Large Language Models (LLMs) have recently emerged as a potentially transformative technology in clinical oncology, offering promising opportunities for advancing precision medicine. This comprehensive review critically examines the multifaceted applications of LLMs across the cancer care continuum, encompassing diagnosis, treatment planning, and supportive care. In cancer diagnosis, LLMs exhibit promising potential for enhancing the accuracy and efficiency of pathological and radiological assessments, as well as facilitating the interpretation of complex genomic data. In the context of treatment decision-making, these models demonstrate the capability to integrate vast amounts of clinical data, research findings, and patient characteristics to propose personalized treatment strategies. LLMs also demonstrate potential for enhancing patient education, symptom management, and psychological support. Notwithstanding their potential, significant challenges persist, including the requirement for high-quality annotated datasets, limitations in model interpretability, and ethical considerations regarding medical practice and patient privacy. This review provides a critical analysis of the current state of LLM applications in oncology, elucidating both their transformative potential and the obstacles impeding their widespread clinical implementation. The review examines emerging trends, including the integration of multimodal data analysis and advancements in explainable artificial intelligence (AI), which have the potential to address some of these challenges. Through providing a comprehensive overview of the opportunities and challenges presented by LLMs in oncology, this review seeks to inform clinicians, researchers, and policymakers, thereby promoting further innovation and responsible integration of AI technologies in cancer care with the ultimate goal of improving patient outcomes.
Keywords: Large Language Models | Generative AI | Precision Oncology | Diagnosis | Prognosis
Junyi Shen, Rong Hu, Qiliang Huang, Kailai Li, Jian Zhang
Published online:
Received: · Accepted:
This article: Open access under CC BY-ND 4.0.
Abstract
Bispecific T-cell engagers (BiTEs) are pivotal for hematological malignancies. However, real-world safety assessments often emphasize adverse event (AE) incidence while overlooking severity and fatality rates. Furthermore, toxicity heterogeneity across BiTE targets and their comparative safety against chimeric antigen receptor T-cell (CAR-T) therapy remain incompletely elucidated. To address these gaps, a retrospective pharmacovigilance study using the FAERS and VigiBase databases (2015–2025) employed disproportionality analysis (reporting odds ratios) to identify BiTE-associated high-occurrence (BAHO) and high-fatality (BAHF) AEs. By integrating incidence and fatality data, we developed a clinical risk stratification framework. Time-to-onset kinetics, co-occurrence patterns, and comparative safety versus CAR-T therapies were also evaluated. Analyzing 18,928 reports (9,350 FAERS; 9,578 VigiBase) identified 198 BAHO-AEs and 36 BAHF-AEs. Cytokine release syndrome (CRS) emerged as a central toxicity hub, frequently co-occurring with ICANS and infections. AEs were classified into three risk patterns: BAHO-nonBAHF (e.g., CRS), BAHO-BAHF (e.g., sepsis, ICANS), and an underrecognized nonBAHO-BAHF category (e.g., cardiac failure, pleural effusion). BiTE toxicities showed an early-onset trend, with immune disorders typically manifesting within seven days, whereas infections and respiratory disorders peaked later (28–90 days). Distinct toxicity profiles were observed across CD19, BCMA, GPRC5D, and CD20-directed agents. Additionally, BiTEs showed a stronger fatality signal for ICANS in FAERS than CAR-T therapies, despite CAR-T's higher overall fatal burden. Ultimately, this large-scale analysis comprehensively maps the BiTE safety landscape. The proposed incidence-fatality framework highlights underrecognized severe toxicities, synergistic lethal risks, and distinct temporal dynamics, providing an evidence-based foundation to optimize precision safety management.
