How Advanced AI Can Improve Cancer Clinical Trial Decision-Making: The 2026 Evidence Guide
Last updated: October 15, 2026
Cancer research is becoming increasingly complex. Tumor sequencing, molecular pathology, immunotherapy, antibody-drug conjugates, cell therapies, targeted medicines and combination regimens have created more opportunities for individualized treatment. At the same time, researchers must make difficult decisions about which patients to recruit, which biomarkers to study, how to measure treatment response and when the available evidence justifies advancing a drug to the next stage of development.
Advanced artificial intelligence (AI) can help improve cancer clinical trial decision-making by connecting fragmented clinical, molecular, imaging and research data. Machine learning, large language models, natural language processing, computer vision, causal inference and multimodal AI can assist researchers in translating complex information into structured, reviewable evidence.
The most immediate opportunities involve identifying potentially eligible patients, interpreting trial eligibility criteria, improving recruitment forecasts, extracting information from medical records and supporting clinical trial monitoring. More ambitious applications include biomarker-based treatment selection, predicting resistance, optimizing adaptive trials, assessing imaging endpoints and constructing external control groups.
However, cancer research requires an important distinction between predicting what might happen and establishing what actually happened because of a treatment. An AI model may predict that a patient is likely to respond, but this does not prove that an investigational medicine causes the response or improves survival.
This guide examines how AI can strengthen oncology research, the evidence available as of October 2026, and the scientific and regulatory safeguards required to use it responsibly.
- Improve trial access: AI can help identify potentially eligible cancer patients and match molecular profiles against complex trial criteria.
- Improve study design: AI can help analyze eligibility requirements, historical outcomes, recruitment feasibility and candidate biomarkers.
- Improve evidence quality: validated systems can assist with imaging, pathology, safety surveillance, data extraction and missing-data detection.
- Support earlier decisions: AI can help investigators explore whether an oncology development program should continue, change direction or stop.
- Preserve scientific standards: randomized comparisons, clinically meaningful endpoints, validated biomarkers and independent statistical analysis remain essential.
- AI in cancer clinical trials: the big picture
- Improving cancer trial design
- Biomarker-driven patient matching
- Patient recruitment and eligibility screening
- Site selection and equitable enrollment
- Precision oncology and treatment-response prediction
- Adaptive trials and dose optimization
- AI in cancer endpoint assessment
- Safety monitoring and treatment resistance
- Digital twins and synthetic control arms
- Improving go/no-go decisions
- Integrating real-world evidence
- Evidence from published research
- Bias, limitations and overconfidence
- A practical implementation framework
- The oncology clinical research knowledge graph
- What comes next
- Conclusion
- Frequently asked questions
- References and further reading
1. AI in Cancer Clinical Trials: The Big Picture
Clinical trials generate large volumes of information that are often difficult to integrate. Relevant evidence may be distributed across pathology reports, genomic testing, electronic health records, imaging studies, laboratory results, adverse-event reports, patient questionnaires and published research.
AI can help organize this information and provide structured support for decisions made throughout the oncology research lifecycle.
| Decision area | Potential AI application | Scientific or operational value | Principal limitation |
|---|---|---|---|
| Trial design | Analyze historical protocols, outcomes and recruitment data. | Identify design weaknesses and unrealistic assumptions before recruitment. | Historical patterns may not generalize to a new drug or population. |
| Biomarker matching | Standardize mutations, protein expression, fusions and eligibility criteria. | Identify trials potentially relevant to a patient's tumor biology. | A molecular alteration does not automatically establish treatment sensitivity. |
| Recruitment | Screen clinical records against trial eligibility requirements. | Reduce repetitive manual screening and identify overlooked candidates. | Prescreening is not final eligibility assessment or informed consent. |
| Treatment stratification | Analyze molecular, pathological and clinical variables. | Generate hypotheses about treatment-effect differences among patients. | Predicted response does not prove comparative treatment benefit. |
| Imaging and pathology | Quantify lesions, characterize tissue and support image review. | Improve consistency and reduce repetitive measurement work. | Algorithms may fail across scanners, institutions or tumor types. |
| Safety monitoring | Detect unusual patterns in laboratory values, symptoms and adverse events. | Direct attention to signals that require clinical investigation. | Statistical associations do not establish causation. |
| Development decisions | Model scenarios, uncertainty and expected trial outcomes. | Support transparent decisions about whether to continue development. | Predictions cannot replace prespecified statistical criteria. |
A major review published in Nature Reviews Clinical Oncology on August 7, 2026, examined AI applications across pretrial planning, clinical trial conduct and post-trial inference. The authors identified candidate-patient identification, eligibility assessment, structured data extraction and trial monitoring as particularly promising operational applications.
The same review cautioned that digital twins, synthetic control arms and other methods intended to replace or substantially alter conventional evidence generation remain less mature and face important validation and regulatory challenges.
2. Improving Cancer Clinical Trial Design
A cancer trial can fail for multiple reasons: the drug lacks meaningful activity, the target population is incorrectly defined, the dose is unsuitable, the endpoint does not capture benefit, recruitment is inadequate or the study cannot distinguish the treatment effect from the natural course of the disease.
 |
| Diverse cancer hallmarks targeted by repurposed non-oncology drugs. This figure was created with Biorender.com. Source: Nature 2024 |
AI can help investigate several of these problems before a trial begins. Its value is greatest when it exposes assumptions that researchers can verify, rather than simply producing a more elaborate protocol.
2.1 Testing whether a target population is recruitable
Researchers can analyze de-identified clinical records, cancer registries and historical trial datasets to estimate how many patients are likely to meet a proposed study's eligibility criteria.
