Engaging Preference Optimization Alignment in Large Language Model for Continual Radiology Report Generation: A Hybrid Approach

Cognitive Computation, Год журнала: 2025, Номер 17(1)

Опубликована: Янв. 27, 2025

Язык: Английский

---

The data scientist as a mainstay of the tumor board: global implications and opportunities for the global south

Frontiers in Digital Health, Год журнала: 2025, Номер 7

Опубликована: Фев. 6, 2025

Tumor boards are multidisciplinary teams of healthcare professionals that working together to encompass the full spectrum care around diagnosing, planning treatment, and advising outcomes for individual cancer patients. These typically consist oncologists, radiologists, pathologists, geneticists, surgeons, nurse practitioners, other palliative (National Cancer Institute, 2024). create a collaborative space experts from various disciplines assess clinical factors patient circumstances, ensuring application appropriate standards personalized recommendations National Comprehensive Network (NCCN) Guidelines enhance treatment met. Since no fits "textbook" profile, oncologists benefit discussing tailored plans learning their colleagues' experiences. When tumor functioning well, they can have significant impact on (NCCN, 2025). For instance, thoracic oncology board in Munich, Germany, found 90% met or exceeded standards, with nearly being implemented practice (Walter et al, 2023).Tumor increasingly used worldwide, but expertise resources conducting still limited Global South. However, this does not mean cannot be developing countries. A 2020 survey Southeast Asia 80.4% pediatric solid units had pediatric-trained specialists, including radiation nuclear medicine physicians, nurses. This indicates already place these specialists play critical role (Ottman, 2020). With implementation global south, data scientists further AI analytics improve decision-making personalize care.Advances big data, machine (ML), artificial intelligence (AI) provide more precise, evidence-based, patient-specific care, thus, giving different approach as how diagnose, treat, manage patients (Alowais 2023). there is growing number complexity industry such Electronic Health Records (EHRs), next-generation genomic sequencing (NGS), advanced imaging modalities like X-ray Radiography, Magnetic Resonance Imaging (MRI), Computed Tomography (CT) scans. analyzing individually manually, time-consuming considerably impractical. where decision support systems (CDSS) powered by ML put into action. predictive analysis disease progression prognosis, based patients' drug-drug interaction alerts (Wang 2023;Alowais As precision continue evolve, will rely data-driven tools reduce errors, overall health (Khalifa Albadawy, Data process analyze large datasets identify biomarkers predict respond specific treatments (Nardone In addition, algorithms interpret radiological images, detect early signs cancer, progression. becoming standard boards, especially high-income countries (Bi 2019;El Saghir 2015).For oncology, most commonly diagnostic guide targeted therapies Polymerase Chain Reaction (PCR), fluorescent situ hybridization (FISH), immunohistochemistry (IHC) (Goosens 2015). high-throughput (NGS)-based diagnostics, which somatic mutations tumors, proven clinically useful identifying single-nucleotide mutations, insertions, deletions, rearrangements (Kamps 2017). Thus, multigene NGS testing oncologist picture molecular profile utilized best option (Mehta 2020).As continues gain prominence characterization cancers becomes complex (Specchia al., 2020;Nardone 2024), incorporating essential. bring ML, analysis, bioinformatics, enabling make accurate, evidence-based decisions lead improved 2024;Rodriguez Ruiz 2022). They synthesizing diverse generated uncovering actionable insights, informing strategies. particularly crucial shifts focus toward approaches genetic characteristics tumors (Subrahmanya, 2022).Specifically, apply statistical techniques survival clustering, modeling uncover insights inform decisions.Their knowledge foundation models, Generative Pre-trained Transformer (GPT), Bidirectional Encoder Representations Transformers (BERT), memory-augmented neural networks enables them extract valuable unstructured medical records pathology reports 2024;Wang 2023).Globally, trend towards integrating care. countries, AI-based assist clinicians interpreting predicting outcomes, optimal United States, example, institutions Memorial Sloan Kettering Center using tools, algorithms, real time during discussions. integration varies across regions, some low-and middle-income facing barriers adoption due lack infrastructure (Zuhair 2024).In States Europe, key members boards. at University Florida collaborate develop models (UF Health, AI-driven responses potential trials. Similarly, work real-world integrate it decision-making, each receives unique (Harris 2023).Data-driven provided