JCO Clinical Cancer Informatics最新文献

Purpose: The increasing use of patient portal messages has enhanced patient-provider communication. However, the high volume of these messages has also contributed to physician burnout.

Methods: Patient-generated portal messages sent to a single cancer center from 2011 to 2023 were extracted. BERTopic, a natural language processing topic modeling technique based on large language models, was optimized. For further categorization, the topic words were labeled using GPT-4, followed by review by two oncologists. Uniform Manifold Approximation and Projection was used for dimensionality reduction and visualizing topics. Message volume changes over time were assessed using a Student's t test.

Results: A total of 2,280,851 messages were analyzed. The monthly average number of messages increased from 2,071 in 2012 to 43,430 in 2022 (P < .001). There was a significant rise in message volume after the COVID-19 pandemic, with a posterior probability of a causal effect of 96.4% (P = .04). Scheduling-related messages were the most frequent across departments, whereas symptoms and health concerns were second or third most common topics. In medical oncology and surgical oncology, topics on prescriptions and medications were more common compared with radiation oncology and gynecologic oncology. Despite concurrent institutional changes in self-scheduling systems, scheduling-related messages did not decrease over time.

Conclusion: The substantial increase in patient portal messages, particularly scheduling-related inquiries, underscores the need for streamlined communication to reduce the burden on health care providers. These findings highlight the need for strategies to manage message volume and mitigate physician burnout, laying groundwork for artificial intelligence-driven future triage systems to improve message management and patient care.

Purpose: Flow cytometry is a key diagnostic technique in hematology that provides protein information at a single-cell level. Traditionally interpreted manually in a sequence of two-dimensional plots, automated analysis techniques have grown in significance in both research and clinics improving interrater reliability and speeding up analysis. Published tools usually require a specific diagnostic setup, which hinders widespread implementation.

Methods: In this paper, we present the development of a software package and web app (diagnFlow) for the automated analysis of any in-house clinical flow cytometry data set. We exemplify the application of this classifier and its clinical benefit in lymphoma diagnosis and other settings.

Results: Routine performance for the focused diagnostic task was evaluated in a blinded one-examiner setup. Multiple customary workflows solving the task in an automated manner were designed using diagnFlow. Each workflow could improve on the performance of the manual interpretation. The most easily interpretable and computationally efficient workflow out-performed more complicated approaches and was made available as an easy-to-use web app. Same-sample wet laboratory data further elucidated the biological signal the classifier is based on. The approach made available as a web app was validated in additional data sets outperforming a competition-winning clustering-based approach.

Conclusion: diagnFlow provides a valuable data set-agnostic approach to flow cytometry data sets previously not leveraged for automatic analysis while maintaining interpretability and resource efficiency.

Purpose: Goals-of-care (GOC) discussions during advanced serious illness and end-of-life (EOL) care are critical. Institutions are increasingly tracking the frequency and timing of GOC documentation, but large-scale content assessments have been limited. We aimed to use natural language processing (NLP) to assess GOC documentation quality and associations with EOL care for patients with cancer.

Methods: This is a retrospective review of patients at a single US center who died with cancer between 2018 and 2022, and had documented GOC notes in the last 12 months of life. Eight GOC components were identified: current understanding of illness, information preferences, prognostic disclosure, goals, fears, acceptable function, trade-offs, and family involvement. NLP software searched for the aggregate presence of these components at the patient level within extracted GOC notes. We evaluated associations between these eight components and receipt of aggressive EOL care (chemotherapy within 14 days of death, no hospice care, or hospice admission ≤3 days of death).

Results: Two thousand thirty-one patients met inclusion criteria. The most common GOC component addressed was family involvement (75.0%) and the least common was fears (21.1%). Only 5.4% had all eight components documented. More comprehensive GOC notes were associated with lower rates of aggressive EOL care; 73.2% received aggressive care when 0/8 components were documented, compared with 56.8% and 50.3% with six or seven components discussed, respectively. In multivariate logistic regression, GOC components documented (≤6 v ≥7: OR, 2.13; P < .0001) and primary tumor site (lymphoma: OR, 2.86; P < .0001) were independent predictors of aggressive EOL care.

Conclusion: Increasingly comprehensive and higher-quality GOC documentation is associated with a lower likelihood of receiving aggressive EOL care. Opportunities to improve the quality and documentation of GOC conversations may affect EOL care for patients with cancer.

Purpose: Lung cancer is a leading cause of death in Chile, where late-stage diagnoses and high mortality rates prevail. Here, we describe the development of OncovigIA, a novel digital tool powered by natural language processing that enhances the identification of potential lung cancer cases by surveilling computed tomography (CT) reports in a large public Hospital in Santiago, Chile.