For example, a trial may require all of the following:
- A specific cancer type and histological subtype.
- A defined disease stage or treatment setting.
- A particular mutation, gene fusion, protein-expression level or other biomarker.
- Previous exposure to specified treatments.
- Acceptable organ function and performance status.
- No prohibited comorbidity, prior therapy or competing clinical condition.
Individually, these requirements may appear reasonable. Together, they may substantially narrow the eligible population.
An AI-supported feasibility analysis could estimate the number of potentially eligible patients at participating hospitals, identify the criteria that most frequently exclude patients and show whether recruitment assumptions depend on incomplete biomarker testing.
2.2 Identifying overly restrictive eligibility criteria
AI can help investigators assess whether each eligibility criterion is necessary for participant safety, interpretation of the primary endpoint or the scientific objective.
For instance, a trial might exclude patients with a common, stable comorbidity even though that exclusion has little relevance to the investigational treatment's known risks. The model may identify this criterion as a major recruitment barrier.
That finding should trigger scientific and clinical review, not automatic removal of the restriction. Broadening eligibility can improve access and real-world relevance, but safety protections must remain appropriate to the drug, disease and study design.
2.3 Comparing alternative trial designs
Researchers can use historical evidence and statistical simulation to explore alternative sample sizes, stratification factors, recruitment assumptions, assessment schedules and endpoint definitions.
In oncology, these analyses may be particularly valuable when the disease is uncommon, the molecular subgroup is small, several treatment arms compete for the same patient population or long-term follow-up is required.
AI can help generate and analyze scenarios, while validated statistical methods determine whether the proposed design has acceptable operating characteristics. Simulations are only as credible as the data, assumptions and model structure underlying them.
3. Biomarker-Driven Patient Matching: Connecting Tumor Biology to Trial Eligibility
Precision oncology has created an increasing need to match patients to trials using more than a broad cancer diagnosis.
Depending on the treatment and cancer type, eligibility may depend on genomic alterations, gene fusions, protein expression, amplification, tumor mutational burden, mismatch repair status, microsatellite instability, homologous recombination deficiency or other molecular and pathological features.
AI can help translate complex trial descriptions into structured, searchable eligibility requirements.
3.1 From free text to structured biomarkers
Trial registries frequently describe molecular eligibility criteria in natural language. The terminology may vary among studies, laboratories and researchers.
A large language model may help identify relevant biomarker terms and organize them into a structured format, including whether a feature is required, prohibited or relevant only to a specific study cohort.
However, correct biomarker extraction requires more than recognizing a gene name. The system must preserve:
- Biological meaning: distinguish mutations, gene amplification, gene fusions, protein expression and other molecular events.
- Variant specificity: preserve the exact alteration when the protocol distinguishes among variants.
- Logical relationships: retain whether criteria require A AND B, A OR B, or different combinations for separate cohorts.
- Clinical context: distinguish biomarkers used for diagnosis, prognosis, eligibility or prediction of treatment response.
- Exclusion conditions: retain negative requirements and disqualifying findings.
A study published in npj Digital Medicine in May 2025 investigated large language models for extracting and structuring genetic biomarkers from oncology trial descriptions. The researchers emphasized that this was a trial-document processing task rather than direct processing of individual patient records.
The distinction matters: correctly extracting a biomarker from a protocol is a necessary building block for patient matching, but it does not by itself demonstrate accurate patient-level eligibility decisions.
3.2 Connecting biomarkers with treatment and resistance
A molecular alteration should not be interpreted in isolation. Its significance depends on the cancer type, histology, disease stage, treatment setting, specific therapy and available clinical evidence.
For example, a genomic alteration may support enrollment in a trial testing a targeted therapy, yet response could be affected by co-occurring mutations, prior exposure to drugs from the same class, tumor heterogeneity or acquired resistance.
An AI research platform can help organize these relationships and highlight important differences between superficially similar cases.
Cancer type → Histology → Stage → Biomarker or molecular profile → Prior treatments → Investigational therapy → Comparator → Response → Duration of response → Progression → Resistance mechanism → Toxicity → Survival and quality-of-life outcomes.
This structure helps prevent a common analytical error: treating all patients with the same cancer label or biomarker as if they had identical treatment prospects.
3.3 Biomarker relevance is not the same as treatment effectiveness
An alteration may be biologically interesting without being clinically predictive for a particular drug. A biomarker may also be prognostic without identifying which treatment produces better outcomes.
Investigators should distinguish three questions:
- Biological question: Is the alteration involved in a relevant cancer process?
- Predictive question: Does the biomarker identify a difference in benefit from a particular treatment compared with a suitable alternative?
- Clinical question: Does using this biomarker to guide treatment improve outcomes for the intended patient population?
AI can help generate and prioritize hypotheses for all three questions, but each requires its own evidence. A plausible biological mechanism is not sufficient proof of clinical benefit.
4. Improving Patient Recruitment and Eligibility Screening
Recruitment is one of the most practical applications of AI in cancer research because eligibility information is often distributed across numerous medical records and trial documents.
A patient may have the appropriate cancer diagnosis but require further assessment of histology, treatment history, imaging, molecular results, laboratory measurements and clinical status before a study team can determine eligibility.
4.1 How an AI-powered screening workflow works
- Identify the candidate population: search authorized clinical records for relevant cancer diagnoses and treatment settings.
- Extract clinical information: identify pathology, molecular findings, laboratory results, prior therapies and other potentially relevant details.
- Interpret the trial protocol: transform inclusion and exclusion criteria into structured requirements.