revolutionized treatment. By large-scale datasets, driving suggest likely effective than (Berger Mardis, 2018). Predictive also forecast allowing tailor predicted response. trial-and-error nature efficient 2023).Al promising innovation imaging, applications ranging image acquisition processing reporting, follow-up planning, management. Given broad scope applications, anticipated daily radiologists (Pesapane The challenge, however, lies AI-powered equipment information training many professionals, radiologists. preparation may contribute reluctance adopt radiology fields (Waymel, 2019;Pesapane Nevertheless, transformative advancements only occurring academic hospitals highly facilities, regions communities grappling greatest challenges disparities (Sitek, 2024).While made strides science LMICs face challenges. include computational infrastructure, insufficient access high-quality shortage trained capable (Alami Additionally, concerns about algorithmic bias ethical implications healthcare, populations (Siala Wang, Overcoming require investment both technology human capital, well development frameworks use settings.In South, includes Asia, Africa, Latin America Caribbean (United Nations Development Programme (UNDP), 2004), often hampered outdated infrastructure. higher rates late-stage diagnoses poorer compared (Bamodu Chung, parts Sub-Saharan Africa travel long distances leading delays diagnosis (Mwamba Moreover, underfunded latest advances Just Philippines, main challenge difficulty financial toxicity brings family (Fernandez & Ting, Thus despite national incorporate country (Loong 2023) very challenging day-to-day practice. If cost were limiting factor, Philippines would managing (Catedral 2020).Data address South leveraging optimize resource allocation accuracy. telemedicine platforms mobile (mHealth) real-time rural areas (Haleem 2021;Akingbola progression, high risk complications, prioritize those need (Alowais, IBM's Watson Oncology (WFO), an CDSS therapy selection (Liu 2018), beneficial tool Hence, track monitor responses, 2019). applied even absence equipment, help regions.While studies settings LMICs, specifically limited. Kenya, screen cervical areas, significantly reducing While Ethiopia, been blood smear images diagnose leukemia accuracy (Akingbola examples demonstrate revolutionize providing affordable, scalable solutions pressing Academy International (USAID) has making efforts gap highlighting actions effectively promote (USAID, 2022).Transdisciplinarity emerged multiple integrated tackle problems angles. incorporates domains medicine, science, social sciences, ethics. pooling fields, providers offer comprehensive patients, (Van Bewer, Complex Hospital S.G. Bosco Turin, nurses psychologists, workers worked unmet needs innovative projects (Clementi Transdisciplinary successful strategy expediting emergency department (ED) flow. Through collaboration allied team was able efficiently, prompt delivery (Innes 2016). secondary BRIGHT Study chronic illness management after heart transplant revealed centers dedicated achieved better (p=0.042) (Cajita, Similar disciplines, leverage resources, project linking 54 million electronic England (Wood, 2021). highlight within transdisciplinary sectors.To must meet range technical, domain-specific, interpersonal requirements. modeling, essential, oncology-related omic records. Candidates should hold graduate-level degree discipline strong emphasis statistics mathematics, statistics, physics, biology, computer electrical engineering, biomedical engineering (BME), related fields. level ensures ability handle heterogeneous while adhering regulations similar Insurance Portability Accountability Act (HIPAA) maintain privacy confidentiality.A robust understanding terminology workflows seamless communication professionals. Furthermore, excel translating findings employing visualization facilitate disciplines. Beyond technical skills, abilities vital environment To ensure quality consistency contributions, eligibility regulation international professional bodies. Lastly, commitment continuous adapt emerging innovations medicine.Insights study Fermin Tan (2021), BME formal discipline, importance formalized educational pathways recognition healthcare. research demonstrated recognizing field achieve impactful innovations. Applying lessons emphasizes structured education programs regulatory LMIC contexts.Efforts standardize qualifications competencies EDISON Science Framework (EDSF), provides professionalization comprising components Competence (CF-DS), Body Knowledge (DS-BoK), Model Curriculum (MC-DS), Professional Profiles (DSPP) (Demchenko 2017a(Demchenko , 2017b(Demchenko 2017c(Demchenko 2017d)). American Medical Informatics Association (AMIA) competency-based accreditation informatics, aligns closely roles (Valenta Computing Machinery (ACM) supports computing undergraduate curricula, detailing essential skills (ACM, 2021).Country-specific vary; skills-based hiring under Executive Order 14110 practical over (US Office Personnel Management [OPM], Occupational Standards (NOS) Kingdom outlines detailed performance criteria life sciences (Unique Registration Number [URN] COGBIO-05), applicable specialized Standards, 2018).The scientist plays synthesizer knowledge, patterns large, disparate domains, clinical, genomic, environmental (Hassan Within do employ variety breadth contribution extends beyond reinforcement learning, Bayesian networks, simulation-based approaches, others.Reinforcement type algorithm learns sequences maximizing cumulative rewards, (Coronato account differences between classic RL following Markov assumption future state system depends its current (Kuznetsov 2010). suggesting adaptation model. dynamic strategies patient's response ongoing treatments. continuously adjust dosages chemotherapy minimize effectiveness (Eckardt Tempo, novel framework screening, context breast cancer. Tempo policy, combined model, outperforms practices detection adapted screening preferences. It improves overscreening (Yala allows time, adaptive new emerges. tailors assessments profiles, enhancing precision.Data probabilistic graphical represent set variables conditional dependencies. likelihood observed data. setting, sources-clinical biomarkers, history-to estimate probabilities uncertainty quantifications, helping doctors informed cases ambiguity (Polotskaya Huehn al developed digital model relevant head neck squamous cell carcinoma (HNSCC). Validation showed guides immunotherapy decisions, 84% concordance (Cohen's κ = 0.505, p 0.009) when actual 25 created physician's patient.Simulation-based enable virtual scenarios evaluate outcomes. simulating strategies, explore consequences before applying simulations options, long-term effects profiles (Nave, Federov (2020), method optimizing Treating Fields (TTFields) brain tumors. TTFields, delivered through transducer arrays skin, inhibit growth, distribution varying array placement, anatomy, characteristics. Incorporating expected physician TTFields ultimately improving outcomes.In addition dose optimization amount timing radiation. aim balance efficacy minimizing side effects, involve toxic agents. adjusting dosing schedules metabolism characteristics, duration frequency increase probability success without compromising (Bräutigam, emergence anti-tumor complicates this, creating urgent optimized dose-schedule designs doses concurrently single trial (Chen recent deep (DL) led DL-based prediction models. Unlike traditional methods, DL automatically extracts features CT, MRI, PET scans map values, guiding final distribution. distributions anatomical prescriptions (Jiang 2024).Multi-omics another important facet planning. combine genomics, transcriptomics, proteomics, metabolomics tumor. therapies. multi-omics might reveal just mutation interacts pathways, treatments, targeting metabolic aberration (Babu Snyder, Multi-omics offers view volumes pose analytical helps extracting omics advancing (Li Cai al. (2022), explored methods research, general-purpose task-specific approaches. benchmarked five Cell Line Encyclopedia, assessing classification, drug prediction, runtime efficiency. Their paper selecting encourages advance discovery treatments.Radiomics involves quantitative (e.g., scans) heterogeneity pathomics analyzes histopathological discernible pathologist alone. image-derived predictions select therapeutic Like (2024), radiopathomics classify stage I, II, III gastric Other researchers prognosis colorectal lung cancers, 2020a(Wang 2020b)). Radiomics situations available, offering non-invasive options (Gillies 2016;Brancato .Spatial biology technologies, GeoMx® (NanoString Technologies®) 1 CosMx™ 2 Visium® (10x Genomics®) 3 Xenium™ 4 revolutionizing profiling spatial context. mapping heterogeneity, microenvironment, cell-cell interactions, bulk offer. transcriptomics (spTx) 5 6 combining high-resolution RNA profiling, capturing cellular organization biomarker localization tissue samples (Cook HD 7 sub-cellular resolution, reconstruction morphology expression (Polanski 2024) .One notable spTx Despite faces intratumoral (ITH), tumour differently drugs. Using spTx, shows sensitivity tumor, core periphery. finds genetically identical cells depending location (Jimenez-Santos consider surrounding microenvironment. could addressing tumor's complexity, chances failure.Causal focuses determining cause-and-effect relationships, going correlation interventions Peter-Clark (PC) (Spirtes 1993) latent Gaussian causal (Cai SHapley Additive exPlanations (SHAP) Local Interpretable Agnostic Explanation (LIME) primarily explain correlations rather causation (Ladbury 2022).For language (LLM) impacting Non Small Lung (NSCLC), revealing potentially unexpected relationships smoking status having effect choice (Naik further, infe

