Artificial Intelligence in Oncology: Current Landscape, Challenges, and Future Directions

May combine two claims.

A ClinicalTrials.gov search for cancer and artificial intelligence at the time of writing found 526 studies, 81 of which were interventional and actively recruiting.

“A search of ClinicalTrials.gov of”

May combine two claims.

The digitalization of oncology, including electronic medical records, radiology and pathology imaging, and routine genomic profiling, is advancing but is uneven across data types and clinical sites.

“This includes the storage of patient data in electronic medical record (EMR) systems, the digitization of radiology and pathology imaging (3), and the increasing adoption of routine genomic profiling (4).”

• cites [3] Niazi, 2019, Digital pathology and artificial intelligence, Lancet Oncol sourceNot checked yet

  Not checked in the free run.

• cites [4] AACR Project GENIE Consortium, 2017, AACR Project GENIE: powering precision medicine through an international consortium, Cancer Discov sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

Breast, prostate, lung, and colorectal cancers accounted for 50% of all new cancer cases in 2023.

“the review focuses on the four most common cancer types: breast, prostate, lung, and colorectal, accounting for 50% of all new cancer cases in 2023 (7).”

• cites [7] Surveillance, Epidemiology, and End Results (SEER): cancer stat facts: common cancer sitesNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

AI is being applied with increasing frequency across all areas of oncology.

“Artificial intelligence (AI) is increasingly being applied to all aspects of oncology.”

May combine two claims.

Deep learning advances over the past decade have made it possible to learn complex patterns directly from real-world data, and this has been a main driver of AI progress in healthcare and elsewhere.

“In particular, advances in deep learning (1) over the past decade have made it feasible to learn complex patterns directly from real-world data, making it the core driver of AI advances in and out of healthcare (2).”

• cites [1] LeCun, 2015, Deep learning, Nature sourceNot checked yet

  Not checked in the free run.

• cites [2] Topol, 2019, High-performance medicine: the convergence of human and artificial intelligence, Nat Med sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Multiple commercial AI devices for breast cancer detection in mammography have received FDA clearance and/or EU certification to aid radiologists.

“In the case of mammography, there are now multiple commercial AI devices cleared by the U.S. Federal Drug Administration (FDA) and/or the European Union (EU) that are designed to aid radiologists in breast cancer detection.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Multiple FDA-cleared and/or EU-certified commercial AI devices exist in cancer detection, and multiple clinical trials have been reported.

“Although not yet uniformly adopted across imaging providers or cancer hospitals, there are multiple FDAcleared and/or EU-certified commercial devices in the cancer detection domain and multiple reported clinical trials.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Multiple clinical trials are evaluating the real-world effectiveness of AI systems for breast cancer detection in mammography.

“There are also multiple ongoing clinical trials evaluating the real-world effectiveness of such systems.”

May combine two claims.

The Mirai system predicts 5-year breast cancer risk directly from mammograms and has been retrospectively evaluated across multiple hospitals.

“the Mirai system developed by Yala and colleagues ( 12) is designed to predict 5-year cancer risk directly from mammograms and has been retrospectively evaluated across multiple hospitals.”

• cites [12] Yala, 2022, Multi-institutional validation of a mammography-based breast cancer risk model, J Clin Oncol sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Increasing availability of detailed longitudinal cancer patient data can be used to train AI models, and real-time access to such data can enable individualized clinically relevant predictions for precision oncology.

“These data can be used to build and train AI models, and critically, the real-time availability of these data can enable individualized, clinically relevant AI predictions to further the goal of precision oncology (5).”

• cites [5] Garraway, 2013, Precision oncology: an overview, J Clin Oncol sourceNot checked yet

  Not checked in the free run.

May combine two claims.

AI applications for breast, prostate, lung, and colorectal cancers are a good representation of the current state of the field because their prevalence supports large-scale data collection and encourages clinical application.

“AI applications for these cancer types serve as a good representation of the current state of the field, especially as their prevalence facilitates large-scale data collection and encourages clinical application.”

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing).

A longer-term prospective trial is assessing the link between colonoscopy CADe systems and colorectal cancer incidence and mortality after 10-year follow-up.

“a longer-term, prospective clinical trial is now seeking to assess the connection between colonoscopy CADe systems and incidence and mortality from colorectal cancer after a 10-year follow-up period (25).”

• cites [25] Mori, 2022, Clinical trial designs for artificial intelligence in gastrointestinal endoscopy, Lancet Gastroenterol Hepatol sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Groups have reported deep learning for breast cancer diagnosis and subtyping from H&E slides and for quantifying features such as mitotic cell counting.

“groups have reported on the use of deep learning for breast cancer diagnosis and subtyping from H&E slides (49)(50)(51), and quantifying other clinically relevant features such as mitotic cell counting (52,53).”

• cites [49] Sui, 2021, A pyramid architecture-based deep learning framework for breast cancer detection, BioMed Res Int sourceNot checked yet

  Not checked in the free run.

• cites [50] Sandbank, 2022, Validation and real-world clinical application of an artificial intelligence algorithm for breast cancer detection in biopsies, NPJ Breast Cancer sourceNot checked yet

  Not checked in the free run.