Materials and methods: We combined natural language processing and large language models with state-of-the-art machine learning techniques and approaches to treat unbalanced data sets and determine the best solution to implement in OncovigIA. Focusing on key sections of the reports and using various machine learning models, including a balanced Random Forest, the tool achieved high performance with 0.90 accuracy and 0.84 F1-score on the test set.

Results: When applied to 13,326 CT chest reports from 2022, it successfully identified 377 CTs of patients with suspected lung cancer previously undetected and not managed by the multidisciplinary local lung cancer team.

Conclusion: This study underscores the potential of artificial intelligence in early cancer detection and highlights the importance of its integration into local health care ecosystems. By promptly increasing the number of patients referred for specialized management, the tool OncovigIA offers a promising path toward improving lung cancer survival rates in Chile and beyond. Moreover, this article provides avenues for its broader implementation, extending it to other cancer types and/or health care-related texts for continuous surveillance, aiming at the early referral and treatment of cancer in low-resource settings.

Purpose: Clinical guidelines recommend offering individualized physical activity prescriptions to cancer survivors. However, there are limited tools to support individualized physical activity discussions and prescriptions. We developed and validated a simulation model-based tool to estimate individualized survival outcomes for postdiagnosis physical activity among postmenopausal breast cancer survivors.

Methods: We adapted an established simulation modeling approach developed within the Cancer Intervention and Surveillance Modeling Network to estimate breast cancer-specific and all-cause survival associated with postdiagnosis physical activity for 50- to 75-year-old (postmenopausal) women with stage I to III invasive breast cancer. Model estimates were generated for 60,480 subgroups based on age, weight status (BMI), stage, tumor subtype, treatment, aerobic (<30 min/wk [no/minimal], ≥30 to <150 min/wk [insufficient], ≥150 to <300 min/wk [active], ≥300 min/wk [highly active]), and muscle-strengthening (<2 or ≥2 d/wk) activity. The outcomes were 10-year survival and absolute survival benefits for different levels of physical activity by individual characteristics and treatment. Model inputs were derived from trials, cohort studies, registry, and surveillance data. External validation used independent data.

Results: Survival rates and absolute benefits for physical activity varied by age, weight status, stage, tumor subtype, and amount and type of activity. For example, the 10-year breast cancer-specific and all-cause survival for no/minimal activity in a 65- to 69-year-old-woman with stage II, hormone receptor-positive, human epidermal growth factor receptor 2-negative breast cancer with obesity was 79.2% and 72.2%, respectively. Increasing aerobic activity from no/minimal to insufficient activity with <2 d/wk of muscle-strengthening was associated with absolute increases in 10-year breast cancer-specific and all-cause survival by 2.8 and 3.4 percentage points, respectively. The model closely replicated survival rates in independent data.

Conclusion: Simulation model-based estimates could support clinical tools for guideline-recommended individualized discussions and physical activity prescriptions for breast cancer survivors.

Purpose: Cancer remains a leading cause of death worldwide. The growing volume of high-throughput single-cell and spatial transcriptomic data sets-particularly those related to cancer-offers immense opportunities as well as analytical challenges for effective data analysis and interpretation. Large language models (LLMs), pretrained on vast data sets and capable of various biomedical tasks, offer a promising solution. This review explores the application of LLMs in cancer research from both cellular and pathologic perspectives, aiming to showcase their potential in advancing precision oncology.

Materials and methods: We systematically review current LLMs in analyzing single-cell RNA sequencing, spatial transcriptomic, and histology image data, emphasizing their relevance to cancer biology and translational research.

Results: A total of 24 LLMs, published or in preprint between 2022 and 2025, were selected for review. In single-cell transcriptomics, LLMs have primarily been used for cell type annotation, batch integration, and drug-response prediction. In spatial transcriptomics, LLMs support multislide and multimodal spatial data integration, gene expression imputation, niche and region label prediction, spatial domain identification, cell-cell communication inference, and marker gene detection. In computational pathology, LLMs have been applied to cancer subtyping, detection of rare malignancies, genomic mutation prediction, image segmentation, as well as cross-modal retrieval. Despite these advances, many models remain underoptimized for cancer-specific applications, highlighting the need for domain-specific fine-tuning and scalable adaptation strategies.

Conclusion: LLMs have the potential to significantly advance cancer research by providing scalable and effective tools for analyzing and interpreting single-cell, spatial transcriptomic, and pathology data. Future efforts should prioritize tailoring these models to cancer-specific contexts to enhance their utility in uncovering disease mechanisms, identifying biomarkers, and informing therapeutic strategies.