- Compare patient records: distinguish supported matches, possible matches, confirmed exclusions and criteria requiring additional information.
- Prioritize manual review: present potentially eligible patients with traceable supporting evidence.
- Complete clinical screening: allow authorized study staff to verify eligibility, discuss the study and document informed consent.
The strongest design combines AI with conventional software. Language models can help interpret unstructured text, while deterministic rules should enforce critical requirements that can be checked precisely. Every consequential match should be linked to its supporting source data.
4.2 Real-world example: AI-assisted cancer trial prescreening
A study published in JCO Clinical Cancer Informatics in January 2026 evaluated a real-time, AI-driven patient-matching system using a common clinical data model.
In an initial metastatic colorectal cancer trial evaluation, the system achieved 94% retrospective accuracy and 88% prospective accuracy against the study's reference chart review, with 100% sensitivity in the reported evaluation. The authors also reported a tenfold reduction in chart-review workload in their comparison.
Across implementation involving 29 clinical trials, the system had screened 98,348 patients, identified 825 eligible candidates and facilitated 117 enrollments at the time reported.
These results illustrate an important distinction between model performance and clinical utility. A practical system must not only classify eligibility reasonably well but also work within the research team's normal workflow and contribute to real recruitment.
The results come from a particular health-system implementation and should not be assumed to generalize unchanged to every cancer center, trial design or patient population.
4.3 Patient matching must not become automated exclusion
An AI model may miss an eligible patient because the medical record uses different terminology, a genomic result is stored in a scanned report, a test has not yet been completed or the model misinterprets a historical treatment.
For this reason, an uncertain or incomplete match should not automatically exclude a patient from consideration. The system should identify the missing information and allow the clinical team to resolve it.
Prescreening must also respect patient privacy, applicable data-protection laws, institutional access controls and research ethics requirements. Finding a potentially relevant trial does not authorize contacting a patient without appropriate procedures.
5. AI for Trial Site Selection and Equitable Enrollment
Trial performance depends partly on where research takes place and whether sites can reach the population the study is intended to investigate.
AI-supported site selection may integrate historical recruitment, investigator experience, patient populations, facility capabilities, referral pathways, competing studies and geographic access.
A useful model should distinguish between a site that previously recruited many patients in a common cancer and a site that has access to the particular molecular subgroup required for a new trial.
5.1 Beyond historical enrollment volume
Site models should evaluate:
- The estimated number of potentially eligible patients, not merely the total number of cancer patients.
- Access to relevant molecular testing and specialist pathology services.
- Experience with the investigational treatment, imaging requirements and safety procedures.
- Competing trials targeting similar patients.
- Distance to the study site, visit burden and the availability of appropriate support services.
- Potential barriers affecting older adults, rural communities and groups historically underrepresented in cancer research.
A March 2026 study in Nature Health described DocTr, a cross-modal deep-learning framework designed to recommend clinicians and investigators for clinical trials using patient encounter data, trial documents and historical enrollment relationships.
The work demonstrates how AI can approach trial site and investigator selection as a structured recommendation problem. However, recommendation accuracy is not equivalent to improved recruitment, better representation or superior clinical outcomes. Those outcomes require their own evaluation.
Source: Matching clinicians with clinical trials using AI. Nature Health, 2026.
6. AI for Precision Oncology and Treatment-Response Prediction
Cancer is not a single disease, and even patients with the same histological diagnosis may have different molecular alterations, tumor microenvironments, previous treatments and resistance mechanisms.
AI can integrate clinical variables, genomic profiles, pathology slides, radiology images and other data to develop hypotheses about treatment response and patient heterogeneity.
6.1 Potential applications
- Biomarker discovery: identify molecular or pathological features associated with response, resistance or prognosis.
- Patient stratification: identify subgroups that may require separate analysis or deserve further study.
- Trial enrichment: assess whether a biologically defined population is suitable for evaluating a targeted therapy.
- Combination therapy research: identify hypotheses about complementary mechanisms or potential resistance pathways.
- Treatment-effect heterogeneity: investigate whether the relative effects of treatments vary across clinically important subgroups.
6.2 Predicting response is not proving benefit
Suppose an AI model predicts that patients whose tumors contain a particular molecular alteration are more likely to respond to an investigational medicine.
That prediction may be useful for selecting hypotheses, planning a study or identifying patients for further review. But it does not establish that the medicine works better than standard treatment, that the observed responses are caused by the medicine, or that patients live longer.
To establish a predictive biomarker, researchers generally need to evaluate whether treatment effects differ according to biomarker status using appropriate comparative data and statistical methods.
External validation is important because an association discovered in one dataset may disappear in another population or treatment setting.
6.3 Immunotherapy requires particular care
Immunotherapy studies can be difficult to analyze because immune-related responses, delayed responses, atypical imaging patterns and immune-mediated adverse events may complicate interpretation.
For example, radiological enlargement of a lesion does not always have the same implication in every treatment context. Researchers must follow the assessment rules specified in the protocol, including an appropriate response-assessment framework when relevant.
AI may assist with image measurements, lesion tracking and detection of unusual trajectories. It should not independently redefine progression or overrule the study's endpoint rules merely because a prediction model expects a later response.
7. Adaptive Cancer Trials, Bayesian Methods and Dose Optimization
Adaptive trials permit specified changes as data accumulate, using methods designed to preserve the validity of the study. Depending on the protocol, adaptations may involve sample size, dose selection, treatment-arm allocation or decisions about whether further enrollment is justified.
AI can support simulation, recruitment forecasting, data review and the interpretation of complex patterns. However, AI and adaptive trial methodology are not interchangeable: the statistical design determines which adaptations are permissible and how evidence should be evaluated.
