Clinical Research with AI
AI can accelerate search, screening, extraction, prediction modeling, and analysis. It can also scale citation errors, selection bias, data leakage, and irreproducible decisions. Research speed is valuable only when provenance, analytic code, and human adjudication remain auditable.
After completing this chapter, researchers should be able to:
- Leverage AI tools for literature review, systematic reviews, and evidence synthesis
- Critically evaluate AI-generated real-world evidence and comparative effectiveness studies
- Understand AI’s role in drug discovery, target identification, and precision medicine research
- Implement TRIPOD+AI and design-appropriate reporting guidance for AI research
- Ensure reproducibility, transparency, and ethical conduct in AI research
- Recognize the potential and limitations of AI in advancing medical knowledge
Introduction
The scale of medical publishing makes search, screening, and updating difficult. AI tools can prioritize records and extract candidate fields, but the same speed can scale citation errors, biased exclusions, and confounded analyses. The LLM evidence base illustrates the denominator problem: an LLM-assisted systematic review identified 4,609 peer-reviewed clinical LLM studies from January 2022 through September 2025, but only 1,048 used real-world patient data and only 19 were prospective randomized trials (Chen et al., 2026). Publication volume is not accumulation of clinical evidence. Reporting standards including TRIPOD+AI, TRIPOD-LLM, and STARD-AI improve transparency when applied to the correct study design. Clinical-trial design and reporting are covered separately in Clinical Trials and AI.
AI for Literature Synthesis and Knowledge Discovery
The Information Overload Problem
Medical literature grows exponentially: - PubMed and other databases add a large and continuously changing literature - Systematic-review timelines vary with topic, corpus, team, methods, and update model - Reviews can become outdated when evidence changes faster than the review process - Physicians cannot keep up with literature in their own specialty, let alone broader medicine
AI offers tools to manage this information deluge.
AI-Powered Literature Search
Traditional Search Limitations: - Keyword-based (PubMed, Google Scholar) - Requires perfect query formulation - Misses relevant articles using different terminology - High sensitivity (finds many articles) but low precision (many irrelevant)
AI-Enhanced Search:
1. Semantic Search: - Understands meaning, not just keywords - “Finds papers about X” not just “papers containing word X” - Examples: - Semantic Scholar: AI-powered search engine for scientific literature - PubMed with Best Match: Uses ML to rank results by relevance - Elicit: AI research assistant that answers questions using literature
2. Citation Network Analysis: - AI maps relationships between papers (who cites whom) - Identifies seminal papers and research trends - Tools: Connected Papers, Inciteful
3. Question-Answering Systems: - Ask natural language questions, get answers with citations - Examples: - Consensus: “What are the effects of X on Y?” returns an integrated answer from multiple studies - Scite.ai: Shows how a paper has been cited (supporting vs. contrasting evidence)
Evidence questions: - Did the evaluation measure recall, precision, time, workload, or final review conclusions? - Was the corpus representative of the intended review and was the comparator reproducible? - Did the workflow preserve records that the model deprioritized or excluded? - Human adjudication and source verification remain essential
Automated Systematic Reviews
Traditional Systematic Review Process: 1. Define research question (PICO format) 2. Search multiple databases (PubMed, Embase, Cochrane, etc.) 3. Screen titles/abstracts for relevance (2 independent reviewers) 4. Full-text review of included studies 5. Data extraction 6. Risk of bias assessment 7. Meta-analysis (if appropriate) 8. Write report
Bottleneck: Steps 3-6 are labor-intensive, taking months.
