JCO Clinical Cancer Informatics最新文献

Purpose: To evaluate the appropriateness of a cure model when analyzing right-censored end points of a randomized clinical trial (RCT) in malignancy in the presence of long-term survivors. We aim to derive how the ratio estimation of censored cured subjects (RECeUS), previously proposed for a homogeneous population, could be extended for use in RCTs.

Methods: Based on the RECeUS method, four decision rules were considered to assess the appropriateness of a cure model. They considered the eligibility conditions to be met: in both arms, in at least one randomized arm, in the entire sample, or when only considering an average of the conditions, respectively. A simulation study was performed to evaluate their performance and the impact of the link function when considering the appropriateness of cure models. We also illustrate the method using two real data examples from two RCTs conducted in patients with acute leukemia and COVID-19 disease.

Results: Simulation results show that the best decision rule that can be applied in all considered treatment effect scenarios might be to check the criteria in at least one randomized arm. Regardless of the rules, the cure model appeared to be appropriate in both RCT data.

Conclusion: When analyzing survival data from RCTs, the appropriateness of a cure model could be considered in the face of a plateau shape of the survival curves. To ensure that the presence of such a plateau in the survival curves is a reliable indicator of the presence of cured patients in the population, the RECeUS method should be used in each randomized arm separately, with criteria met in at least one randomized arm.

Purpose: Tobacco use is a major risk factor for diseases such as cancer. Granular quantitative details of smoking (eg, pack years and years since quitting) are essential for assessing disease risk and determining eligibility for lung cancer screening (LCS). However, existing natural language processing (NLP) tools struggle to extract detailed quantitative smoking data from clinical narratives.

Methods: We cross-validated four pretrained Bidirectional Encoder Representations from Transformers (BERT)-based models-BERT, BioBERT, ClinicalBERT, and MedBERT-by fine-tuning them on 90% of 3,261 sentences mentioning smoking history to extract six quantitative smoking history variables from clinical narratives. The model with the highest cross-validated micro-averaged F1 scores across most variables was selected as the final SmokeBERT model and was further fine-tuned on the 90% training data. Model performance was evaluated on a 10% holdout test set and an external validation set containing 3,191 sentences.

Results: ClinicalBERT was selected as the final model based on cross-validation and was fine-tuned on the training data to create the SmokeBERT model. Compared with the state-of-the-art rule-based NLP model and the Generative Pre-trained Transformer Open Source Series 20 billion parameter model, SmokeBERT demonstrated superior performance in smoking data extraction (overall F1 score, holdout test: 0.97 v 0.88-0.90; external validation: 0.86 v 0.72-0.79) and in identifying LCS-eligible patients (97% v 59%-97% for ≥20 pack-years and 100% v 60%-84% for ≤15 years since quitting).

Conclusion: We developed SmokeBERT, a fine-tuned BERT-based model optimized for extracting detailed quantitative smoking histories. Future work includes evaluating performance on larger clinical data sets and developing a multilingual, language-agnostic version of SmokeBERT.

Purpose: Cancer centers and health systems are tasked with deciding where to deploy community interventions to reduce the burden of cancer within their catchment areas. Few methods exist to prioritize communities in a systematic manner, considering features of individuals, populations, systems, and policies. We developed a geographically informed index to prioritize census tracts based on community need, with an initial focus on identifying communities in need of breast cancer screening (BCS) interventions.

Methods: This study used publicly available data to select variables known to be associated with disparities in BCS rates. Variables were identified from five categories: economic stability, education access and quality, neighborhood and built environment, social and community context, and health status and health care access and quality. Data were analyzed at the census tract level across the Sidney Kimmel Comprehensive Cancer Center catchment (N = 1,216). Principal component analysis was applied to 23 variables, and five principal components were selected to construct a composite measure using a weighted sum. The resulting index values were used to stratify the data set for further analysis and mapped for visualization.

Results: The analysis produced the Community Need Priority Index (CNPI)-BCS, with values ranging from 0 to 1 (mean, 0.259; standard deviation [SD], 0.161). The top quintile (Q5, n = 243) represented the highest-need communities. Q5 tracts were primarily concentrated in Philadelphia, Camden, and Delaware counties. Philadelphia County had the highest average (mean, 0.364; SD, 1.78) and the most tracts in the top quintile (45%, n = 175). Montgomery county had the lowest average (mean, 0.169; SD, 0.092).

Conclusion: This novel methodological approach considered the complex nature of multiple, intersectional barriers to good health to identify priority areas of need within cancer center catchment areas.

Purpose: The rapid expansion of scientific literature has made it increasingly challenging for clinicians and researchers to efficiently identify relevant evidence. While large language models (LLMs) offer promising solutions for automating literature review tasks, few tools support integrated workflows that enable trend analysis as well. This study aimed to develop and evaluate Rapid Clinical Evidence eXplorer (RaCE-X), a Generative Pre-trained Transformer (GPT)-based automated pipeline designed to streamline abstract screening, extract structured information, and visualize key trends in clinical research.

Methods: We used GPT-4.1 mini to screen 865 PubMed abstracts based on predefined screening criteria. Structured information was then extracted from the 87 relevant abstracts based on a predefined information model covering nine fields. A gold standard data set was created through expert review to assess model performance. The extracted information was visualized through an interactive dashboard. Usability was evaluated using the Post-Study System Usability Questionnaire (PSSUQ) and open-ended feedback from five clinical research coordinators.

