Across 157 trials, researchers are testing artificial intelligence (AI) in oncology primarily as a reader and communicator, not as a treatment decision-maker. In our analysis of clinical trials that started on or after January 1, 2026, twelve trials place AI within treatment selection or planning workflows, one uses an AI-derived therapeutic approach as the investigational product, and none of the five hematology trials identified in this 2026-start search use AI for diagnosis or treatment selection.
Key Takeaways
- AI is being tested across the cancer care pathway, but its clinical roles remain unevenly distributed. Most registered trials focus on detection, diagnosis, prediction, communication, and supportive care rather than autonomous treatment decisions.
- The gap between what AI can analyze and what it is being asked to decide remains substantial. Only a small proportion of the 2026-start trials identified in this analysis place AI directly within treatment selection or treatment-planning workflows.
- The strongest evidence for AI in oncology remains task-specific. Imaging, pathology, structured information retrieval, and other well-defined tasks are more readily evaluated than complex clinical decisions requiring integration of multiple data sources.
- Clinical validation—not technical capability—is the critical next step. The growing number of oncology AI studies shows where the technology is being tested, but trial registration or high performance on a defined task does not establish clinical benefit, safety, or readiness for routine care.
Our Review – Scope and Methods
To better understand where and how artificial intelligence is being used in oncology, the editorial teams of Onco’Zine and ADC Review | Journal of Antibody-Drug Conjugates reviewed records in PubMed and ClinicalTrials.gov. The review was designed as an editorial analysis of the current evidence and registered clinical-trial landscape, not a systematic review or meta-analysis.
We examined PubMed to identify published research describing clinical and translational applications of AI in cancer. [1][2]We also searched ClinicalTrials.gov to identify registered oncology studies in which AI is evaluated as part of diagnosis, prediction, treatment planning or selection, procedures, patient communication and supportive care, workflow, or clinical research.[1]
For the ClinicalTrials.gov analysis, we searched on September 28, 2026, using the condition term ‘Oncology’ and the search term ‘artificial intelligence,’ restricted to studies with a start date on or after January 1, 2026. The search returned 195 records. We excluded records that concerned non-oncologic conditions captured by the search terms or that included an oncology study with no identifiable AI component in its registered title, conditions, or interventions. This left 157 studies for classification.
We categorized studies by the function AI performs, based on information in the registered title, conditions, and intervention descriptions. We did not review protocol documents for classification. Because some studies use AI for more than one purpose, individual studies could reasonably fit more than one functional category; the analysis therefore reflects the principal AI application identified in the registry record.
This approach did not measure whether AI works, but rather where researchers are currently putting AI to work in oncology and what role they are asking it to perform.[3][4]
Six findings
Diagnosis and patient communication account for about 60% of the trial set. These are the two largest categories by a wide margin, and they sit at opposite ends of the care pathway. Diagnostic applications are concentrated in imaging and pathology, including chest CT and pulmonary nodules, mammography and breast ultrasound, dermoscopy, endoscopic imaging, frozen-section assessment and whole-slide pathology. Patient-facing applications include chatbots, symptom-tracking applications, education modules, and prehabilitation or rehabilitation support. Neither group necessarily requires AI to select a treatment.
Twelve trials place AI output within treatment selection or treatment-planning workflows. These include AI-generated contours for prostate brachytherapy (NCT06964412), adaptive radiotherapy workflow in glioma (NCT06492486), AI-informed biopsy decisions in indeterminate prostate MRI (NCT07231627), risk-adapted surveillance after hepatocellular carcinoma resection (NCT07732504), AI-assisted staging and treatment decisions in HCC (NCT07538882), and CDK4/6 inhibitor rechallenge in advanced breast cancer (NCT07227233). Radiotherapy planning is among the more defined applications in this group, involving a repeatable task for which established clinical planning workflows provide a reference for comparison.
One trial tests a therapeutic product in which AI is part of the product-development process. PAN-NeoVax (NCT07560943), a pan-cancer neoantigen mRNA vaccine being studied at West China Hospital, is a dose-escalation safety and tolerability study in patients with advanced solid tumors. In this search, it is the only trial in which the registered AI contribution appears to be the investigational therapeutic product itself rather than a diagnostic, predictive or supportive function.
Concordance with human experts is being measured directly. At least four trials compare AI-generated output with multidisciplinary team or specialist decisions: LLMs benchmarked against tumor boards across 12 tumor types (NCT07739121), multimodal AI compared with MDT decisions in gastrointestinal oncology (NCT07746830), an LLM compared with MDT recommendations in rectal cancer (NCT07595107), and AI-augmented pathology review compared with independent pathologist assessment under different orders of presentation (NCT07741058). These studies address whether AI recommendations agree with established clinical decision-making. The pathology study also examines whether the order in which AI is consulted changes the reader’s conclusion.
