This pragmatic, multicenter, cluster-randomized trial will evaluate whether a locked artificial intelligence (AI) clinical decision-support system can improve outcomes by helping multidisciplinary teams select first-line immune checkpoint inhibitor (ICI)-based systemic treatment for adults with unresectable hepatocellular carcinoma (HCC). Twenty-six hospitals or independent HCC multidisciplinary teams will be randomly assigned in a 1:1 ratio to AI-assisted treatment selection or usual-care treatment selection. Approximately 1,800 participants will be enrolled. Eligible participants must already be considered suitable for first-line ICI-based systemic therapy; the study does not compare immunotherapy with no immunotherapy. At AI-assisted sites, the system will use prespecified pretreatment information to estimate and compare expected outcomes across clinically appropriate, locally available, guideline-concordant ICI-based regimens. The AI output is advisory. Treating clinicians and patients retain responsibility for the final treatment decision, and reasons for not following an AI recommendation will be recorded. At usual-care sites, treatment will be selected through the standard multidisciplinary decision-making process without access to the AI output. Both groups will receive approved standard-of-care treatments. The AI model, input definitions, preprocessing pipeline, decision rules, thresholds, and software version will be locked before enrollment of the first participant and will not be retrained or modified using trial outcome data. Both groups will use the same eligibility criteria, patient-registration time point, imaging schedule, follow-up schedule, and outcome definitions. The primary outcome is progression-free survival assessed by blinded independent central imaging review. Overall survival is a key secondary outcome. Additional outcomes include tumor response, duration of response, safety, quality of life, treatment delivery, and implementation measures. This trial evaluates the clinical utility of a prespecified AI system rather than developing or optimizing another prediction model.
Background and Rationale Several first-line immune checkpoint inhibitor (ICI)-based systemic treatment options are available for patients with unresectable hepatocellular carcinoma (HCC), but their relative benefits and toxicity profiles vary between patients. The artificial intelligence (AI) clinical decision-support system evaluated in this trial was developed and externally validated in separate, completed retrospective multicenter studies before initiation of the registered trial. However, predictive performance in retrospective data alone cannot establish whether using the system improves treatment selection or patient outcomes. This prospective trial is therefore designed to evaluate the clinical utility of a prespecified and locked AI system when integrated into routine multidisciplinary care. This ClinicalTrials.gov record pertains exclusively to the prospective cluster-randomized evaluation. The preceding retrospective model-development and external-validation studies are outside the scope of this registry record and are not included in the registered enrollment, study dates, treatment groups, or outcomes. Study Design This is a prospective, multicenter, pragmatic, open-label, two-arm, parallel-group, superiority, cluster-randomized controlled trial with blinded independent central review of the primary imaging endpoint. Twenty-six participating hospitals, each constituting one cluster, will be randomized centrally in a 1:1 ratio to either an AI-assisted treatment-selection strategy or a usual-care treatment-selection strategy. Cluster randomization is used to minimize contamination between treatment strategies and to evaluate the AI system as a component of the multidisciplinary clinical workflow. Randomization will be performed by an independent statistician using a prespecified restricted randomization procedure intended to improve balance in important cluster-level characteristics. All participating hospitals will be identified, activated, and committed to participation before cluster randomization. Approximately 1,800 adults with unresectable HCC will be prospectively enrolled. Participants must be eligible for and intended to receive first-line ICI-based systemic therapy. The clinical determination that a participant is eligible for ICI-based therapy must be made before generation or disclosure of any AI output. Consecutive potentially eligible patients will be screened at each participating hospital using identical prespecified criteria. Eligibility will be confirmed through a centralized process before any AI output is generated or disclosed and before the final first-line regimen is selected. Participants will be centrally registered only after eligibility confirmation and informed consent. All screened patients and reasons for non-enrollment will be recorded to assess the completeness and representativeness of recruitment. AI-Assisted Treatment-Selection Strategy At hospitals assigned to the AI-assisted strategy, prespecified pretreatment information available at the time of participant enrollment will be entered into the locked AI clinical decision-support system. Only variables collected before treatment initiation will be used. The system will generate patient-specific estimates and a comparative ranking of prespecified, clinically appropriate, locally available, and guideline-concordant first-line ICI-based treatment options. Clinical eligibility, contraindications, and feasibility of the candidate treatment options will be assessed before the AI output is generated. Treatments that are contraindicated or otherwise clinically inappropriate for an individual participant will not be presented as eligible recommendations. The AI output will be provided to the treating multidisciplinary team as decision support and will not replace clinical judgment, safety assessment, or shared decision-making. Treating clinicians and participants will retain responsibility for the final treatment decision. Clinicians may select a regimen other than the highest-ranked AI recommendation when clinically appropriate. The AI recommendation, final treatment selection, recommendation concordance, and prespecified reasons for deviation from the recommendation will be prospectively recorded. Usual-Care Treatment-Selection Strategy At hospitals assigned to the usual-care strategy, participants will undergo the same screening, eligibility assessment, informed-consent process, central registration, baseline assessment, and prospective follow-up as participants at AI-assisted hospitals. The treating multidisciplinary team will not receive or have access to the AI output. First-line treatment will be selected through the hospital's