AI Assistance:
1. Abstract Screening (High Maturity): - AI trained on prior systematic reviews learns inclusion/exclusion patterns - Screens abstracts, ranks by relevance - Sensitivity and workload reduction depend on the corpus, prevalence, labels, threshold, stopping rule, and review design - Workload can fall while clinically important studies are missed, so both burden and recall require measurement - Critical: Human review of borderline cases essential Nagendran et al., 2020
Tools: - Rayyan AI: Collaborative systematic review platform with AI screening - Covidence: Cochrane-affiliated platform with ML prioritization - DistillerSR: Advanced features for complex reviews - ASReview: Open-source active learning for screening - LEADS (research): A specialized foundation model (fine-tuned Mistral-7B) for systematic review tasks. In a user study with 16 clinicians, human-AI collaboration achieved higher recall (0.81 vs 0.77) with 22.6% time savings for screening, and higher extraction accuracy (0.85 vs 0.80) with 26.9% time savings. Not commercially available; demonstrates the potential of domain-specific LLM fine-tuning (Wang et al., Nat Commun 2025)
2. Data Extraction (Moderate Maturity): - AI extracts structured data from full-text articles - Sample size, intervention details, outcomes, effect sizes - Extraction accuracy varies by field, document format, model, prompt, and adjudication rule - Works best for structured data (tables, standardized reporting) - Struggles with narrative syntheses, complex interventions - Human verification required for all extracted data
3. Risk of Bias Assessment (Low Maturity): - AI can identify presence/absence of features (randomization mentioned, blinding reported) - Cannot make nuanced quality judgments (adequacy of randomization, likelihood of selective reporting) - Current tools provide “flags” for human reviewers, not autonomous assessment
4. Evidence Synthesis (Emerging): - Large language models (GPT-4, Claude) can summarize findings across studies - Generate draft GRADE evidence summaries - Caution: May miss contradictions, overstate consistency, or misinterpret statistical significance - Human expert oversight non-negotiable
WHO’s 2025 guidance on LMMs raises a concern specific to scientific research: generative AI may undermine peer review (WHO, 2025).
Mechanisms of harm:
- Reviewers using AI: Peer reviewers delegating evaluation to LLMs, accepting AI judgments without critical assessment
- AI-generated manuscripts: LLMs producing plausible-sounding but unverified content that passes superficial review
- Fabricated citations: LLMs generating non-existent references that reviewers do not verify
- Homogenized perspectives: AI trained predominantly on high-income country literature may systematically undervalue research from underrepresented regions
Safeguards for physician-researchers:
- Disclose AI use in submissions: Many journals now require disclosure of AI assistance in manuscript preparation
- Verify all AI-generated content: Never submit AI-drafted text without thorough verification of claims and citations
- Decline AI-only review: If asked to peer review, personally read and evaluate the manuscript rather than delegating to AI
- Check citations: Verify that every cited paper exists and supports the claimed assertion
Outside clinical corpora, Epoch AI reports a sharp 2026 rise in math-arXiv AI acknowledgments, including ~12% of established-author math papers disclosing research-assistance-class use by August 2026 (Epoch AI, 2026). Treat as a disclosure-norm signal in one STEM field, not as a rate for clinical manuscripts.
Bottom Line on AI Systematic Reviews: - AI accelerates screening and data extraction (saves time) - Cannot replace human judgment on study quality and synthesis - Best use: Human-AI collaboration (AI does first pass, humans verify and finalize) - General guidance supports AI-assisted screening with human oversight, not for final inclusion decisions Topol, 2019
Knowledge Graph and Trend Analysis
AI for Research Trend Identification: - Analyzes citation networks, topic modeling, and publication patterns - Identifies emerging research areas before they become mainstream - Detects “sleeping beauties” (important papers that were initially overlooked) - Predicts future research directions
Applications: - Funding agencies use AI to identify promising research areas - Researchers identify gaps in literature - Journal editors spot emerging topics for special issues - Industry tracks competitive landscape
Tools: - Dimensions AI: Research intelligence platform - Lens.org: Patent and literature linkage - ResearchRabbit: AI-powered literature exploration
AI in Clinical Trial Recruitment and Design
Clinical-trial design, matching, recruitment, endpoint assessment, outcome prediction, and trials of AI interventions now have a dedicated evidence review in Clinical Trials and AI. The separate chapter preserves the distinctions among retrieval, prescreening, eligibility, consent, enrollment, retention, and clinical outcomes.
The Clinical Trial Recruitment Crisis
No universal percentage describes trial recruitment performance across sponsors, phases, diseases, and health systems. The dedicated chapter replaces unsupported headline estimates with study-specific denominators and measures the complete pathway from candidate identification to enrollment and retention.