Results: RaCE-X demonstrated high screening performance (precision = 0.954, recall = 0.988, F1 = 0.971) and achieved strong average performance in information extraction (precision = 0.977, recall = 0.989, F1 = 0.983), with no hallucinations identified. Usability testing indicated generally positive feedback (overall PSSUQ score = 2.8), with users noting that RaCE-X was intuitive and effective for data interpretation.

Conclusion: RaCE-X enables efficient GPT-based abstract screening, structured information extraction, and research trend exploration, thereby facilitating the summary of clinically relevant evidence from the biomedical literature. This study demonstrates the feasibility of using LLMs to reduce manual workload and accelerate evidence-based research practices.

Purpose: Analyzing complex medical data sets is specialized and time-consuming. This study aimed to develop and evaluate a novel multiagent artificial intelligence (AI) framework for automating medical data analysis workflows and to compare its performance against nonagent-based approaches using large language models (LLMs).

Methods: A six-party AI agent system was developed using the AutoGen platform, with specialized agents for planning, data retrieval, cleaning, statistical analysis, and review, powered by OpenAI gpt-4o. This framework was applied to deidentified single patient-level data sets from 20 recent studies in the field of bone marrow transplantation (2021-2023). The primary objective was to evaluate its accuracy in replicating published primary outcomes, benchmarked against direct use of the Web site-based ChatGPT 4o.

Results: The multiagent framework successfully replicated 53.3% (95% CI, 40.7 to 66.0) of primary outcomes, significantly outperforming ChatGPT 4o (35.0% [95% CI, 22.9 to 47.1]; P = .04). The agent framework's failures were predominantly due to data transformation issues (46.4%) and analysis code errors (21.4%). In contrast, ChatGPT 4o failures largely stemmed from incorrect statistical method application (38.4%) and data transformation (45.6%), often attempting to resolve code errors by switching to alternative, incorrect statistical methods. Hallucinations of variables or results were not observed in the multiagent approach.

Conclusion: Our multiagent AI framework demonstrated superior accuracy and robustness in automating biomedical data analysis compared with a generalized LLM.

Purpose: Electronic patient portals can promote patient-centered care, but determinants of engagement remain underexplored in oncology. This study examines sociodemographic and clinical factors associated with engagement with four portal features, including invitations to complete patient-reported outcome (PRO) measures before appointments.

Methods: Secondary analysis of the Northwestern University IMproving the Management of symPtoms during and following Cancer Treatment study, a stepped-wedge cluster randomized trial to promote symptom management using PROs in adult oncology care was performed. For each enrolled participant, we examined portal usage across 1 year.

Results: A total of 3,457 patients were enrolled between April 2020 and April 2023 from 30 Northwestern Medicine ambulatory oncology clinics. Patients were 65% female, 85% White, and 85% non-Hispanic/Latino, with a mean age of 60.8 years. Cancer diagnoses were 30% breast, 12% lymphoma, and all other types accounted for <10% of the sample. Patients accessed laboratory results most frequently (median 23 days in the year), followed by messaging (median 11 days) and physician notes (median 2 days). A total of 62.6% of patients completed at least one invited PRO. Controlling for sociodemographic factors, patient characteristics that were associated with greater engagement across three or more features included more oncology appointments, high health literacy, high anxiety, one or more severe physical symptoms, and high shared decision making with their health care team. Black race, Hispanic/Latino ethnicity, and Medicaid insurance were associated with lower portal engagement. Patients who used any other portal features were more likely to complete PROs. In contrast to other portal features, patients with at least one severe physical symptom were less likely to complete PROs (incidence rate ratio, 0.87 [95% CI, 0.81 to 0.93]; P < .001).

Conclusion: Portal use among patients with cancer varies by sociodemographic and clinical characteristics. Findings suggest a need for targeted interventions to promote equitable use among under-represented groups and promote portal-based PRO completion for patients with higher symptom burden.

Purpose: The objective of this study was to develop a flexible risk stratification strategy for AML that is specific for venetoclax plus azacitidine (ven/aza), addresses real-world data (RWD) issues, and is also adaptable to different use cases.

Methods: A series of tunable risk models (RMs) were generated from a dynamic counterfactual machine learning (ML) strategy. These used a range of features from diagnostic AML samples and were tested using objective metrics on a single-institution cohort of 316 newly diagnosed patients treated with ven/aza. RM performance was tested using various model assumptions, data elements, and end points and with applications to an external AML real-world cohort (RWC).

Results: Favorable, intermediate, and adverse risk groups were identified in a series of ML-based RMs using different assumptions, for genetic-only or genetic-plus-phenotypic data elements and with overall survival and event-free survival as end points. Most RMs demonstrated equitable patient distribution (approximately 20%-40% in each risk group), significant separation between risk strata (log-rank-based P values <0.001), and predictability computed by time-dependent survival AUC values of 0.60-0.70. Similar performance was observed when the proposed RM strategy was adapted and compared with the European Leukemia Net 2022 using the external RWC.

Conclusion: The proposed ML strategy addresses a variety of RWD considerations and is readily tunable through coding and parameter updates for different contexts and use case needs. This strategy represents a novel approach to developing more effective RMs for AML and possibly other diseases.