• cites [51] Couture, 2018, Image analysis with deep learning to predict breast cancer grade, ER status, histologic subtype, and intrinsic subtype, NPJ Breast Cancer sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

CAMELYON16/17 data have become a standard benchmark for deep learning models in pathology.

“The CAMELYON16/17 data have subsequently become a standard benchmark for deep learning models in pathology.”

May combine two claims.

In a retrospective study, pathologists using Paige Prostate Alpha had improved sensitivity for prostate cancer diagnosis, reduced review time, and no statistically significant change in specificity.

“In a retrospective study, pathologists demonstrated improved sensitivity in diagnosing prostate cancer when using the AI platform, along with a reduction in review time and no statistically significant change in specificity.”

• cites [60] Raciti, 2020, Novel artificial intelligence system increases the detection of prostate cancer in whole slide images of core needle biopsies, Mod Pathol sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Since 2013, the U.S. Preventive Services Task Force has recommended low-dose CT screening for lung cancer in high-risk populations.

“Since 2013, the U.S. Preventive Services Task Force has recommended low-dose computed tomography to screen for lung cancer in high-risk populations (26).”

• cites [26] US Preventive Services Task Force, 2021, Screening for lung cancer: US Preventive Services Task Force Recommendation Statement, JAMA sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Paige has received FDA Breakthrough Device Designation for Paige Lymph Node, providing a path for accelerated FDA clearance.

“The AI company Paige has also recently been granted FDA Breakthrough Device Designation status for Paige Lymph Node, providing a path for accelerated FDA clearance (48).”

• cites [48] U.S. FDA grants Paige breakthrough device designation for cancer detection in breast lymph nodesNot checked yet

  Not checked in the free run.

May combine two claims.

Many AI efforts in prostate cancer aim to classify clinically significant disease and quantify features such as prostate gland segmentation and volume.

“There have been many AI efforts in this domain, including those aimed at directly classifying clinically significant prostate cancer and approaches to quantify clinically relevant features such as segmenting and measuring the volume of the prostate gland (35).”

• cites [35] Li, 2022, Machine learning in prostate MRI for prostate cancer: current status and future opportunities, Diagnostics sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Paige Prostate Alpha was validated in an external cohort and cleared by the FDA.

“The platform was subsequently validated in an external cohort (61) and eventually cleared by the FDA (DEN200080, RRID:SCR_012945).”

• cites [61] da Silva, 2021, Independent real-world application of a clinical-grade automated prostate cancer detection system, J Pathol sourceNot checked yet

  Not checked in the free run.

• cites DEN200080Not checked yet

  Not checked in the free run.

May combine two claims.

Google developed LYNA and reported that pathologists using it detected significantly more micrometastases with shorter review time.

“Google developed LYmph Node Assistant (LYNA; ref. 46) and reported that pathologists who used the system were able to detect significantly more micrometastases with shorter review time.”

• cites [46] Steiner, 2018, Impact of deep learning assistance on the histopathologic review of lymph nodes for metastatic breast cancer, Am J Surg Pathol sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Treatment decisions in breast cancer often depend on detecting sentinel lymph node metastases, but pathology review is labor intensive and has significant interreader variability.

“Treatment decisions in breast cancer frequently hinge on the detection of such lesions (44), but pathology review remains labor intensive and subject to significant interreader variability (arXiv:1703.02442).”

• cites [44] Giuliano, 2018, Eighth edition of the AJCC cancer staging manual: breast cancer, Ann Surg Oncol sourceNot checked yet

  Not checked in the free run.

• cites arXiv:1703.02442Not checked yet

  Not checked in the free run.

May combine two claims.

Sandbank and colleagues presented a deep learning model that classifies invasive and noninvasive breast cancer subtypes and generates 51 predicted clinical and morphologic features.

“Sandbank and colleagues (50) have recently presented a deep learning model for classifying invasive and noninvasive breast cancer subtypes and generating 51 predicted clinical and morphologic features.”

• cites [50] Sandbank, 2022, Validation and real-world clinical application of an artificial intelligence algorithm for breast cancer detection in biopsies, NPJ Breast Cancer sourceNot checked yet

  Not checked in the free run.

May combine two claims.

AI has been applied to histopathologic tasks of cancer subtype, grade, and stage classification, but these applications are generally further from clinical adoption and have fewer FDA/EU clearances than imaging-based detection.

“AI has been applied to each of these tasks, though these applications are generally further from clinical adoption with fewer FDA or EU clearances than the imaging-based detection applications described above.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Ardila and colleagues developed an approach to localize lung nodules and predict malignancy likelihood; the model is now licensed by Aidence.

“Ardila and colleagues (29) developed an approach to localize lung nodules and predict their likelihood of malignancy, where this model is now licensed by the AI company Aidence (30).”

• cites [29] Ardila, 2019, End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography, Nat Med sourceNot checked yet

  Not checked in the free run.

• cites [30] 2022, Aidence and Google Health enter into collaboration to help improve lung cancer screening with AINot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The PI-CAI challenge provides a multicenter, multivendor dataset of 1,500 prostate MRI exams for training AI models.

“the Prostate Imaging-Cancer AI (PI-CAI) challenge enables participants to train on a multicenter, multivendor data set of 1,500 MRI prostate exams (36).”

• cites [36] Sunoqrot, 2022, Artificial intelligence for prostate MRI: open datasets, available applications, and grand challenges, Eur Radiol Exp sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Much of this work has been integrated into Stratipath, a Swedish AI start-up focused on digital pathology.

