JCO Clinical Cancer Informatics最新文献

Advancements in oncology are accelerating in the fields of artificial intelligence (AI) and machine learning. The complexity and multidisciplinary nature of oncology necessitate a cautious approach to evaluating AI models. The surge in development of AI tools highlights a need for organized evaluation methods. Currently, widely accepted guidelines are aimed at developers and do not provide necessary technical background for clinicians. Additionally, published guides introducing clinicians to AI in medicine often lack user-friendly evaluation tools or lack specificity to oncology. This paper provides background on model development and proposes a yes/no checklist and questionnaire designed to help oncologists effectively assess AI models. The yes/no checklist is intended to be used as a more efficient scan of whether the model conforms to published best standards. The open-ended questionnaire is intended for a more in-depth survey. The checklist and the questionnaire were developed by clinical and AI researchers. Initial discussions identified broad domains, gradually narrowing to model development points relevant to clinical practice. The development process included two literature searches to align with current best practices. Insights from 24 articles were integrated to refine the questionnaire and the checklist. The developed tools are intended for use by clinicians in the field of oncology looking to evaluate AI models. Cases of four AI applications in oncology are analyzed, demonstrating utility in real-world scenarios and enhancing case-based learning for clinicians. These tools highlight the interdisciplinary nature of effective AI integration in oncology.

Purpose: To evaluate a generative artificial intelligence (GAI) framework for creating readable lay abstracts and summaries (LASs) of urologic oncology research, while maintaining accuracy, completeness, and clarity, for the purpose of assessing their comprehension and perception among patients and caregivers.

Methods: Forty original abstracts (OAs) on prostate, bladder, kidney, and testis cancers from leading journals were selected. LASs were generated using a free GAI tool, with three versions per abstract for consistency. Readability was compared with OAs using validated metrics. Two independent reviewers assessed accuracy, completeness, and clarity and identified AI hallucinations. A pilot study was conducted with 277 patients and caregivers randomly assigned to receive either OAs or LASs and complete comprehension and perception assessments.

Results: Mean GAI-generated LAS generation time was <10 seconds. Across 600 sections generated, readability and quality metrics were consistent (P > .05). Quality scores ranged from 85% to 100%, with hallucinations in 1% of sections. The best test showed significantly better readability (68.9 v 25.3; P < .001), grade level, and text metrics compared with the OA. Methods sections had slightly lower accuracy (85% v 100%; P = .03) and trifecta achievement (82.5% v 100%; P = .01), but other sections retained high quality (≥92.5%; P > .05). GAI-generated LAS recipients scored significantly better in comprehension and most perception-based questions (P < .001) with LAS being the only consistently significant predictor (P < .001).

Conclusion: GAI-generated LASs for urologic oncology research are highly readable and generally preserve the quality of the OAs. Patients and caregivers demonstrated improved comprehension and more favorable perceptions of LASs compared with OAs. Human oversight remains essential to ensure the accurate, complete, and clear representations of the original research.

Purpose: Cancer and its treatment cause symptoms. In this study, we aimed to develop machine learning (ML) systems that predict future symptom deterioration among people receiving treatment for cancer and then validate the systems in a simulated deployment across an entire health care system.

Methods: We trained and tested ML systems that predict a deterioration in nine patient-reported symptoms within 30 days after treatments for aerodigestive cancers, using internal electronic health record (EHR) data at Princess Margaret Cancer Centre (3,229 patients; 20,267 treatments). The primary performance metric was the area under the receiver operating characteristic curve (AUROC). The best-performing systems in the held-out internal test set were then externally validated across 82 cancer centers in Ontario (12,079 patients; 77,003 treatments) by adapting techniques from meta-analysis.

Results: The best ML systems predicted symptom deterioration with AUROCs ranging from 0.66 (95% CI, 0.63 to 0.69) for dyspnea to 0.73 (95% CI, 0.71 to 0.75) for drowsiness in the internal test cohort. Treatments flagged as high-risk were significantly associated with future symptom deterioration (odds ratios [ORs], 2.53-6.56; all P < .001) and emergency department visits for dyspnea (OR, 1.85; P = .008), depression (OR, 1.84; P = .04), and anxiety (OR, 2.66; P < .001). In the external validation cohort, the AUROCs for different symptoms meta-analyzed across centers ranged from 0.67 (95% CI, 0.66 to 0.68) to 0.73 (95% CI, 0.72 to 0.74). Performance across centers displayed significant heterogeneity for six of nine symptoms (I², 46.4%-66.9%; P = .004 for dyspnea, P < .001 for the rest).