Hematologic malignancy is largely absent from the diagnostic and therapeutic categories in this search. Five trials involve hematologic conditions, and all fall outside diagnosis and treatment selection: a generative AI and propensity-score model used to construct a comparator for a phase II venetoclax-decitabine study in older adults with AML (NCT07420790); prediction of non-adherence to oral therapy in lymphoma (NCT07546188); LLM-generated plain-language patient synopses in hematology and oncology (NCT07519811); and two supportive-care studies that include hematologic patients among mixed populations (NCT07772596 and NCT07553572). None of the five hematology trials identified in this 2026-start search used AI for hematologic diagnosis or treatment selection. This finding is specific to the registry search and does not represent the broader literature on AI in flow cytometry, marrow morphology or measurable residual disease.
AI is being placed upstream of antibody-drug conjugate (ADC) eligibility and selection. Three trials place an AI model at the point where an ADC-related treatment decision is made. NCT07782086 studies trastuzumab deruxtecan in metastatic breast cancer with HER2-ultralow status determined by AI assessment—a range of immunohistochemical expression in which classification can affect treatment eligibility. NCT07528716 uses an AI model to select among neoadjuvant anti-HER2 regimens, two of which contain the ADC SHR-A1811. NCT07454941 applies AI to HER2 assessment in urothelial carcinoma, where HER2-directed ADCs are being investigated and assessment is less standardized than in some established breast cancer applications. These studies place AI at a clinically consequential point: the assessment of a biomarker or treatment pathway that can determine whether an ADC is considered.
What may account for the distribution?
The concentration in imaging and endoscopy is consistent with the characteristics of these data: large volumes of digital images can be linked to defined clinical or pathological outcomes, creating datasets suitable for model development and validation. Gastrointestinal, thoracic, breast, prostate and skin cancers therefore appear prominently in this group.
The hematology gap presents a different data environment. Hematologic diagnosis can integrate morphology, immunophenotype, cytogenetics, and molecular findings generated across different platforms and institutions. That heterogeneity may present additional challenges for developing and validating models that can be applied consistently across settings. In this search, the hematology trial that uses generative AI applies it substantively to constructing a comparator rather than interpreting a patient’s marrow or other diagnostic specimen.
The patient-facing category is notably large. Forty-four trials test whether AI can perform or support parts of the communication and supportive-care workload in oncology, including education, symptom reporting, follow-up contact, and psychological support. Their endpoints include symptom burden, anxiety, comprehension, and quality of life rather than diagnostic accuracy.
Our conclusion
This nine-month review shows that AI’s role in oncology is already shifting—from tools that primarily read, predict, and communicate to applications that increasingly influence clinical decisions and treatment pathways. The picture is still early, but the direction is becoming clearer. The next nine-month analysis will be critical, as we believe it will show whether these early signals represent the beginning of a fundamental shift in how AI is integrated into cancer care—or simply the first stage of a longer evolution.
References
[1] Saady M, Eissa M, Yacoub AS, Hamed AB, Azzazy HME. Implementation of artificial intelligence approaches in oncology clinical trials: A systematic review. Artif Intell Med. 2025 Mar;161:103066. doi: 10.1016/j.artmed.2025.103066. Epub 2025 Jan 18. PMID: 39837136.
[2] Hao Y, Qiu Z, Holmes J, Löckenhoff CE, Liu W, Ghassemi M, Kalantari S. Large language model integrations in cancer decision-making: a systematic review and meta-analysis. NPJ Digit Med. 2025 Jul 17;8(1):450. doi: 10.1038/s41746-025-01824-7. PMID: 40676129; PMCID: PMC12271406.
[3] Wong EYT, Verlingue L, Aldea M, Franzoi MA, Umeton R, Halabi S, Harbeck N, Indini A, Prelaj A, Romano E, Smyth E, Tan IB, Valachis A, Vibert J, Wiest IC, Yang YH, Gilbert S, Kapetanakis G, Pentheroudakis G, Koopman M, Kather JN. ESMO guidance on the use of Large Language Models in Clinical Practice (ELCAP). Ann Oncol. 2025 Dec;36(12):1447-1457. doi: 10.1016/j.annonc.2025.09.001. Epub 2025 Oct 18. PMID: 41111032.
[4] Adams C. Opportunities and Challenges in Implementing Large Language Models (LLMs) in Oncology. JCO Clin Cancer Inform. 2026 Mar;10:e2600031. doi: 10.1200/CCI-26-00031. Epub 2026 Mar 10. PMID: 41805367.
Featured image: © 2926 Ch LLM and AI in Oncology. Used with permission.
DOI