“Much of this work has now been integrated into Stratipath (59), a Swedishbased AI start-up focused on digital pathology.”

• cites [59] Stratipath BreastNot checked yet

  Not checked in the free run.

May combine two claims.

Radiotherapy is a critical therapeutic approach for multiple cancer types and an intense area of AI research, development, and commercialization.

“Radiotherapy is a critical therapeutic approach for multiple cancer types, and an intense area of AI research, development, and commercialization.”

May combine two claims.

These efforts include histology and CT-based models for survival prediction and correlation of CT imaging features with histology risk factors.

“These efforts include histology (102) and CT-based (103) models for survival prediction, and the correlation of CT imaging features with histology risk factors (104).”

• cites [102] Yu, 2016, Predicting non-small cell lung cancer prognosis by fully automated microscopic pathology image features, Nat Commun sourceNot checked yet

  Not checked in the free run.

• cites [103] Saad, 2023, Predicting benefit from immune checkpoint inhibitors in patients with non-small-cell lung cancer by CT-based ensemble deep learning: a retrospective study, Lancet Digit Health sourceNot checked yet

  Not checked in the free run.

• cites [104] Nam, 2022, Histopathologic basis for a chest CT deep learning survival prediction model in patients with lung adenocarcinoma, Radiology sourceNot checked yet

  Not checked in the free run.

May combine two claims.

AI is used at all steps of radiotherapy, including image segmentation, treatment planning, outcome prediction, quality assurance, and AI-guided radiosurgery.

“This includes the use of AI at all steps of radiotherapy, including image segmentation, treatment planning, outcome prediction, quality assurance, and AIguided radiosurgery (116).”

• cites [116] Li, 2022, Artificial intelligence in radiotherapy, Semin Cancer Biol sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

CE-marked products exist for quantifying PD-L1 expression from IHC slides.

“Additionally, CEmarked products exist for quantifying PD-L1 expression from IHC slides (109-111).”

• cites [109] Aiforia achieves CE-IVD mark for clinical PD-L1 AI model in lung cancer diagnosticsNot checked yet

  Not checked in the free run.

• cites [110] Mindpeak Solutions: plug-and-play AI solution to detect and quantify PD-L1 (SP263) within predefined ROIs in non-small-cell lung carcinoma (NSCLC)Not checked yet

  Not checked in the free run.

• cites [111]Not checked yet

  Not checked in the free run.

May combine two claims.

Several validation studies of histopathology-based prognostic models for colorectal cancer exist, including one showing a prognostic feature previously identified through AI can be learned and used by pathologists.

“There have additionally been several validation studies of histopathology-based prognostic models (95,96), including a study demonstrating that a prognostic feature previously identified through AI analysis can be learned and used by pathologists (95).”

• cites [95] L'Imperio, 2023, Pathologist validation of a machine learning–derived feature for colon cancer risk stratification, JAMA Netw Open sourceNot checked yet

  Not checked in the free run.

• cites [96] Skrede, 2020, Deep learning for prediction of colorectal cancer outcome: a discovery and validation study, Lancet Lond Engl sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Some clinically available devices exist for AI-based treatment selection, but they are generally further from clinical maturity than cancer detection and diagnosis tasks.

“Although there are some clinically available devices for these tasks, they tend to be further from clinical maturity than cancer detection and diagnosis tasks.”

May combine two claims.

Nagpal and colleagues used a deep learning model to predict Gleason scores from TCGA H&E images and reported significantly higher diagnostic accuracy than a cohort of 29 board-certified pathologists.

“Nagpal and colleagues (62) used a deep learning model to predict Gleason scores using H&E images from The Cancer Genome Atlas (TCGA) and reported significantly higher diagnostic accuracy compared with a cohort of 29 board-certified pathologists.”

• cites [62] Nagpal, 2019, Development and validation of a deep learning algorithm for improving Gleason scoring of prostate cancer, NPJ Digit Med sourceNot checked yet

  Not checked in the free run.

May combine two claims.

P-NET is a sparse deep learning model that processes molecular profiling data in a pathway-driven fashion to predict prostate disease states such as metastases, and its prediction scores independently correlate with prognosis.

“P-NET (112) is a sparse deep learning model that processes molecular profiling data in a biologically informed, pathway-driven fashion to predict prostate disease states such as metastases, where the resulting prediction scores were shown to independently correlate with prognosis.”

• cites [112] Elmarakeby, 2021, Biologically informed deep neural network for prostate cancer discovery, Nature sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Significant efforts have been made to use AI for histopathologic diagnosis in lung and colorectal cancers, but no FDA-cleared devices exist for these tasks at the time of writing.

“Significant efforts have also been made in using AI for histopathologic diagnosis in lung and colorectal cancers, though FDA-cleared devices for these tasks do not exist at the time of writing.”

May combine two claims.

That approach used weakly supervised training to predict response from H&E whole slide images and interpretability methods identified features correlating with higher response predictions, such as TILs and necrosis.

“The approach included weakly supervised training to predict response from H&E whole slide images, with subsequent application of interpretability approaches to identify features that correlate with higher response predictions, such as the presence of tumor-infiltrating lymphocytes (TILs) and necrosis.”