7.1 Dose optimization
In oncology, a dose that produces substantial biological activity may not provide the best overall balance of efficacy, tolerability and treatment burden.
This is particularly important for targeted medicines, immunotherapies, antibody-drug conjugates, radiopharmaceuticals and combination regimens, where dose-response relationships may not follow a simple pattern.
AI-supported models can help explore exposure-response relationships, summarize toxicity patterns and simulate candidate dose schedules. Decisions should still be grounded in pharmacology, clinical observations, prespecified analysis and appropriate safety oversight.
7.2 Adaptive designs and interim analyses
Researchers may use prespecified interim analyses to investigate whether a trial is progressing as expected, whether a treatment arm meets a stopping criterion, or whether the study should continue.
AI can help identify potential data problems or simulate alternative scenarios before a study begins. Nevertheless, informal model predictions should not introduce unplanned changes into a trial in ways that undermine its statistical validity.
Investigators should specify the decision rules, responsible decision-makers, access to unblinded data and documentation requirements before implementing any consequential adaptation.
For guidance on Bayesian methods, consult the FDA's Bayesian methodology guidance for clinical trials of drugs and biological products. Its status and applicability should be checked before use; the publication listed by the agency in January 2026 is draft guidance.
8. AI in Cancer Treatment-Response and Endpoint Assessment
Clinical endpoints determine what a cancer trial can legitimately conclude. AI can help measure or interpret certain endpoints, but the endpoint itself must remain clinically meaningful and fit for the intended regulatory or scientific claim.
| Endpoint | Meaning | Possible AI contribution | Key safeguard |
|---|---|---|---|
| Objective response rate (ORR) | Proportion of patients meeting the specified criteria for complete or partial response. | Assist with lesion measurement and image review. | Use a prespecified, appropriate response-assessment standard. |
| Duration of response (DoR) | Time from the documented response until the defined endpoint event. | Help organize response timelines and flag inconsistent assessments. | Preserve endpoint definitions and handle censoring correctly. |
| Progression-free survival (PFS) | Time until defined disease progression or death, according to the protocol. | Support imaging review and data-quality checks. | Maintain valid progression assessments and statistical methods. |
| Overall survival (OS) | Time until death from any cause. | Reconcile records and support follow-up workflows. | AI predictions cannot substitute for observed survival evidence. |
| Pathological complete response (pCR) | A protocol-defined absence of residual invasive cancer in specified surgical specimens after treatment. | Assist with pathology analysis and measurement. | Use validated pathological criteria and appropriate endpoint interpretation. |
| Minimal residual disease (MRD) | Evidence of residual disease measured using a specified sensitive assay. | Help analyze assay data and longitudinal patterns. | Validate the assay and the endpoint's intended clinical use. |
| Patient-reported outcomes | Patient reports of symptoms, function and health-related quality of life. | Summarize trends and flag missing questionnaires. | Use appropriate validated instruments and protect patients' own reports from algorithmic distortion. |
8.1 Imaging and pathology
Computer-vision algorithms can assist with lesion measurements, tumor segmentation, tissue characterization and the review of pathology images.
In a multicenter oncology trial, however, the same algorithm may encounter scanners, image acquisition procedures, staining methods and patient populations that differ from those used to develop it.
Researchers must establish the algorithm's intended use, evaluate its performance under relevant conditions and document how disagreements between AI-assisted assessments and qualified reviewers will be resolved.
When imaging contributes to a primary endpoint, the study design should address acquisition quality, image handling, interpretation procedures and any applicable requirements for independent review.
Reference: FDA, Clinical Trial Imaging Endpoint Process Standards Guidance for Industry.
8.2 Biomarker change is not automatically clinical benefit
An intervention might alter a circulating biomarker, molecular signature, radiographic measurement or other intermediate measure without improving survival or quality of life.
AI can help identify relationships between intermediate measures and outcomes, but the existence of a correlation does not establish that the biomarker reliably captures the treatment effect.
Researchers must distinguish exploratory biomarkers from validated surrogate endpoints and explain exactly what their study results do and do not demonstrate.
8.3 Including the patient's experience
Imaging and laboratory results cannot capture every meaningful effect of treatment. Fatigue, pain, functional limitations, treatment burden and the ability to maintain daily activities may be important to patients.
AI can assist in summarizing patient-reported outcome data and identifying incomplete assessments. It should not replace validated instruments or infer that a patient feels better simply because clinical or imaging indicators improve.
Reference: FDA, Core Patient-Reported Outcomes in Cancer Clinical Trials, October 2024.
9. AI for Safety Monitoring, Progression and Treatment Resistance
Oncology drugs can produce complex safety profiles, particularly when used in combinations or over prolonged periods. AI can assist study teams in identifying unusual patterns across adverse events, clinical notes, laboratory values, medications and treatment exposure.
9.1 Safety signal detection
Potential applications include:
- Identifying clusters of similar adverse events across patients or study sites.
- Flagging unexpected laboratory changes following treatment exposure.
- Detecting possible differences in tolerability between doses or patient subgroups.
- Reconciling safety reports with clinical records and treatment timelines.
- Highlighting missing assessments or potential reporting delays.
An AI-generated signal is a reason to investigate, not proof that a treatment caused the event. Cancer severity, pre-existing conditions, concomitant medicines, prior therapies and differences in monitoring may explain some apparent associations.
Safety decisions must follow the study's procedures, applicable reporting obligations and responsibilities of investigators, sponsors and independent monitoring bodies. Automated systems must not delay urgent clinical assessment.