Язык: Английский

---

JAVIS Chat: A Seamless Open-Source Multi-LLM/VLM Deployment System to Be Utilized in Single Computers and Hospital-Wide Systems with Real-Time User Feedback

Applied Sciences, Год журнала: 2025, Номер 15(4), С. 1796 - 1796

Опубликована: Фев. 10, 2025

The rapid advancement of large language models (LLMs) and vision-language (VLMs) holds enormous promise across industries, including healthcare but hospitals face unique barriers, such as stringent privacy regulations, heterogeneous IT infrastructures, limited customization. To address these challenges, we present the joint AI versatile implementation system chat (JAVIS chat), an open-source framework for deploying LLMs VLMs within secure hospital networks. JAVIS features a modular architecture, real-time feedback mechanisms, customizable components, scalable containerized workflows. It integrates Ray distributed computing vLLM optimized model inference, delivering smooth scaling from single workstations to hospital-wide systems. consistently demonstrates robust scalability significantly reduces response times on legacy servers through Ray-managed multiple-instance models, operating seamlessly diverse hardware configurations enabling departmental By ensuring compliance with global data protection laws solely closed networks, safeguards patient while facilitating adoption in clinical This paradigm shift supports care operational efficiency by bridging potential utility, future developments speech-to-text integration, further enhancing its versatility.

Язык: Английский

---

The foundational capabilities of large language models in predicting postoperative risks using clinical notes

npj Digital Medicine, Год журнала: 2025, Номер 8(1)

Опубликована: Фев. 11, 2025

Clinical notes recorded during a patient's perioperative journey holds immense informational value. Advances in large language models (LLMs) offer opportunities for bridging this gap. Using 84,875 preoperative and its associated surgical cases from 2018 to 2021, we examine the performance of LLMs predicting six postoperative risks using various fine-tuning strategies. Pretrained outperformed traditional word embeddings by an absolute AUROC 38.3% AUPRC 33.2%. Self-supervised further improved 3.2% 1.5%. Incorporating labels into training increased 1.8% 2%. The highest was achieved with unified foundation model, improvements 3.6% 2.6% compared self-supervision, highlighting foundational capabilities risks, which could be potentially beneficial when deployed care.

Язык: Английский

---

Performance of ChatGPT versus Google Bard on Answering Postgraduate-Level Surgical Examination Questions: A Meta-Analysis

Indian Journal of Surgery, Год журнала: 2025, Номер unknown

Опубликована: Фев. 17, 2025

Язык: Английский

---

Building an intelligent diabetes Q&A system with knowledge graphs and large language models

Frontiers in Public Health, Год журнала: 2025, Номер 13

Опубликована: Фев. 20, 2025

This paper introduces an intelligent question-answering system designed to deliver personalized medical information diabetic patients. By integrating large language models with knowledge graphs, the aims provide more accurate and contextually relevant guidance, addressing limitations of traditional healthcare systems in handling complex queries. The combines a Neo4j-based graph Baichuan2-13B Qwen2.5-7B models. To enhance performance, Low-Rank Adaptation (LoRA) prompt-based learning techniques are applied. These methods improve system's semantic understanding ability generate high-quality responses. performance is evaluated using entity recognition intent classification tasks. achieves 85.91% precision 88.55% classification. integration structured significantly improves accuracy clinical relevance, enhancing its responses for diabetes management. study demonstrates effectiveness graphs systems. proposed approach offers promising framework advancing management other applications, providing solid foundation future interventions.