• cites [83] Ogier du Terrail, 2023, Federated learning for predicting histological response to neoadjuvant chemotherapy in triple-negative breast cancer, Nat Med sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Large language models such as BERT, GPT, and ChatGPT are rapidly diffusing into industry, and healthcare-specific models are being developed.

“Models such as BERT (80), GPT (81), and ChatGPT (arXiv:2303.08774)are rapidly diffusing into industry, and healthcare-specific models are also being developed (ref.82; arXiv:2106.03598,arXiv:1904.05342).”

• cites [80] Devlin, 2019, Pre-training of deep bidirectional transformers for language understandingNot checked yet

  Not checked in the free run.

• cites [81] Brown, 2020, Language models are few-shot learnersNot checked yet

  Not checked in the free run.

• cites arXiv:2303.08774Not checked yet

  Not checked in the free run.

May combine two claims.

Amgad and colleagues used deep learning to extract human-interpretable features and then generated a risk score for breast cancer prognosis.

“Amgad and colleagues (84) explicitly enforce interpretability by using deep learning to extract human-interpretable features, and subsequently use these features to generate a risk score for breast cancer prognosis.”

• cites [84] Amgad, 2024, A population-level digital histologic biomarker for enhanced prognosis of invasive breast cancer, Nat Med sourceNot checked yet

  Not checked in the free run.

May combine two claims.

In an evaluation of four large language models asked to suggest treatment options for fictional advanced cancer patients, the models produced a wider range of options than human experts but their recommendations often disagreed with expert consensus.

“Although the LLMs generated a wider range of treatment options compared with human experts, their recommendations often deviated from expert consensus.”

• cites [151] Benary, 2023, Leveraging large language models for decision support in personalized oncology, JAMA Netw Open sourceNot checked yet

  Not checked in the free run.

May combine two claims.

CancerVar is an AI platform that uses deep learning to predict oncogenicity of somatic variants using functional and clinical features and may provide a robust alternative to manually curated databases such as CiVIC and OncoKB.

“CancerVar (131) is an AI platform that uses a deep learning framework to predict the oncogenicity of somatic variants using both functional and clinical features and may provide a robust alternative to manually curated databases, such as CiVIC (132) and OncoKB ( 133).”

• cites [131] Li, 2022, CancerVar: an artificial intelligence-empowered platform for clinical interpretation of somatic mutations in cancer, Sci Adv sourceNot checked yet

  Not checked in the free run.

• cites [132] Griffith, 2017, CIViC is a community knowledgebase for expert crowdsourcing the clinical interpretation of variants in cancer, Nat Genet sourceNot checked yet

  Not checked in the free run.

• cites [133] Chakravarty, 2017, OncoKB: a precision oncology knowledge base, JCO Precis OncolNot checked yet

  Not checked in the free run.

May combine two claims.

Personalized cancer vaccines are being tested in a wide variety of clinical indications, and recent trial data indicate their potential role in combination with existing immune-checkpoint blockade.

“personalized cancer vaccines are being tested in a wide variety of clinical indications, and recent trial data indicate their potential role in combination with existing immune-checkpoint blockade (139).”

• cites [139] Khattak, 2023, Abstract CT001: a personalized cancer vaccine, mRNA-4157, combined with pembrolizumab versus pembrolizumab in patients with resected high-risk melanoma: efficacy and safety results from the randomized, open-label Phase 2 mRNA-4157-P201/Keynote-942 trial, Cancer Res sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Applied to 37 cancer types from the Pan-Cancer Analysis of Whole Genomes dataset, Dig identified potential cryptic splice sites, recurrent variants in noncoding regions, and mutations in rarely mutated genes.

“Applied to 37 cancer types from the Pan-Cancer Analysis of Whole Genomes data set (128), Dig identified several potential cryptic splice sites, recurrent variants in noncoding regions, as well as mutations in rarely mutated genes.”

• cites [128] ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Consortium, 2020, Pan-cancer analysis of whole genomes, Nature sourceNot checked yet

  Not checked in the free run.

May combine two claims.

AI in oncology faces significant challenges in obtaining and sharing robust medical data, protecting patient privacy, and ensuring fairness across all patients.

“These challenges include obtaining and sharing robust medical data, risks to patient privacy, and ensuring fairness in AI performance across all patients.”

May combine two claims.

AI applications are advancing in image-guided radiotherapy, adaptive radiotherapy, stereotactic radiosurgery, and surgery.

“AI applications are also now advancing in the fields of image-guided radiotherapy, adaptive radiotherapy, stereotactic radiosurgery (119), and surgery in general, where autocontouring is similarly helpful (120).”

• cites [119] Kawamura, 2024, Revolutionizing radiation therapy: the role of AI in clinical practice, J Radiat Res (Tokyo) sourceNot checked yet

  Not checked in the free run.

• cites [120] Mascagni, 2022, Computer vision in surgery: from potential to clinical value, NPJ Digit Med sourceNot checked yet

  Not checked in the free run.

May combine two claims.

A 2022 meta-analysis of 10 published studies across commercial and academic platforms found that most clinical trial matching applications offered comparable or superior performance to manual screening, with dramatic efficiency improvements.

“A 2022 meta-analysis reported on 10 published studies across multiple commercial and academic platforms and found that most applications offered comparable, if not superior performance to manual screening, with dramatic improvements in efficiency (149).”