usual multidisciplinary decision-making process, taking into account contemporary clinical guidelines, clinical characteristics, contraindications, treatment availability, clinician judgment, and participant preferences. The control group is a concurrent, prospectively enrolled usual-care group and is not a historical, external, or retrospectively constructed comparison cohort. Treatment and Follow-up The trial evaluates a treatment-selection strategy and does not mandate the use of an investigational drug. Participants in both groups will receive approved, guideline-concordant standard-of-care treatment selected from the prespecified trial-eligible ICI-based regimens. Eligibility criteria, baseline assessment requirements, imaging frequency, follow-up procedures, criteria for treatment modification or discontinuation, subsequent-treatment documentation, and outcome definitions will be harmonized between the two randomized groups. Imaging assessments will follow the same prespecified schedule, anchored to participant enrollment, irrespective of the assigned strategy, treatment selected, treatment discontinuation, or clinical response. Whenever clinically appropriate, scheduled outcome assessment will continue after treatment discontinuation until documented disease progression, death, withdrawal of consent, or the end of follow-up. Model Locking, Deployment, and Auditability Before enrollment of the first participant, the final model parameters, eligible input variables, variable definitions, preprocessing procedures, handling of missing inputs, treatment-eligibility rules, output format, decision rules, thresholds, and software version will be finalized and locked. The deployed model and software will be documented using prespecified version identifiers, timestamps, and a reproducible code or model hash. Trial outcome data will not be used to retrain, recalibrate, fine-tune, or otherwise modify the model during the primary trial period. Any necessary software maintenance that does not alter the model, input processing, treatment ranking, or decision logic will be documented and version controlled. System access, input data, missing-data handling, generated outputs, timing of recommendation disclosure, clinician access, final treatment selection, recommendation concordance, and reasons for clinician override will be recorded in prospective audit logs. Hospitals assigned to usual care will not have access to the AI system during the primary trial period. Outcomes The primary hypothesis is that the AI-assisted treatment-selection strategy is superior to the usual-care treatment-selection strategy with respect to progression-free survival. The primary outcome is progression-free survival, measured from the date of prospective participant enrollment to the first occurrence of radiographic disease progression or death from any cause, whichever occurs first. Radiographic progression will be assessed according to Response Evaluation Criteria in Solid Tumors version 1.1 (RECIST 1.1) by blinded independent central review. Modified RECIST for HCC may be used in prespecified secondary or sensitivity analyses. Imaging data submitted for central review will be de-identified. Central imaging reviewers will be unaware of cluster assignment, AI recommendations, treatment-selection rationale, and treating hospital and, where feasible, the treatment regimen received. A common imaging schedule and prespecified review procedures will be used in both groups to minimize differential surveillance and outcome-assessment bias. Overall survival is the key secondary outcome. Other secondary outcomes include objective response rate, duration of response, disease control rate, time to treatment failure, treatment exposure, treatment modification and discontinuation, subsequent anticancer therapy, adverse events, patient-reported quality of life, and health-care utilization. Implementation outcomes include the proportion of participants for whom an AI recommendation is successfully generated, recommendation acceptance, concordance between the highest-ranked recommendation and the treatment received, reasons for clinician override, time required for treatment selection, workflow integration, and variation in implementation across hospitals. Prespecified exploratory analyses will evaluate prospective model calibration, treatment-effect heterogeneity, and performance across clinically relevant patient and center subgroups. These analyses will not be used to modify the locked model during the primary trial period. Statistical Analysis The primary analysis will follow the intention-to-treat principle. All prospectively registered participants will be analyzed according to the randomized assignment of their hospital, regardless of whether the AI system was accessed, whether an AI recommendation was successfully generated, whether the recommendation was followed, or which treatment was ultimately received. The primary treatment effect will be estimated using a prespecified time-to-event model that accounts for clustering of participants within hospitals. The analysis will incorporate the cluster-randomized design and prespecified cluster-level and participant-level prognostic factors. Statistical inference will account for the number of randomized clusters, intracluster correlation, variation in cluster size, and potential center-level heterogeneity. Prespecified sensitivity analyses will evaluate the effects of missing outcome data, treatment discontinuation, loss to follow-up, cluster withdrawal, deviations from the assigned decision strategy, and alternative assumptions regarding censoring. Per-protocol and as-treated analyses may be performed as secondary supportive analyses but will not replace the primary intention-to-treat analysis. Oversight Trial conduct, participant safety, data quality, protocol adherence, model-version integrity, and completeness of outcome ascertainment will be centrally monitored. An independent data monitoring committee will review accumulating safety and efficacy information and make recommendations regarding trial continuation, modification, or early termination in accordance with a prespecified charter. The committee will operate independently of the investigators responsible for model development, trial management, treatment decisions, and primary statistical analysis.
Study Type
INTERVENTIONAL
Allocation
RANDOMIZED
Purpose
TREATMENT
Masking
SINGLE
Enrollment
1,800
A locked AI clinical decision-support system will analyze prespecified pretreatment information and provide the multidisciplinary team with patient-specific estimates and a comparative ranking of clinically appropriate first-line ICI-based treatment options. The recommendation is advisory and does not replace clinical judgment or shared decision-making.