Conclusion: ML can predict future symptoms among people with cancer from routine EHR data, which could guide personalized interventions. Heterogeneous performance across centers must be considered when systems are deployed across a health care system.

Purpose: Patients receiving systemic therapy (ST) for non-small cell lung cancer (NSCLC) experience toxicities that negatively affect patient outcomes. This study aimed to test an approach for prospectively collecting patient-reported outcome (PRO) data, wearable sensor data (WSD), and clinical data, and develop a machine learning (ML) algorithm to predict health care utilization, specifically urgent care (UC) visits.

Materials and methods: Patients with NSCLC completed the PROMIS-57 PRO quality-of-life measure and wore a Fitbit to monitor patient-generated health data from ST initiation through day 60. Demographic and clinical data were abstracted from the medical record. ML explainable models on the basis of Bayesian Networks (BNs) were used to develop predictive models for UC visits.

Results: Patients in the training data set (N = 58) were age 69 years on average (range, 35-89) and mostly female (57%), White (88%), and non-Hispanic (95%) patients with adenocarcinoma (69%). Initial BN models trained on demographic and clinical data demonstrated moderate predictive accuracy on cross-validation for UC visits before ST (AUC, 0.72 [95% CI, 0.57 to 0.80]) and during ST (AUC, 0.81 [95% CI, 0.63 to 0.89]). Incorporating PRO and WSD during ST yielded enhanced models with significantly improved performance (final AUC, 0.86 [95% CI, 0.76 to 0.95]) via DeLong test (P < .001).

Conclusion: Multidimensional data sources, including demographic, clinical, PRO, and WSD, can enhance ML predictive models to elucidate complex, interactive factors influencing health care utilization during the first 60 days of ST. Use of explainable ML to predict and prevent treatment toxicities and health care utilization could improve patient outcomes and enhance the quality of cancer care delivery.

Purpose: Ipsilateral breast tumor recurrence (IBTR) remains a critical concern for patients undergoing breast-conserving surgery (BCS). Reliable risk estimation tools for IBTR risk can support personalized surgical and adjuvant treatment decisions, especially in the era of evolving systemic therapies. We aimed to develop and validate models to estimate IBTR risk.

Patients and methods: This multicenter retrospective cohort study included 8,938 women who underwent partial mastectomy for invasive breast cancer between 2008 and 2017. Prediction models were developed using Cox proportional hazards regression and validated via bootstrap resampling. Model performance was assessed using Harrell's C-index, Brier scores, calibration plots, and goodness-of-fit tests.

Results: During a median follow-up of 9.0 years (IQR, 6.6-10.9), IBTR occurred in 320 patients (3.6%). The initial model, based on variables from Sanghani et al, achieved a Harrell's C-index of 0.74. Incorporating hormonal receptor status, human epidermal growth factor receptor 2 status, radiotherapy, and targeted therapy as predictors reduced the C-index to 0.65, despite their clinical relevance. Importantly, the inclusion of these factors improved calibration, demonstrating better alignment between predicted and observed IBTR probabilities. Although the hazard ratios (HRs) for radiotherapy aligned with the Early Breast Cancer Trialists' Collaborative Group meta-analyses (MA), those for chemotherapy and endocrine therapy showed slight differences. Therefore, HRs from the MA were used to represent treatment effects in our model.

Conclusion: We have developed and internally validated a new risk estimation model for IBTR using Cox regression and bootstrap methods. A Web-based risk estimation tool is now available to facilitate individualized risk assessment and treatment planning.