• cites [149] Chow, 2023, Use of artificial intelligence for cancer clinical trial enrollment: a systematic review and meta-analysis, J Natl Cancer Inst sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Radici and colleagues evaluated the FDA-cleared and CE-marked Limbus Contour AI platform for autocontouring at a single European tertiary cancer hospital and found it significantly reduced procedure time and interobserver variability.

“A recent study by Radici and colleagues (117) evaluated the FDAcleared and CE-marked Limbus Contour (LC) AI platform for autocontouring within a single European tertiary cancer hospital and found that the technology significantly reduced the procedure time and interobserver variability.”

• cites [117] Radici, 2022, Implementation of a commercial deep learning-based auto segmentation software in radiotherapy: evaluation of effectiveness and impact on workflow, Life Basel SwitzNot checked yet

  Not checked in the free run.

May combine two claims.

Only 6% of patients with cancer enroll in clinical trials, and race-, sex-, and age-based disparities persist.

“only 6% of patients with cancer enroll in clinical trials (144) and there continue to be race-, sex-, and age-based disparities within clinical trials (145, 146).”

• cites [144] Unger, 2021, Nationally representative estimates of the participation of cancer patients in clinical research studies according to the commission on cancer, J Clin Oncol sourceNot checked yet

  Not checked in the free run.

• cites [145] Murthy, 2004, Participation in cancer clini­cal trials: race-, sex-, and age-based disparities, JAMA sourceNot checked yet

  Not checked in the free run.

• cites [146] Kwiatkowski, 2013, Inclusion of minorities and women in cancer clinical trials, a decade later: have we improved?, Cancer sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Transcriptomic-based algorithms for cell-of-origin prediction have been developed, but they require expansion of clinical transcriptome profiling for generalizability assessments.

“Transcriptomic-based algorithms have also been developed (136, 137), although these approaches will require expansion of clinical transcriptome profiling for assessments of generalizability.”

• cites [136] Grewal, 2019, Application of a neural network whole transcriptome-based pan-cancer method for diagnosis of primary and metastatic cancers, JAMA Netw Open sourceNot checked yet

  Not checked in the free run.

• cites [137] Divate, 2022, Deep learning-based pan-cancer classification model reveals tissue-of-origin specific gene expression signatures, Cancers sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

As more AI models reach the clinic, challenges remain in integrating AI applications into clinical workflows and assessing their ultimate impact on patients and healthcare systems.

“As more AI models reach the clinic, there are also challenges in effectively integrating AI applications into clinical workflows and assessing the ultimate impact on patients and healthcare systems.”

May combine two claims.

In the same evaluation, large language models correctly identified several important treatment strategies and proposed some reasonable options that experts had not readily seen.

“Despite these limitations, LLMs were able to correctly identify several important treatment strategies and even suggested some reasonable options that were not readily apparent to experts (151).”

• cites [151] Benary, 2023, Leveraging large language models for decision support in personalized oncology, JAMA Netw Open sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Dig outperformed multiple other methods for driver identification in accuracy and power and ran multiple orders of magnitude faster.

“The method outperformed multiple other methods for driver identification in accuracy and power and ran multiple orders of magnitude faster.”

• cites [127] Sherman, 2022, Genome-wide mapping of somatic mutation rates uncovers drivers of cancer, Nat Biotechnol sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Multiple groups have reported using LLMs to extract structured data from clinical notes and to match patients to clinical trials.

“multiple groups have now reported on the use of LLMs to effectively extract structured data from clinical notes (150) and to match patients to clinical trials (arXiv:2303.16756,arXiv:2306.02077,arXiv:2304.07396, and arXiv:2308.02180).”

• cites [150] Agrawal, Large language models are few-shot clinical information extractors sourceNot checked yet

  Not checked in the free run.

• cites arXiv:2303.16756Not checked yet

  Not checked in the free run.

• cites arXiv:2306.02077Not checked yet

  Not checked in the free run.

May combine two claims.

In cancer, individual sites are limited in patient numbers, patient diversity, and local clinical practices.

“This is particularly important in cancer, where each site is limited in patient size, patient diversity, and its own set of embedded clinical practices.”

May combine two claims.

Interpretation of germline and somatic mutations in cancer remains a challenge, and new AI tools are emerging to automatically annotate and interpret variants.

“The interpretation of germline and somatic mutations observed in cancer continues to be a challenge, and new AI tools are also emerging to automatically annotate and interpret variants.”

May combine two claims.

ChatGPT's first-line treatment recommendations for advanced solid tumors were generally promising, but in some cases it gave outdated information such as investigational drug names.

“Performance was generally promising, though there were cases in which the model arguably provided outdated information, such as investigational drug names (152).”

• cites [152] Schulte, 2023, Capacity of ChatGPT to identify guideline-based treatments for advanced solid tumors, CureusNot checked yet

  Not checked in the free run.

May combine two claims.

Clinician burnout risks driving providers away from clinical practice and impairing access to care.

“that risks driving providers away from clinical practice and impairing access to care.”

May combine two claims.

The clinical trial system has inefficiencies, with 20% of trials prematurely terminated, many due to insufficient accrual.

“The clinical trial system also suffers from overall inefficiencies, where 20% of trials are prematurely terminated, many due to insufficient accrual of patients (147).”