Язык: Английский

---

Advancing Ophthalmology With Large Language Models: Applications, Challenges, and Future Directions

Survey of Ophthalmology, Год журнала: 2025, Номер unknown

Опубликована: Март 1, 2025

Язык: Английский

---

Appropriateness of Thyroid Nodule Cancer Risk Assessment and Management Recommendations Provided by Large Language Models

Deleted Journal, Год журнала: 2025, Номер unknown

Опубликована: Март 3, 2025

The study evaluates the appropriateness and reliability of thyroid nodule cancer risk assessment recommendations provided by large language models (LLMs) ChatGPT, Gemini, Claude in alignment with clinical guidelines from American Thyroid Association (ATA) National Comprehensive Cancer Network (NCCN). A team comprising a medical imaging informatics specialist two radiologists developed 24 clinically relevant questions based on ATA NCCN guidelines. readability AI-generated responses was evaluated using Readability Scoring System. total 322 training or practice United States, recruited via Amazon Mechanical Turk, assessed AI responses. Quantitative analysis SPSS measured recommendations, while qualitative feedback analyzed through Dedoose. compared performance three providing appropriate recommendations. Paired samples t-tests showed no statistically significant differences overall among models. achieved highest mean score (21.84), followed closely ChatGPT (21.83) Gemini (21.47). Inappropriate response rates did not differ significantly, though trend toward higher rates. However, accuracy (92.5%) responses, (92.1%) (90.4%). Qualitative highlighted ChatGPT's clarity structure, Gemini's accessibility but shallowness, Claude's organization occasional divergence focus. LLMs like show potential supporting require oversight to ensure performed nearly identically overall, having score, difference marginal. Further development is necessary enhance their for use.

Язык: Английский

---

Application of large language models in healthcare: A bibliometric analysis

Digital Health, Год журнала: 2025, Номер 11

Опубликована: Янв. 1, 2025

The objective is to provide an overview of the application large language models (LLMs) in healthcare by employing a bibliometric analysis methodology. We performed comprehensive search for peer-reviewed English-language articles using PubMed and Web Science. selected were subsequently clustered analyzed textually, with focus on lexical co-occurrences, country-level inter-author collaborations, other relevant factors. This textual produced high-level concept maps that illustrate specific terms their interconnections. Our final sample comprised 371 journal articles. study revealed sharp rise number publications related LLMs healthcare. However, development geographically imbalanced, higher concentration originating from developed countries like United States, Italy, Germany, which also exhibit strong inter-country collaboration. are applied across various specialties, researchers investigating use medical education, diagnosis, treatment, administrative reporting, enhancing doctor-patient communication. Nonetheless, significant concerns persist regarding risks ethical implications LLMs, including potential gender racial bias, as well lack transparency training datasets, can lead inaccurate or misleading responses. While promising, widespread adoption practice requires further improvements standardization accuracy. It critical establish clear accountability guidelines, develop robust regulatory framework, ensure datasets based evidence-based sources minimize risk reliable use.

Язык: Английский

---

Assessing the performance of large language models (GPT-3.5 and GPT-4) and accurate clinical information for pediatric nephrology

Pediatric Nephrology, Год журнала: 2025, Номер unknown

Опубликована: Март 5, 2025

Artificial intelligence (AI) has emerged as a transformative tool in healthcare, offering significant advancements providing accurate clinical information. However, the performance and applicability of AI models specialized fields such pediatric nephrology remain underexplored. This study is aimed at evaluating ability two AI-based language models, GPT-3.5 GPT-4, to provide reliable information nephrology. The were evaluated on four criteria: accuracy, scope, patient friendliness, applicability. Forty specialists with ≥ 5 years experience rated GPT-4 responses 10 questions using 1-5 scale via Google Forms. Ethical approval was obtained, informed consent secured from all participants. Both demonstrated comparable across criteria, no statistically differences observed (p > 0.05). exhibited slightly higher mean scores parameters, but negligible (Cohen's d < 0.1 for criteria). Reliability analysis revealed low internal consistency both (Cronbach's alpha ranged between 0.019 0.162). Correlation indicated relationship participants' professional their evaluations (correlation coefficients - 0.026 0.074). While provided foundational level support, neither model superior addressing unique challenges findings highlight need domain-specific training integration updated guidelines enhance reliability fields. underscores potential while emphasizing importance human oversight further refinements applications.

Язык: Английский

---