9.2 Predicting and studying resistance
Acquired resistance is a major challenge in targeted therapy and other cancer treatments. Resistance may be associated with new molecular alterations, clonal selection, pathway reactivation, changes in the tumor microenvironment or other mechanisms.
AI may help researchers combine serial tissue biopsies, circulating tumor DNA, imaging, clinical response and prior-treatment records to generate hypotheses about resistance.
Such analyses can guide the design of subsequent studies or identify questions worth testing in a new cohort. They do not necessarily establish the causal mechanism of resistance or demonstrate that switching treatment will improve outcomes.
9.3 Early futility decisions
Models may help investigate whether a trial remains likely to answer its research question under specified assumptions. They can also flag problems such as slow recruitment, missing data or response rates that differ from expectations.
However, the apparent lack of activity may arise from several different causes:
- The medicine does not have meaningful activity in the studied population.
- The biological target was not present or was not adequately inhibited.
- The dose or treatment schedule was unsuitable.
- The target population was too heterogeneous.
- The study's endpoint, follow-up period or data collection was inadequate.
- Recruitment or treatment delivery differed materially from the protocol's assumptions.
AI can help distinguish these possibilities for investigation, but the decision to stop or continue a program should consider clinical evidence, uncertainty, safety, data integrity and the prespecified statistical framework.
10. Digital Twins and Synthetic Control Arms in Oncology
Researchers are investigating whether AI-generated patient representations, simulated outcomes and external comparator groups could help answer selected research questions when conventional randomized studies are difficult or impractical.
These approaches have potential value, particularly in rare cancers, very small molecular subgroups or settings where relevant historical data exist. They also raise substantial methodological challenges.
10.1 What is a digital twin?
In clinical research, a digital twin is a computational representation of an individual, disease process or biological system intended to model relevant characteristics or behavior.
An oncology-oriented model might use longitudinal clinical records, tumor characteristics, treatment history and observed outcomes to simulate possible disease trajectories.
Researchers could use such a model to explore hypotheses, compare design scenarios or estimate how different assumptions might affect trial planning.
However, a simulated trajectory is a model output, not a directly observed counterfactual. Its reliability depends on its assumptions, the quality of the underlying data and whether it remains accurate for the patient population and treatment context in which it is used.
10.2 Can synthetic controls replace randomized controls?
Not as a general rule. External control groups may be appropriate in selected circumstances, but comparisons with historical or real-world patients are vulnerable to confounding and differences in patient selection, treatment standards, endpoint measurement and follow-up.
Randomization helps balance measured and unmeasured baseline characteristics in expectation. Statistical adjustment cannot guarantee that an external control eliminates all relevant differences.
AI-generated synthetic data introduces additional concerns: a model may reproduce selection bias, omit uncommon clinical events, distort correlations or generate implausible observations.
The 2026 review in Nature Reviews Clinical Oncology identifies synthetic controls and digital twins as promising but less mature applications requiring further prospective validation. They should currently be treated as research methods with context-specific roles, not universal substitutes for conventional clinical evidence.
11. Improving Cancer Drug Development Go/No-Go Decisions
One of the most consequential decisions in drug development is whether to advance, modify, expand or discontinue an investigational treatment.
These decisions involve more than predicting the probability of success. Investigators must assess the magnitude and reliability of the treatment effect, safety, biological plausibility, patient population, competitive treatment landscape, operational feasibility and remaining uncertainty.
11.1 The AI-assisted evidence review
A decision-support system could assemble a structured review containing:
- Target validity: whether the biological hypothesis remains supported by the available evidence.
- Pharmacological activity: whether the treatment achieves the intended exposure and biological effect.
- Clinical activity: observed response rate, response duration and other relevant outcomes.
- Comparative benefit: evidence relative to an appropriate control or standard treatment.
- Safety: adverse events, dose interruptions, discontinuations and treatment-related mortality where relevant.
- Population relevance: whether results apply to the intended cancer subtype and molecular subgroup.
- Data reliability: missingness, protocol deviations, assessment quality and uncertainty in the results.
- Next experiment: what additional study would most directly resolve the remaining uncertainty.
This approach is especially useful when a development program produces mixed results. A treatment may show biological activity but lack sufficient response durability, or appear promising in a small subgroup without reliable comparative evidence.
11.2 A practical decision framework
| Decision | Evidence supporting consideration | What AI can contribute |
|---|---|---|
| Advance | Acceptable safety, credible activity and a scientifically justified next-stage study. | Summarize evidence, quantify uncertainty and simulate development scenarios. |
| Refine | Evidence suggests that dose, schedule, population or study methods may be suboptimal. | Identify candidate explanations and compare revised design assumptions. |
| Expand a cohort | A promising signal requires confirmation in a defined population. | Review subgroup evidence and estimate recruitment feasibility. |
| Pause | Important safety, data-integrity or methodological concerns require resolution. | Identify discrepancies and prioritize records or questions for investigation. |
| Stop | The evidence and prespecified criteria indicate an unfavorable benefit-risk profile or inadequate justification for further development. | Support transparent evidence synthesis; final responsibility remains with authorized decision-makers. |
AI should not be assigned a universal threshold that decides whether every cancer drug succeeds or fails. The appropriate decision criteria differ between early-phase dose-finding studies, randomized phase II trials, confirmatory phase III studies and trials in rare or biomarker-defined diseases.
12. Integrating Real-World Evidence with Cancer Clinical Trials
Real-world evidence may help researchers understand patient characteristics, treatment patterns, disease progression, outcomes and the feasibility of future studies.
Relevant sources can include cancer registries, electronic health records, claims databases, molecular testing systems, pathology archives and post-treatment follow-up records.