• cites [147] Stensland, 2014, Adult cancer clinical trials that fail to complete: an epidemic?, J Natl Cancer Inst sourceNot checked yet

  Not checked in the free run.

May combine two claims.

MD Anderson Cancer Center is creating a comprehensive AI-based Radiation Planning Assistant for autocontouring and radiotherapy planning, primarily targeting low- and middle-income countries.

“MD Anderson Cancer Center is also creating a comprehensive AI-based Radiation Planning Assistant for autocontouring and radiotherapy planning, and will primarily target low-and middleincome countries, where AI can provide scalable assistance (118).”

• cites [118] MD Anderson Cancer Center: Radiation Planning AssistantNot checked yet

  Not checked in the free run.

May combine two claims.

Genomics-based targeted NGS algorithms for cell-of-origin prediction have been rigorously evaluated and are being deployed and assessed for clinical utility.

“This includes genomics-based algorithms, where targeted NGS algorithms have been rigorously evaluated and are currently being deployed and assessed for clinical utility (134, 135).”

• cites [134] Penson, 2020, Development of genome-derived tumor type prediction to inform clinical cancer care, JAMA Oncol sourceNot checked yet

  Not checked in the free run.

• cites [135] Moon, 2023, Machine learning for genetics-based classification and treatment response prediction in cancer of unknown primary, Nat Med sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim. May combine two claims.

AI models must be tested across multiple independent sites and evaluated for overall accuracy and potential bias.

“Beyond training, AI models must also be tested across multiple independent sites and evaluated for overall accuracy and potential bias (162).”

• cites [162] Karargyris, 2023, Federated benchmarking of medical artificial intelligence with MedPerf, Nat Mach Intell sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

Rich molecular data are increasingly critical for personalized cancer treatment across cancer types.

“Rich molecular data are increasingly critical for personalized cancer treatment across cancer types.”

Classified as probably not a substantial claim. May combine two claims.

Cell-of-origin prediction is a long-standing diagnostic challenge, particularly for cancers of unknown primary where accurate origin information could significantly affect treatment strategy.

“An additional long-standing diagnostic challenge is cell-oforigin prediction, which is particularly relevant for cancers of unknown primary where accurate origin information could significantly affect treatment strategy.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Properly deployed, AI could increase efficiencies and patient diversity within the clinical trial system and assist physicians in matching patients to trials.

“Properly deployed, AI could increase efficiencies and patient diversity within the clinical trial system and assist physicians in matching patients to clinical trials.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Federated learning keeps data private at each institution while jointly updating and sharing machine learning models in a distributed manner.

“The third option is federated learning (FL), where data remain private at each institution, but machine learning models are jointly updated and shared in a distributed fashion (166, 167).”

• cites [166] Rieke, 2020, The future of digital health with federated learning, NPJ Digit Med sourceNot checked yet

  Not checked in the free run.

• cites [167] Saldanha, 2022, Swarm learning for decentralized artificial intelligence in cancer histopathology, Nat Med sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Different data modalities likely contain distinct and complementary information, making multimodal AI an active area of oncology research.

“These different data modalities are likely to contain distinct and complementary information; thus, multimodal AI is an active area of research in oncology (168, 169).”

• cites [168] Boehm, 2022, Harnessing multimodal data integration to advance precision oncology, Nat Rev Cancer sourceNot checked yet

  Not checked in the free run.

• cites [169] Lipkova, 2022, Artificial intelligence for multimodal data integration in oncology, Cancer Cell sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Because of this classification, randomized controlled clinical trials are not typically required by the FDA, though this can depend on intended use.

“Importantly, given this classification, randomized controlled clinical trials are not typically required by the FDA, though this can depend on the device's intended use.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Neither the United States nor the EU currently has AI-specific regulatory approval pathways for AI-based medical tools.

“Neither the United States nor the EU currently has AI-specific regulatory approval pathways for AI-based medical tools.”

May combine two claims.

Clinical trials in AI oncology remain rare, but there is now an uptick in such trials.

“Although clinical trials in AI oncology remain rare, we are now starting to see an uptick in such trials.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Regulation of large language models is evolving and debated, including which contexts and intended uses qualify them as medical devices.

“A particularly evolving and debated topic is the regulation of LLMs, and in which contexts and intended uses do they qualify as medical devices (191,192).”

• cites [191] Meskó, 2023, The imperative for regulatory oversight of large language models (or generative AI) in healthcare, NPJ Digit Med sourceNot checked yet

  Not checked in the free run.

• cites [192] Gilbert, 2023, Large language model AI chatbots require approval as medical devices, Nat Med sourceNot checked yet

  Not checked in the free run.

May combine two claims.

The same dataset and even model architecture can be used to support multiple clinical tasks, emphasizing the need for robust datasets.

“These efforts further emphasize the need for robust data sets, as the same data set and even model architecture can be used to support multiple clinical tasks.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Both the US and EU regulate AI healthcare applications as medical devices and categorize them by risk to individual patients.

“Rather, both regulate AI healthcare applications as medical devices and categorize such devices based on the level of risk posed to individual patients.”

Probably not a substantial claim: may not pass S3 (load-bearing).

In oncology, potential sources of bias span the full patient care and AI life-cycle continuums.

