Cancer clinical trials are entering an era in which artificial intelligence could influence nearly every stage of the research process, from deciding which patients are most likely to qualify to determining how widely study results can be applied. A new Review in Nature Reviews Clinical Oncology examines how AI is being integrated into oncology trials and argues that its most immediate value is not replacing clinicians or generating evidence independently, but augmenting human decision-making in complex, data-intensive workflows.
The need for new tools is substantial. Oncology trials frequently struggle with slow patient recruitment, high rates of failure and findings that do not fully represent the wider cancer population. These problems arise from a combination of biological complexity, rapidly changing treatment strategies and operational barriers. Trials may require patients to meet highly specific molecular, clinical or previous-treatment criteria, yet identifying eligible individuals across fragmented electronic health records can be difficult and time-consuming. AI systems are increasingly being developed to search these records and help research teams find potential participants more efficiently.
The technology behind these applications includes machine-learning algorithms, natural-language processing and large-scale analysis of electronic health record data. Natural-language processing can interpret information stored in clinical notes, pathology reports, imaging summaries and treatment histories—data that are often unstructured and difficult to analyze using conventional software. Machine-learning models can then compare these details with a trial’s inclusion and exclusion criteria, flagging patients who may be suitable for review by a clinician or research coordinator. This approach does not eliminate the need for human assessment, but it can reduce the time required to locate candidates and identify missing information.
AI may also accelerate eligibility assessment, one of the most labor-intensive steps in trial recruitment. Conventional screening can involve manually reviewing laboratory results, medication histories, imaging records, performance status and prior therapies. An AI system can organize these data and identify potential mismatches or unanswered questions. In principle, this could shorten the interval between a patient’s diagnosis and a trial invitation, which is especially important in cancers where treatment decisions must be made quickly. However, the Review emphasizes that automated suggestions should remain subject to clinical verification because medical records can be incomplete, contradictory or outdated.
Beyond recruitment, AI is being used to improve trial conduct. Algorithms can extract relevant information from clinical records, support case-report-form completion and assist with safety monitoring. Remote systems may help track symptoms, treatment adherence or patient-reported outcomes between clinic visits. These tools could make participation less burdensome by allowing some information to be collected digitally rather than requiring repeated in-person assessments. At the same time, remote monitoring introduces new responsibilities, including ensuring that concerning symptoms are recognized promptly and that patients understand how digital systems fit into their care.
AI can also monitor the trial itself. Real-time analysis of accumulating data may help research teams detect operational problems, such as delays in data entry, uneven recruitment across sites or unexpected patterns in adverse events. Models can be designed to watch for changes in performance, a process known as model monitoring, because an algorithm that works well in one setting may become less reliable when patient populations, clinical practices or data formats change. This phenomenon, often called data drift, is a major concern in oncology, where diagnostic technologies and treatment standards evolve rapidly.
The Review draws a critical distinction between AI that supports existing research processes and AI intended to substitute for parts of clinical evidence generation. Operational tools, such as candidate identification, data extraction and trial monitoring, are already being implemented at selected cancer centers, although their effectiveness and generalizability still require careful evaluation. By contrast, synthetic control arms, outcome-prediction simulations and digital twins remain at an earlier stage. Synthetic control arms use data from previous or external patients to estimate what might have happened without the experimental treatment. Digital twins attempt to model an individual patient’s likely disease course under different interventions. Both concepts are scientifically appealing, but they depend on assumptions that may be difficult to validate prospectively.
The central methodological challenge is causality. A clinical trial is designed to compare outcomes under controlled conditions, while AI models often learn associations from observational data. Even highly accurate prediction does not prove that a treatment caused a particular outcome. Patients represented in historical databases may differ from trial participants in age, disease severity, access to care, genetic background or supportive treatment. If these differences are not properly addressed, an AI-generated comparison could produce misleading estimates of benefit or risk. For this reason, the authors describe prospective validation, transparent methods and regulatory review as essential before AI-generated evidence can influence major treatment decisions.
Equity is another defining issue. Electronic health record datasets may underrepresent people who receive care outside major academic hospitals or who face barriers to diagnosis and treatment. If an algorithm learns from incomplete data, it may perform less accurately for racial, ethnic, socioeconomic or geographically underserved groups. In a trial setting, that could reinforce existing disparities by directing opportunities toward patients who are already easier to identify. Developers and trial sponsors will therefore need to evaluate model performance across diverse populations, document data limitations and design systems that broaden access rather than simply optimize recruitment efficiency.
The emerging message is neither that AI will solve the longstanding problems of cancer trials nor that it should be excluded from clinical research. Instead, AI is most likely to have its earliest and safest impact as a supervised partner for clinicians, trialists and research staff. Its long-term promise will depend on whether tools are tested in real-world prospective studies, governed by harmonized standards and continuously evaluated after deployment. Coordination among patients, healthcare professionals, regulators, technology developers and industry will be crucial. If that oversight is maintained, AI could help make oncology trials faster, more responsive and more representative—while preserving the human judgment required to determine whether new treatments truly work.
Subject of Research: Artificial intelligence applications in oncology clinical trials
Article Title: AI-based augmentation of oncology clinical trials
Article References: Villa, A., Eadie, A.L., Synnott, D. et al. AI-based augmentation of oncology clinical trials. Nat Rev Clin Oncol (2026). https://doi.org/10.1038/s41571-026-01189-0
Image Credits: AI Generated
DOI: 10.1038/s41571-026-01189-0
Keywords: artificial intelligence, oncology, clinical trials, machine learning, electronic health records, patient recruitment, eligibility assessment, trial monitoring, synthetic control arms, digital twins, clinical research, healthcare equity