AI-Powered Patient Recruitment
TrialGPT demonstrated strong retrieval and criterion-level matching performance on synthetic patient summaries and reduced screening time in a small physician study. Randomized and prospective evaluations now provide a more complete, mixed evidence base, including faster prescreening without a higher eligible proportion and AI notifications that did not significantly increase enrollment. See Patient Matching, Prescreening, and Recruitment.
Real-World Deployment: Cleveland Clinic ATTR-CM Trial Screening
This nonrandomized deployment is not retained as core evidence because it compared unequal, nonconcurrent time windows and included vendor involvement. The clinical-trials chapter prioritizes peer-reviewed prospective and randomized evidence with explicit denominators.
Eligibility Criteria Optimization: Trial Pathfinder
Trial Pathfinder retrospectively evaluated eligibility changes using records from 61,094 patients with non-small-cell lung cancer. Its modeled broadening results support empirical review of inherited exclusions, not automatic relaxation or prospective safety equivalence (Liu et al., 2021). Detailed appraisal appears under Protocol Development and Eligibility Design.
Adaptive Trial Design with AI
Adaptive trials use prespecified statistical rules and should not be relabeled as AI merely because they use computation. AI may support simulation or exploratory subgroup discovery, but estimands, error control, stopping rules, and adaptation timing remain independent design requirements. See Protocol Development and Eligibility Design.
Real-World Evidence and AI
The Promise of RWE
Real-World Evidence (RWE): - Evidence from real-world data (EHRs, claims, registries, wearables) - Observational data from routine clinical practice - Complements RCTs by providing: - Broader patient populations (real-world diversity vs. trial eligibility restrictions) - Longer follow-up - Comparative effectiveness (head-to-head comparisons not feasible in RCTs) - Faster, cheaper than RCTs
AI’s Role: - Analyze large-scale EHR data quickly - Construct matched cohorts (propensity score matching, inverse probability weighting) - Identify confounders and effect modifiers - Generate hypotheses for RCTs Beam et al., 2020
AI for Comparative Effectiveness Research
Typical Use Case: - Compare outcomes of patients receiving Treatment A vs. Treatment B in routine practice - AI identifies patients, extracts outcomes, adjusts for confounders - Generates comparative effectiveness estimate
Illustrative research questions:
1. Antidiabetic medications: - A retrospective database study could compare cardiovascular outcomes across drug classes - The protocol would need an explicit target trial, time zero, eligibility, treatment strategies, follow-up, outcome definitions, and confounder set - Agreement with an RCT would support method calibration, not prove that every new observational comparison is causal - Disagreement should trigger investigation of population, treatment, measurement, adherence, and analytic assumptions
2. Treatments during a rapidly changing outbreak: - Rapid observational analyses can generate time-sensitive signals while randomized evidence is pending - Cohort construction and outcome ascertainment remain vulnerable to time-varying confounding, immortal time, treatment selection, and changing standards of care - Later randomized evidence can agree with or contradict an observational signal - The historical lesson is methodological: speed does not neutralize bias
A 27 August 2026 Cleveland Clinic study ran a locally deployed open-source LLM over 130,509 notes from 8,366 nephrectomy patients (2009–2024) and filled 14 operative, pathology, and radiology fields (136,425 elements) against a manually curated kidney-cancer research database (Knorr, Patel et al., 2026). Overall element agreement was 97.5% (pathology greater than 98%, Cohen’s κ greater than 0.90); of the 2.5% disagreements, arbitration favored the LLM on 10 of 14 variables. This is single-center agreement with an existing research database, not a demonstration that chart review can be retired.
A multi-site preprint scored an LLM on ACC National Cardiovascular Data Registry abstraction after abstractors scoped question-specific documents: 4,214 validation observations from 25 patients, with weighted accuracy 89.6% that fell from 96.1% on Medication/Event Flag questions to 62.0% on Event Timing (Matheson et al., 2026, preprint). The model answered 99% of observations (4,171 of 4,214) and abstained on 1%; frequency weighting raised reported accuracy 5.6 percentage points above the equal-weighted question mean. Carta Healthcare employees conducted the study; 89.6% aggregate accuracy is not evidence that unsupervised registry abstraction is ready.