“Both of these considerations certainly apply in oncology, where potential sources of bias span the full patient care and AI life-cycle continuums.”

May combine two claims.

In clinical practice, there are known disparities in access to newer imaging technologies and in patient outcomes for underserved populations, such as Black women.

“In clinical practice, there are known disparities in access to newer imaging technologies and patient outcomes for underserved populations, such as Black women (176-178).”

• cites [176] Christensen, 2023, Relationship between race and access to newer mammographic technology in women with medicare insurance, Radiology sourceNot checked yet

  Not checked in the free run.

• cites [177] Rauscher, 2012, Disparities in screening mammography services by race/ethnicity and health insurance, J Womens Health sourceNot checked yet

  Not checked in the free run.

• cites [178] Martini, 2022, Breast cancer disparities in outcomes; unmasking biological determinants associated with racial and genetic diversity, Clin Exp Metastasis sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Clinical trials remain the gold standard for assessing new clinical interventions.

“An aspect that is closely tied to regulation is the evaluation of devices to ensure safe and effective performance, where clinical trials remain the gold standard in assessing new clinical interventions.”

May combine two claims.

The same study found breast cancer was the most commonly cited cancer type (31%), most commonly associated with radiology, providing evidence of advancing maturity of AI in mammography.

“The same study also found that breast cancer was the most commonly cited cancer type (31%), most commonly associated with radiology, providing further evidence of the advancing maturity of AI in mammography.”

• cites [194] Luchini, 2022, Artificial intelligence in oncology: current applications and future perspectives, Br J Cancer sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable).

Ensuring transparency and reproducibility is critical for independent research groups to review and build on prior scientific progress.

“A critical component of AI evaluation and development more generally is ensuring transparency and reproducibility, allowing independent research groups to review and build on prior scientific progress (196).”

• cites [196] Haibe-Kains, 2020, Transparency and reproducibility in artificial intelligence, Nature sourceNot checked yet

  Not checked in the free run.

May combine two claims.

FDA clearance requirements depend largely on device class and intended use, with classes ranging from I (lowest risk) to III (highest risk).

“The requirements for FDA clearance depend largely on the class and intended use of the device, with device classes ranging from I (lowest risk) to III (highest risk).”

May combine two claims.

Genetic ancestry may be causally related to some genetic traits and diseases, unlike the social construct of race.

“Fairness and diversity have especially important considerations in such applications given that genetic ancestry may be causally related to some genetic traits and diseases (178, 183-186), as opposed to the social construct of race (173).”

• cites [178] Martini, 2022, Breast cancer disparities in outcomes; unmasking biological determinants associated with racial and genetic diversity, Clin Exp Metastasis sourceNot checked yet

  Not checked in the free run.

• cites [183] Zhang, 2019, Characterization of frequently mutated cancer genes in Chinese breast tumors: a comparison of Chinese and TCGA cohorts, Ann Transl Med sourceNot checked yet

  Not checked in the free run.

• cites [184] Mahal, 2020, Racial differences in genomic profiling of prostate cancer, N Engl J Med sourceNot checked yet

  Not checked in the free run.

• cites [185] Gusev, 2016, Atlas of prostate cancer heritability in European and African-American men pinpoints tissue-specific regulation, Nat Commun sourceNot checked yet

  Not checked in the free run.

• cites [186] Nassar, 2022, Ancestry-driven recalibration of tumor mutational burden and disparate clinical outcomes in response to immune checkpoint inhibitors, Cancer Cell sourceNot checked yet

  Not checked in the free run.

• cites [173] Chen, 2023, Algorithmic fairness in artificial intelligence for medicine and healthcare, Nat Biomed Eng sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Progress has narrowed the performance gap between federated learning and centralized learning, and federated learning has been successfully applied to multiple cancer applications.

“Although FL introduces a number of logistical and algorithmic complexities, progress has been made in narrowing the gap in model performance between FL and centralized learning (160), where FL has now been successfully applied to multiple cancer applications.”

• cites [160] Sheller, 2020, Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data, Sci Rep sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Most AI devices, including oncology devices, are Class II, for which the FDA 510(k) pathway applies and requires substantial equivalence to already-cleared tools.

“The vast majority of AI devices thus far, including those in oncology, fall under Class II, for which the FDA 510(k) pathway applies and requires the device to be”

• cites [187] Harvey, 2020, How the FDA regulates AI, Acad Radiol sourceNot checked yet

  Not checked in the free run.

May combine two claims.

These algorithms are largely designed to assist clinicians in making diagnoses and/or quantify image-based properties, but the clinician makes the final call.

“These algorithms are largely designed to assist clinicians in making diagnoses and/or quantify image-based properties (e.g., lesion size), but the clinician makes the final call.”

May combine two claims.

Awareness of AI algorithm fairness is growing, including guidance from regulatory bodies, but much room remains to reduce bias in AI and healthcare.

“Altogether, there is growing awareness of the fairness of AI algorithms, including guidance from regulatory bodies, but there remains much room for improvement in reducing bias in AI and healthcare more broadly.”

May combine two claims.

Deidentification and public release of data, often through consortium models, is an option for data sharing; CAMELYON and BraTS challenges have made large-scale deidentified datasets available, and the Nightingale Open Science Initiative aims to release more.

“The second option is deidentification and public release of data, usually via a consortium model.”