AI can help harmonize information across these sources, identify comparable patient cohorts, summarize historical outcomes and support the design of subsequent studies.
12.1 Why real-world evidence requires careful interpretation
Patients treated in routine practice may differ substantially from participants in clinical trials. Their disease severity, comorbidities, testing patterns, prior therapies and follow-up may be different.
Clinical records may also have missing outcomes, inconsistent terminology and different definitions of disease progression. A lack of recorded progression is not necessarily evidence that progression did not occur.
Methods such as target-trial emulation can help specify the research question, eligibility criteria, treatment strategies, follow-up, outcomes and analysis for an observational comparison that seeks to approximate a target randomized trial.
AI may assist with assembling and analyzing the required data, but it cannot eliminate confounding by itself. The assumptions necessary for a causal conclusion must still be stated and evaluated.
12.2 Appropriate uses of AI-supported real-world evidence
- Estimating potential recruitment populations before launching a trial.
- Identifying gaps in biomarker testing and clinical documentation.
- Understanding treatment sequences before and after the investigational intervention.
- Exploring outcomes in patient groups underrepresented in existing trials.
- Assessing whether an external comparator may be feasible and scientifically defensible.
- Generating hypotheses for prospective testing.
Real-world evidence is most useful when its contribution is clearly defined. Exploratory findings, adjusted observational comparisons and randomized treatment effects should not be presented as if they provide the same strength of causal evidence.
13. What Published Research Demonstrates
The following studies illustrate different ways AI is being applied to oncology research. They also show why the precise evaluation setting matters.
| Study | Research application | Key finding or contribution | Important limitation |
|---|---|---|---|
| Villa et al., 2026 Nature Reviews Clinical Oncology |
Review of AI throughout the oncology trial lifecycle | Identifies practical opportunities in patient matching, eligibility review, data extraction and monitoring. | Reviews multiple applications with different levels of maturity; not a prospective trial of one AI system. |
| Gong et al., 2026 JCO Clinical Cancer Informatics |
AI-assisted clinical trial prescreening | Reports implementation across 29 trials and screening of 98,348 patients, with 825 eligible candidates identified and 117 enrollments facilitated at the time reported. | Evidence from a particular healthcare-system implementation; results may vary elsewhere. |
| Alkhoury et al., 2025 npj Digital Medicine |
Extraction of biomarkers from oncology trial descriptions | Investigates structuring molecular criteria and their logical relationships for downstream trial matching. | Focuses on trial documents, not direct end-to-end verification of patient eligibility. |
| Gao et al., 2026 Nature Health |
AI-assisted investigator recommendations | Explores cross-modal learning for recommending clinicians and investigators for trials. | Matching performance is not equivalent to improved enrollment or patient outcomes. |
| FDA, January 2025 | Credibility assessment of AI models | Proposes a risk-based framework for evaluating models used to support regulatory decisions about drug safety, effectiveness or quality. | Draft guidance; not blanket authorization for AI-generated evidence or autonomous oncology decisions. |
| FDA and EMA, January 2026 | Good AI practice throughout drug development | Sets out ten shared principles covering human-centric design, risk, standards, context of use, governance, validation and lifecycle management. | Principles must be translated into appropriate controls for each system and use case. |
The central lesson is not that AI has already solved cancer drug development. It is that some practical tasks can be made more efficient, while more ambitious claims about treatment effects and replacement of conventional evidence require stronger proof.
14. Bias, Limitations and the Risk of AI Overconfidence
Advanced AI can produce highly plausible explanations even when the supporting data are incomplete or misleading. In oncology research, this is particularly concerning because treatment decisions involve substantial uncertainty and patient outcomes may be influenced by many interdependent factors.
14.1 Selection bias
If the training data mainly contain patients treated at major academic hospitals, the model may perform less well in community hospitals, lower-resource health systems or populations with different patterns of diagnosis and access to molecular testing.
A model may also systematically miss patients whose records are incomplete, even though incomplete documentation does not necessarily mean that the patient is clinically ineligible.
14.2 Biomarker bias and overinterpretation
A biomarker may be more common in datasets from certain institutions or patient populations because testing is not performed uniformly. A model may mistake differences in testing practice for differences in disease biology.
Researchers must also account for multiple testing when a model explores thousands of molecular features and patient subgroups. Without suitable statistical control and independent validation, some apparently favorable associations will arise by chance.
14.3 Confounding and causal inference
Suppose an observational dataset shows that patients receiving an experimental therapy have better outcomes than patients receiving another treatment. The difference may reflect patient selection, disease severity, previous treatment, treatment access or other confounding factors.
AI can adjust for observed variables using suitable methods, but it cannot guarantee that unmeasured confounding has been removed. A convincing association in real-world data does not automatically establish a causal treatment effect.
14.4 Data leakage
Data leakage occurs when information that would not legitimately be available at the intended prediction time influences model training or evaluation.
For example, a model intended to predict treatment response at baseline may inadvertently use information recorded after treatment began. Such a model can appear highly accurate without being useful for the actual decision it is intended to support.
14.5 Model drift
Cancer treatment changes over time. New medicines, biomarker tests, diagnostic procedures and standards of care alter the patient populations and clinical decisions seen in practice.
A model validated on older records may lose accuracy as the clinical environment changes. Ongoing monitoring should examine calibration, error rates, subgroup performance and the need for revalidation.
15. A Practical Implementation Framework for Cancer Research Teams
Oncology research centers, hospitals, sponsors and academic laboratories should start with a clearly defined research problem rather than adopting AI because the technology is available.