Fundamental Limitations of RWE (AI Does Not Solve)
Confounding: - Patients receiving Treatment A differ from those receiving B (not randomized) - Measured confounders: can adjust (age, comorbidities) - Unmeasured confounders: cannot adjust (socioeconomic status, frailty, patient preferences) - AI can only adjust for what’s measured in data - Residual confounding can remain after measured adjustment. The degree and direction are not known merely from the presence of observational data (Fewell et al., 2007)
Selection Bias: - Who gets Treatment A vs. B is non-random - Healthier patients may get newer drugs; sicker patients get older drugs - “Confounding by indication” - Propensity scores and matching reduce but do not eliminate bias
Measurement Error: - EHR data not collected for research (missing data, coding errors) - Outcome misclassification (e.g., cause of death not reliably captured) - Exposure misclassification (medication adherence unknown) - AI cannot create data that was not collected
Causality: - Observational data shows association, not causation - Bradford Hill criteria and causal inference methods (instrumental variables, regression discontinuity) help but have strong assumptions - Causal claims require a defensible identification strategy and explicit assumptions. Randomization is often the strongest available design, but well-designed observational studies can estimate causal effects when their assumptions are credible
Best Practices for AI-Generated RWE
Transparent Methods: - Report data source, cohort construction, confounders adjusted for - Sensitivity analyses (varying analytic choices) - Acknowledge unmeasured confounding
Validation: - Compare RWE findings to known RCT results (does RWE replicate RCT findings?) - External validation in independent datasets
Appropriate Claims: - Avoid causal language (“Treatment A causes better outcomes”) - Use associational language (“Treatment A was associated with better outcomes, after adjusting for measured confounders”) - Acknowledge limitations
Hypothesis Generation: - RWE can generate hypotheses, estimate effects under explicit assumptions, support safety and effectiveness questions, and inform trials; it is not automatically interchangeable with randomized evidence - Inform trial design (endpoints, subgroups, sample size) - Identify promising signals worth testing rigorously Topol, 2019
AI in Drug Discovery and Development
The Drug Development Crisis
Traditional Drug Development: - Development timelines vary by modality, indication, prior knowledge, regulatory pathway, and where the clock starts - Cost estimates vary substantially with accounting method, capital costs, failures, and portfolio selection - Attrition is high, but no single failure percentage represents every phase, therapeutic area, or sponsor. - High failure rate due to: - Wrong target (disease mechanism misunderstood) - Poor pharmacokinetics (drug does not reach target) - Toxicity (unforeseen side effects) - Lack of efficacy (does not work in humans)
AI promises to accelerate early stages and reduce attrition Schork, 2019.