Purpose: Unplanned treatment interruptions represent an important care quality shortfall for patients undergoing cancer radiotherapy. This study aimed to evaluate use of a centralized electronic health record warehouse and large language model-based data preprocessing to facilitate identification of risk factors for radiation therapy interruptions (RTI).

Methods: We analyzed demographic, behavioral, clinical, and neighborhood-level data for 2,130 patients treated with radiotherapy at the University of Tennessee Medical Center in Knoxville. Treatment interruptions were measured as missed days, adjusted for weekends and holidays. Multinomial logistic regression was used to identify factors associated with moderate (2-4 days) and severe (≥5 days) RTI.

Results: Moderate RTI occurred in 15.8% of patients, while 7.7% experienced severe RTI. Moderate delays were associated with genitourinary cancer (adjusted odds ratio (AOR), 3.81; 95% CI, 1.24 to 11.66), prostate cancer (AOR, 2.44; 95% CI, 1.34 to 4.46), and Medicaid coverage (AOR, 2.22; 95% CI, 1.32 to 3.73). Severe RTI was associated with marital status (AOR for divorced or separated patients, 1.86; 95% CI, 1.18 to 2.94), head and neck cancer (AOR, 2.31; 95% CI, 1.10 to 4.87), gynecologic cancer (AOR, 2.97; 95% CI, 1.30 to 6.79), Medicaid insurance (AOR, 3.43; 95% CI, 1.77 to 6.64), daily dose of ≤225 cGy (AOR, 2.55; 95% CI, 1.21 to 5.37), and a total dose of ≥6,000 cGy (AOR, 2.30; 95% CI, 1.09 to 4.88). Severe interruptions were also significantly associated with high neighborhood social vulnerability (AOR, 2.60; 95% CI, 1.32 to 5.09).

Conclusion: Automated data preprocessing permitted efficient identification of treatment course length, marital status, disease site, Medicaid coverage, and socially vulnerable locations as significant factors associated with RTI. These findings underscore the need for data-driven risk assessment and intervention strategies to maintain cancer treatment quality at scale.

Over the past 5 years, large language models (LLMs) have emerged and continued to improve in their generative abilities and are now capable of generating human-understandable text and performing complex data analyses. As these models continue to improve in their capabilities, they are increasingly used to support population oncology, including clinical information extraction, cancer care education, and clinical decision support. This narrative review provides a high-level description of the use of LLMs in cancer with an overview of the current literature, along with research gaps. Despite increasing interest in using LLMs for cancer care, prevention, and research, applied methods in cancer still lag advancements published in the computer science literature. Therefore, we recommend that cancer-focused LLM research and applications better incorporate technical advancements and techniques found in the computer science literature. Additionally, standardized evaluation metrics and approaches need to be better studied and adopted in oncology, along with data governance and computational infrastructure to support state-of-the-art model integration and the use of real-world data. Finally, we describe the need for researchers to incorporate principles and frameworks from implementation and dissemination science to promote LLM-based tool adaptation, effectiveness, fit, and sustainability.

Purpose: Radiotherapy can have an impact on the immune system through lymphocyte depletion. In particular, it can influence tumor control. Accurately predicting radiation-induced lymphopenia (RIL) is key to optimizing treatment strategies. We aimed to evaluate mathematical model structures capable of capturing the kinetics of lymphocyte concentrations after irradiation, with attention to the saturation effect observed during fractionated radiotherapy.

Materials and methods: A meta-analysis of aggregate data from patients with solid tumor treated with specific radiation doses was conducted. We extracted patient and treatment characteristics, lymphocyte counts, including tumor type/location, radiation modality (X-rays v protons), and baseline counts. Various models-including exponential, extended exponential, exponential-quadratic, and saturation models-were tested for their ability to predict lymphocyte kinetics.

Results: Data from 29 studies covering brain, nasopharyngeal, oropharyngeal, esophageal, non-small cell lung, hepatocarcinoma, cervical, pancreatic, rectal cancers, and soft tissue sarcoma were analyzed. Lymphocyte depletion rates varied, with cervical cancer showing the highest reduction, followed by esophageal, hepatocarcinoma, and others. Lymphocyte recovery post-treatment depended heavily on baseline counts and time since radiotherapy completion. Saturation models best fit most cancer types, but for head and neck/central nervous system cancer without nodal involvement, the exponential-quadratic model performed better, reflecting a unique early lymphocyte increase before RIL.

Conclusion: In conclusion, the choice of lymphocyte depletion model should align with cancer type and location. Standardized prospective studies are needed to refine models and enhance radiotherapy strategies.