Step 1: Define the decision
Choose a specific task, such as identifying candidates for a biomarker-driven trial, checking imaging records for missing assessments or forecasting recruitment at proposed sites.
Specify who will use the output, when it will be used and what action it is expected to support.
Step 2: Define a measurable objective
Possible objectives include improving screening efficiency, increasing the proportion of eligible patients identified, reducing the time required to complete chart reviews or improving the consistency of endpoint measurement.
A clinical research team should avoid vague goals such as making trial decisions smarter. It should define measurable outcomes against a credible baseline.
Step 3: Establish a high-quality reference dataset
Clinical records, pathology reports, imaging and molecular results need consistent definitions, traceable sources, appropriate access permissions and suitable quality controls.
For a patient-matching system, the reference dataset should include expert-reviewed eligibility assessments. For an imaging model, it should include appropriately defined reference assessments and an evaluation of reader variability where relevant.
Step 4: Validate against the intended use
Model performance should be assessed in the population, institution, disease setting and workflow where it will actually be used.
Where possible, evaluate the model on independent patients and sites that were not used for development. Examine false positives and false negatives, subgroup performance, calibration and the consequences of erroneous recommendations.
Step 5: Keep a human review pathway
Determine which outputs can be used for administrative support and which require clinical, statistical or regulatory review before an action is taken.
High-risk decisions should have explicit authorization procedures, appropriate escalation routes and documented accountability.
Step 6: Monitor real-world performance
After implementation, measure whether the system achieves its intended goal and whether errors differ across cancer types, sites or patient populations.
Track model changes, data drift, workflow interruptions and unexpected outcomes. Establish criteria for restricting or suspending the system when performance becomes unreliable.
Step 7: Document the evidence
Maintain records of the intended use, training data, validation, model version, known limitations, user instructions and decision history.
Where AI contributes to regulatory evidence, engage appropriate regulatory and quality teams early to determine the required level of model credibility assessment.
16. Building an Oncology Clinical Research Knowledge Graph
One potentially valuable direction is to combine AI with a structured oncology knowledge graph. Rather than storing trial findings as disconnected documents, a knowledge graph represents entities and their relationships in a format that can be queried, evaluated and updated.
This is especially relevant to precision oncology, where the same biomarker may have different significance depending on the cancer type, treatment context, resistance mechanism and quality of supporting evidence.
16.1 Core knowledge graph entities
| Entity | Examples of associated information | Research question supported |
|---|---|---|
| Cancer | Primary site, histology, stage and disease setting. | Which population is relevant to the research question? |
| Biomarker | Gene, variant, protein expression, assay method and result. | What molecular feature defines the study population? |
| Treatment | Investigational agent, class, dose, combination and comparator. | Which intervention was evaluated, and against what? |
| Clinical trial | Phase, protocol, eligibility criteria, site, design and registration. | Under what conditions was the treatment studied? |
| Response | Response category, assessment method, duration and time point. | What response was observed and how was it measured? |
| Resistance | Acquired alteration, potential pathway, prior exposure and timing. | What mechanisms may explain loss of treatment activity? |
| Safety | Adverse event, severity, timing, exposure and attribution assessment. | What harms were observed and how certain is the causal attribution? |
| Evidence | Study design, sample size, comparator, endpoint, confidence interval and risk of bias. | How reliable is the conclusion? |
16.2 Connect each claim to its evidence
Every important clinical claim should connect to the study that supports it. A graph entry should distinguish a randomized trial result from an observational association, a case report, a laboratory experiment or a computational prediction.
For instance, a knowledge graph might record that a molecular alteration was associated with response in a small single-arm study. It should not silently transform that finding into proof that the treatment improves overall survival compared with standard care.
Useful evidence fields include:
- Publication identifier, DOI or trial registration number.
- Cancer type, histology, stage and biomarker definition.
- Investigational treatment, comparator and treatment setting.
- Study design, population size and relevant eligibility criteria.
- Endpoint definitions, effect estimates and uncertainty intervals.
- Response durability, follow-up and available survival data.
- Funding, conflicts of interest and methodological limitations.
- Replication status, evidence certainty and date of last review.
16.3 The role of advanced AI
Large language models can help extract structured information from publications, protocols and trial registries. Machine-learning systems can identify patterns and generate hypotheses, while conventional databases and rules preserve data consistency.
Every extracted finding should be checked against its original source. Entity resolution, terminology normalization, duplicate detection and provenance are necessary if the graph is to function as a dependable research resource.
A well-designed system can help researchers search across cancer types, biomarkers, treatments, responses and resistance mechanisms while retaining the evidence and uncertainty associated with each claim.
17. What Comes Next for AI in Cancer Clinical Research?
Over the coming years, several developments could strengthen the role of AI in oncology trials.
17.1 More reliable multimodal models
Models that combine genomic data, pathology, imaging, longitudinal clinical records and patient-reported outcomes may provide a more complete representation of the disease and treatment context.
The challenge will be to demonstrate that combining additional data improves decision quality rather than merely increasing model complexity.
17.2 Better integration of molecular diagnostics
Improved standardization of genomic terminology, assay interpretation, variant annotation and eligibility requirements could make biomarker-driven trial matching more consistent.
These advances may be especially useful for basket trials, umbrella trials and other designs in which molecular characteristics help determine cohort assignment.
17.3 More prospective evaluation
Retrospective accuracy is useful, but it cannot fully establish whether an AI system improves real clinical research workflows.
Prospective studies should measure not only model performance but also the effect on screening efficiency, completed enrollment, representativeness, data quality, research cost and the frequency of consequential errors.