• cites [45] Ehteshami Bejnordi, 2017, Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer, JAMA sourceNot checked yet

  Not checked in the free run.

• cites [165] Mullainathan, 2022, Solving medicine's data bottleneck: Nightingale Open Science, Nat Med sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Luchini and colleagues summarized 71 FDA-cleared devices in oncology up to 2021 and found radiology accounted for 54.9% and pathology 19.7%.

“For example, Luchini and colleagues (194) summarized 71 FDA-cleared devices in oncology (up to 2021) and found that radiology accounted for the majority of approved devices (54.9%), followed by pathology (19.7%).”

• cites [194] Luchini, 2022, Artificial intelligence in oncology: current applications and future perspectives, Br J Cancer sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

TCGA has become a critical multimodal dataset for advancing AI efforts.

“This work leverages the TCGA database, which has become a critical multimodal data set to advance AI efforts.”

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based).

Foundation models trained on vast quantities of data have shown promise in facilitating downstream tasks with less labeled data, a common situation in oncology.

“Such”

• cites arXiv:2108.07258Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In the EU, no single central agency performs device assessments; accredited organizations issue CE marks.

“Within the EU, there is no one central agency that performs device assessments; rather, assessments are handled by accredited organizations, which are empowered to issue CE marks.”

May combine two claims.

There are efforts to generate more diverse public datasets, such as EMBED, and instances where AI performance has generalized well across different patient populations.

“Conversely, there are efforts to generate more diverse public data sets, such as EMBED (180), and instances where AI performance has generalized well across different patient populations (9,10,12).”

• cites [180] Jeong, 2023, The EMory BrEast imaging Dataset (EMBED): a racially diverse, granular dataset of 3.4 million screening and diagnostic mammographic images, Radiol Artif Intell sourceNot checked yet

  Not checked in the free run.

• cites [9] McKinney, 2020, International evaluation of an AI system for breast cancer screening, Nature sourceNot checked yet

  Not checked in the free run.

• cites [10] Lotter, 2021, Robust breast cancer detection in mammography and digital breast tomosynthesis using an annotation-efficient deep learning approach, Nat Med sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Most described AI applications fall under regulatory agencies, depending on regional regulations and intended use.

“Most of the described AI applications fall under the purview of regulatory agencies, depending on regional regulations and the intended use of the device.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

The precise form of AI results requires careful consideration of the clinical end goal and intended use of the device.

“The precise form of the AI results requires careful consideration of the clinical end goal and the intended use of the device.”

Probably not a substantial claim: may not pass S2 (checkable).

Transparency and reproducibility are especially important for assessing potential biases in datasets and generalizability across clinical settings.

“This is especially important for assessing potential biases in data sets and generalizability across different clinical settings.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Although AI will ideally improve healthcare outcomes, its biggest impact may ultimately be in improving resource allocation and access to quality care.

“Critically, although AI will ideally improve healthcare outcomes, its biggest impact may ultimately be in improving resource allocation and access to quality care.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

The healthcare value of AI deployment needs rigorous assessment and aligned payment structures.

“Second, the healthcare value of AI deployment needs to be rigorously assessed and combined with aligned payment structures.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Integration into clinical systems can be especially challenging for proprietary systems with limited customization.

“Such workflows require integration into clinical systems, which can be especially challenging for proprietary systems with limited customization.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

This assessment will require greater investment in real-world studies and ideally randomized clinical trials, with emphasis on evaluating AI generalizability across patient populations and clinical sites.

“This will require a greater investment in real-world studies and ideally randomized clinical trials, with an emphasis on critically evaluating AI generalizability across patient populations and clinical sites.”

Probably not a substantial claim: may not pass S2 (checkable).

The future of AI oncology will not be driven by technology innovation alone.

“Nonetheless, the future of AI oncology will not be driven by technology innovation alone.”

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims. The quote only partly states this claim.

Several fundamental requirements will drive the future of AI in oncology, including objective assessment of impact on patient outcomes, rigorous assessment of healthcare value, standardized clinical integration, and continuous monitoring of impact and safety.

“Rather, we see several fundamental requirements that will drive the future of AI in oncology (Fig. 4).”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

There is a strong need for objective assessment of AI's impact on patient outcomes using cancer-specific, clinically meaningful metrics.

“First, there is a strong need for an objective assessment of AI's impact on patient outcomes, defined in cancer-specific, clinically meaningful metrics.”

Probably not a substantial claim: may not pass S2 (checkable).

Stakeholders must adopt objective processes for continuously monitoring AI impact and measuring patient safety.

“Finally, stakeholders must adopt objective processes for continuously monitoring AI impact and measuring patient safety.”

Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

Realizing these benefits may require structural changes and new clinical workflows.

“Realizing these benefits may require structural changes and new clinical workflows.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

The greatest strength of data-driven deep learning approaches is their ability to continually improve over time.

“Arguably, the greatest strength of data-driven, deep learning approaches is the ability to continually improve over time.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

Healthcare value is defined as the health outcome achieved per dollar spent and is a critical metric for informed healthcare investments.

“Healthcare value is defined as the health outcome achieved per dollar spent (214) and is a critical metric for making informed healthcare investments.”

• cites [214] Porter, 2010, What is value in health care?, N Engl J Med sourceNot checked yet

  Not checked in the free run.