, cancer progression, disease complications, or biomarkers such as pathogenic gene sequences. Following the re-implementation and optimization of the published models, the head-to-head comparison of the ChatGPT ADA-crafted ML models and their respective manually crafted counterparts revealed no significant differences in traditional performance metrics (p ≥ 0.072). Strikingly, the ChatGPT ADA-crafted ML models often outperformed their counterparts. In conclusion, ChatGPT ADA offers a promising avenue to democratize ML in medicine by simplifying complex data analyses, yet should enhance, not replace, specialized training and resources, to promote broader applications in medical research and practice.