17.4 Stronger regulatory alignment
In January 2026, the FDA and European Medicines Agency published ten shared guiding principles for good AI practice in drug development. These principles emphasize human-centric design, a risk-based approach, a clear context of use, appropriate data governance, performance assessment and lifecycle management.
The FDA also published draft guidance in January 2025 proposing a risk-based credibility assessment framework for AI models intended to support regulatory decisions regarding drug safety, effectiveness or quality.
These frameworks support a more disciplined approach to AI evidence, but sponsors still need to establish whether each model is suitable for its intended application and whether its outputs are sufficiently reliable for the decision at hand.
Sources: FDA and EMA Guiding Principles of Good AI Practice in Drug Development, 2026; FDA draft guidance on AI supporting regulatory decision-making, 2025.
18. Conclusion: AI Should Improve the Quality of Cancer Evidence
Advanced AI has the potential to improve cancer clinical research at several levels. It can help research teams find eligible patients, connect molecular findings with trial criteria, assess recruitment feasibility, organize complex data and identify potential problems earlier.
It may also support more sophisticated research into treatment response, resistance, imaging endpoints, adaptive designs and the use of real-world evidence.
However, not all applications are equally mature. Current evidence provides a stronger basis for AI-assisted operational workflows than for approaches that attempt to replace conventional evidence generation with digital twins or synthetic control groups.
For oncology researchers, the goal should not be to automate every decision. It should be to reduce avoidable errors, improve access to appropriate research, make uncertainty visible and ensure that the resulting evidence is more useful for patients and clinicians.
AI can help identify the right question, the right patient population and the right evidence to examine. It cannot independently prove that a cancer treatment works. Reliable conclusions still depend on sound trial design, appropriate comparators, clinically meaningful outcomes, validated methods and transparent reporting.
19. Frequently Asked Questions
How is AI used in cancer clinical trials?
AI can help design trials, identify potentially eligible patients, match molecular biomarkers with eligibility criteria, forecast recruitment, analyze medical images, monitor safety and organize clinical data. More experimental applications include digital twins and synthetic control groups.
Can AI match cancer patients to clinical trials?
Yes. AI systems can compare patient records with trial eligibility criteria and rank potentially relevant studies. In a 2026 implementation study, an AI-assisted system screened 98,348 patients across 29 trials, identified 825 eligible candidates and facilitated 117 enrollments at the time reported. Such results should not be assumed to generalize to every healthcare system.
Can AI predict which cancer treatment will work best?
AI can analyze molecular, pathological, imaging and clinical information to generate treatment-response predictions. However, a prediction is not proof that a treatment provides greater benefit than an alternative. Clinically useful predictive biomarkers require appropriate validation and evidence.
Can AI improve cancer drug development success rates?
AI may improve specific parts of drug development by helping researchers identify relevant patient populations, refine trial design and detect problems earlier. Whether these improvements translate into higher overall success rates must be demonstrated through appropriate empirical evaluation; it should not be assumed from model accuracy alone.
Can AI replace randomized cancer clinical trials?
No, not as a general replacement. AI can support trial design, analysis and selected uses of external comparator data, but synthetic controls and digital twins have important limitations. Randomized trials remain a central method for establishing comparative treatment effects when they are appropriate and feasible.
Can AI detect treatment resistance?
AI can help analyze genomic data, circulating tumor DNA, imaging, treatment history and clinical outcomes to generate hypotheses about resistance. The findings require biological and clinical validation before they can establish a mechanism or justify a treatment change.
What are the main risks of AI in cancer research?
Important risks include biased or incomplete data, incorrect interpretation of biomarkers, overfitting, data leakage, confounding, poor generalization, model drift and overconfidence in generated explanations. Human oversight, appropriate validation, traceable evidence and monitoring are essential.
What is the best first use of AI for an oncology research center?
A clearly defined operational task, such as trial eligibility prescreening, protocol-data extraction or recruitment feasibility assessment, is often a practical starting point. The system should have measurable performance targets, a reliable reference standard, human review and a plan for monitoring errors.
20. References and Further Reading
- Villa A, Eadie AL, Synnott D, et al. AI-based augmentation of oncology clinical trials. Nature Reviews Clinical Oncology. 2026.
- Gong G, Liu J, Pandya S, et al. Clinical Trial Patient Matching: A Real-Time, Common Data Model and AI-Driven System for Semiautomated Patient Prescreening in Cancer Clinical Trials. JCO Clinical Cancer Informatics. 2026.
- Alkhoury N, Shaik M, Wurmus R, et al. Enhancing biomarker-based oncology trial matching using large language models. npj Digital Medicine. 2025.
- Gao J, Xiao C, Glass LM, et al. Matching clinicians with clinical trials using AI. Nature Health. 2026.
- US Food and Drug Administration. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products. Draft guidance, January 2025.
- US Food and Drug Administration and European Medicines Agency. Guiding Principles of Good AI Practice in Drug Development. January 2026.
- US Food and Drug Administration. Clinical Trial Endpoints for the Approval of Cancer Drugs and Biologics. Guidance for Industry.
- US Food and Drug Administration. Clinical Trial Imaging Endpoint Process Standards Guidance for Industry.
- US Food and Drug Administration. Core Patient-Reported Outcomes in Cancer Clinical Trials. October 2024.
Editorial and medical disclaimer: This article is an educational overview of AI in cancer clinical research. It does not endorse a particular investigational medicine, AI vendor or commercial platform. AI-generated analyses and literature summaries should be checked against original publications and applicable protocols. Clinical trial participation, treatment selection, regulatory decisions and safety management require qualified professionals and the relevant ethical and regulatory oversight.
.png)
.png)
.png)
Comments
Post a Comment