Purpose: Bladder cancer (BLCA) ranks as the tenth most common malignancy worldwide, with rising incidence and mortality rates. Owing to its molecular and clinical heterogeneity, BLCA is associated with high rates of recurrence and metastasis after surgery, contributing to a poor 5-year survival rate. There is a pressing need for highly sensitive and specific molecular biomarkers to enable early identification of high-risk patients, guide clinical management, and improve patient outcomes. This study aimed to develop a prognostic model on the basis of aging-related genes (ARGs) to evaluate survival outcomes and immunotherapy responsiveness in patients with BLCA, and to further explore its relevance to the tumor immune microenvironment and drug sensitivity.

Materials and methods: Transcriptomic and clinical data from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus were used to construct a 12-gene ARG-based prognostic signature through LASSO and Cox regression analyses. Patients were stratified into high-risk and low-risk groups according to the median risk score. Kaplan-Meier survival curves, receiver operating characteristic analyses, and nomograms were used to assess the predictive value of the model. Univariate and multivariate Cox regression analyses were conducted to determine its prognostic independence.

Results: Twelve ARGs were identified. Patients in the low-risk group exhibited significantly better overall survival (P < .0001). In the TCGA cohort, the model yielded AUC values ranging from 0.772 to 0.794 across 1-5 years. Cox regression confirmed the ARG score as an independent prognostic indicator. External validation using the GSE32894 data set supported its clinical reliability. The ARG signature was also associated with immune cell infiltration and predicted chemosensitivity.

Conclusion: The ARG-based risk score independently predicts clinical prognosis in BLCA and correlates with immune microenvironment characteristics, offering potential value in guiding personalized treatment strategies.

Purpose: This article explored how suitable the minimal Common Oncology Data Elements (mCODE) standard is for the real-world evidence research of cancer patient characterization, disease characterization, treatment patterns, and treatment outcomes.

Methods: We identified research questions for each category, broke them down to clinical information elements, and mapped them to the mCODE model. Gaps were further categorized as model deficiencies, clarifying when the mCODE element availability was explicitly specified as an element, derived through external calculation, or implied as part of its support for Fast Healthcare Interoperability Resources.

Results: In our study, 20 research questions were categorized in the following areas: patient characteristics, disease characteristics, treatment patterns, and health outcomes. The mCODE model fully supports patient characterization but shows significant gaps in disease characteristics, treatment patterns, and health outcomes, particularly in areas like treatment regimens and therapy outcomes. Our analysis underscores the need to enhance the mCODE model to better support observational research.

Conclusion: We conclude that mCODE is partially suitable for observational research. Although mCODE shows promise for research purposes in patient and disease characterization, it currently lacks data elements needed to fully support identifying treatment patterns and health outcomes essential for comprehensive observational real-world evidence research.

Purpose: This Special Article provides a comprehensive review and expert commentary on the prospective clinical implementation of artificial intelligence (AI) in the detection of prostate cancer from digital prostate biopsies, as presented in the original research by Flach et al. It contextualizes the study within broader developments in digital pathology and AI, addressing barriers to adoption and the implications for diagnostic workflows and pathology practice.

Design: Drawing on insights from the CONFIDENT-P trial and the author's own experience with digital pathology and AI-assisted workflows, this article critically examines the clinical, regulatory, economic, and operational dimensions of implementing AI in diagnostic pathology. The focus centers on real-world deployment, particularly the integration of Paige Prostate Detect AI (PPD-AI) and its influence on immunohistochemistry (IHC) utilization.

Results: The commentary highlights the trial's prospective design as a significant advancement in AI validation. Key findings include a reduction in IHC use, high diagnostic performance of PPD-AI, and improved diagnostic confidence among AI-assisted pathologists. However, variability in IHC practices across institutions, limitations in AI generalizability, and the need for system integration remain major challenges. The article also addresses practical issues such as automation bias, model drift, and lack of interoperability between viewers and laboratory information systems.

Conclusion: The adoption of AI in digital pathology is accelerating but requires thoughtful integration into clinical workflows. Although prostate biopsies represent an ideal entry point, broader success will depend on regulatory alignment, workforce training, infrastructure readiness, and data governance. This commentary underscores the importance of clinician-AI synergy and provides practical guidance for laboratories navigating the transition from pilot implementations to scalable clinical use.