First-line ICI-based treatment will be selected through the hospital's usual multidisciplinary decision-making process without access to the AI clinical decision-support system. Participants will otherwise undergo the same eligibility assessment, follow-up schedule, and outcome ascertainment as participants in the AI-assisted arm.
Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology
Wuhan, Hubei, China
Progression-Free Survival Assessed by Blinded Independent Central Review Using RECIST Version 1.1
Progression-free survival is defined as the time from prospective participant registration to the first radiographically documented disease progression according to RECIST version 1.1, as determined by blinded independent central review, or death from any cause, whichever occurs first. Participants without either event will be censored at the date of their last adequate radiographic tumor assessment. Imaging assessments will follow the same prespecified schedule in both study groups.
Time frame: From participant registration until radiographic disease progression or death from any cause, whichever occurs first, assessed up to 27 months
Overall Survival
Overall survival is defined as the time from prospective participant registration to death from any cause. Participants who are alive or whose survival status is unknown at the time of analysis will be censored at the date they were last known to be alive. Vital status will be ascertained using harmonized follow-up procedures in both study groups.
Time frame: From participant registration until death from any cause, assessed up to 44 months
Objective Response Rate Assessed by Blinded Independent Central Review Using RECIST Version 1.1
Objective response rate is defined as the proportion of participants with measurable disease at baseline who achieve a confirmed best overall response of complete response or partial response according to RECIST version 1.1, as determined by blinded independent central review. Responses must be confirmed by a subsequent assessment at least 4 weeks after initial documentation. Participants without an adequate post-baseline tumor assessment or without a confirmed response will be classified as non-responders.
Time frame: From participant registration through the last tumor assessment before disease progression or initiation of new anticancer therapy, assessed up to 27 months
Duration of Response
Among participants who achieve a confirmed complete response or partial response, duration of response is defined as the time from the first documented response that is subsequently confirmed to the first radiographically documented disease progression according to RECIST version 1.1, as determined by blinded independent central review, or death from any cause, whichever occurs first. Participants without either event will be censored at the date of their last adequate radiographic tumor assessment before initiation of new anticancer therapy.
Time frame: From the first documented response that is subsequently confirmed until radiographic disease progression or death from any cause, assessed up to 27 months
Disease Control Rate
Disease control rate is defined as the proportion of participants with measurable disease at baseline whose best overall response is a confirmed complete response, confirmed partial response, or stable disease maintained for at least 6 weeks after participant registration according to RECIST version 1.1, as determined by blinded independent central review. Participants without an adequate post-baseline tumor assessment or without documented disease control will be classified as not achieving disease control.
Time frame: From participant registration through the last tumor assessment before disease progression or initiation of new anticancer therapy, assessed up to 27 months
Time to Treatment Failure
Time to treatment failure is defined as the time from prospective participant registration to failure of the initially selected first-line systemic treatment strategy, defined as permanent discontinuation of all active components of the initial first-line therapy, initiation of a new systemic anticancer regimen, or death from any cause, whichever occurs first. Reasons for treatment failure, including disease progression, toxicity, participant preference, and clinician decision, will be prospectively recorded. Participants without treatment failure will be censored at the date of their last documented assessment while receiving the initial first-line treatment strategy. Participants who do not initiate the selected regimen will have treatment failure recorded on the date the decision not to initiate treatment is documented.
Time frame: From participant registration until failure of the initial first-line treatment strategy or death from any cause, assessed up to 27 months
Percentage of Participants With Grade 3 or Higher Treatment-Emergent Adverse Events
The percentage of participants in the safety population who experience at least one grade 3 or higher treatment-emergent adverse event according to the National Cancer Institute Common Terminology Criteria for Adverse Events version 6.0 will be assessed. The safety population includes all registered participants who receive at least one dose of first-line systemic therapy. A treatment-emergent adverse event is an adverse event that first occurs or worsens in severity after initiation of the selected first-line treatment, regardless of causal attribution. Adverse events will be collected and graded using harmonized procedures in both study groups.
Time frame: From the first dose of first-line systemic therapy through 90 days after the last dose, assessed up to 44 months
Change From Baseline in Patient-Reported Global Health Status and Quality of Life at 6 Months
Patient-reported quality of life will be assessed using the Global Health Status/Quality of Life scale of the European Organisation for Research and Treatment of Cancer Quality of Life Questionnaire Core 30, version 3.0 (EORTC QLQ-C30). Scores will be calculated according to the EORTC scoring manual and transformed to a scale ranging from 0 to 100, with higher scores indicating better global health status and quality of life. Change will be calculated as the 6-month score minus the baseline score; a positive change indicates improvement and a negative change indicates deterioration. Questionnaire administration procedures and assessment time points will be harmonized between the two study groups.
Time frame: Baseline and 6 months after participant registration
This platform is for informational purposes only and does not constitute medical advice. Always consult a qualified healthcare professional.