AI Applications in Drug Discovery
1. Target Identification: - Goal: Find disease-relevant proteins/genes to drug - AI Approach: - Integrate multi-omics data (genomics, transcriptomics, proteomics) - Network analysis (protein-protein interaction networks) - Predict which targets are “druggable” and disease-relevant - Examples: - BenevolentAI identified baricitinib (JAK inhibitor) for COVID-19 by AI target analysis - Recursion Pharmaceuticals uses AI on cellular imaging to identify disease mechanisms
2. Lead Optimization: - Goal: Optimize molecular structure for potency, selectivity, pharmacokinetics - AI Approach: - Structure-activity relationship (SAR) modeling - Predict binding affinity, solubility, toxicity from molecular structure - Generative models suggest chemical modifications - Examples: - Atomwise uses deep learning for virtual screening (tests millions of compounds computationally) - Insilico Medicine AI-designed drug for idiopathic pulmonary fibrosis (Phase II trial)
3. De Novo Molecule Design: - Goal: Generate entirely novel molecular structures with desired properties - AI Approach: - Generative adversarial networks (GANs), variational autoencoders (VAEs) - AI “dreams up” molecules that do not exist yet - Filter for drug-like properties, synthesizability - Examples: - Exscientia designed drug for obsessive-compulsive disorder (first AI-designed drug to reach clinical trials) - Generate Biomedicines uses AI for protein therapeutics
4. Drug Repurposing: - Goal: Identify new indications for existing drugs - AI Approach: - Network analysis (drug-disease-gene relationships) - Phenotypic screening data - Real-world data mining (off-label use patterns) - Examples: - BenevolentAI: baricitinib for COVID-19 - Multiple repurposing efforts for cancer (AI identifies oncology drugs for new tumor types)
5. Predictive Toxicology: - Goal: Predict adverse effects before animal/human testing - AI Approach: - Models trained on toxicity databases (ToxCast, Tox21) - Predict hepatotoxicity, cardiotoxicity, genotoxicity from structure - Reduces animal testing, catches problems earlier - Evidence boundary: Predictive performance varies by endpoint, chemical domain, assay, and external validation. Computational toxicology can prioritize testing but does not replace the experimental evidence required for the decision
Reality Check: Hype vs. Progress
Hype: - “AI will reduce drug development time to 1-2 years” - “AI will design perfect drugs with no side effects” - “AI will eliminate need for clinical trials”
Reality: - Clinical and regulatory status changes faster than benchmark publications. - Exscientia, Insilico Medicine, and other companies provide historical examples of computationally supported candidates entering trials - Current phase, discontinuation, partnering, and authorization status must be verified in trial registries and regulator records before publication - AI accelerates early stages (target ID, lead optimization) but not clinical trials (still years) - Experimental validation is required: Computational prioritization does not establish biological activity. - Hit rates depend on the candidate set, assay, threshold, domain, and comparator - Report synthesis success, assay conditions, replication, negative results, and downstream attrition rather than a universal percentage - Agentic discovery remains lab-in-the-loop: Co-Scientist reported in vitro validation for acute myeloid leukemia repurposing candidates, and Robin reported a lab-in-the-loop workflow for dry age-related macular degeneration candidate discovery, but neither removes experimental validation or clinical testing (Gottweis et al., 2026; Ghareeb et al., 2026) - Clinical trial bottleneck remains: Safety and efficacy testing still takes years; AI does not change this - Long-term impact remains an empirical question: Value should be measured at each transition from computation to synthesis, experiment, clinical development, and regulatory use Schork, 2019
Challenges in AI Drug Discovery
Data Limitations: - Drug discovery data is sparse (millions of possible molecules, data on only thousands) - Negative data (compounds that failed) often unpublished - AI models extrapolate from limited data
Biological Complexity: - Human disease is multifactorial (AI trained on single-target assays) - Pharmacokinetics hard to predict (absorption, distribution, metabolism, excretion) - Off-target effects and polypharmacology
Validation Gap: - AI predictions are computational; require wet-lab validation - Many academic AI drug discovery papers do not validate in lab - “Garbage in, garbage out”: low-quality training data = poor predictions
Regulatory evidence: - Drug authorization evaluates the product and supporting evidence; an “AI-designed” label is not separate proof of safety or efficacy - Sponsors should verify current submission expectations for the model’s context of use, data, risk, validation, lifecycle, and documentation - In January 2026, FDA and EMA published Guiding Principles of Good AI Practice in Drug Development
Evaluating a Simulation or “Digital Twin” Validation Claim
Simulation, digital-twin, synthetic-control, and external-control claims are now evaluated in Clinical Trials and AI. The applicable standard requires a named context of use, information available at the prediction time, a prespecified error tolerance, independent validation, uncertainty analysis, and evidence for each important submodel. A headline statement that a simulation “matched” a completed trial is not sufficient evidence.
AI in Genomics and Precision Medicine Research
Genomic Variant Interpretation
Challenge: - Whole-genome sequencing produces millions of variants relative to a reference genome, with counts depending on pipeline and representation - Common population variation is usually not sufficient to explain a rare Mendelian phenotype - The number of clinically relevant variants depends on phenotype, inheritance, assay, family structure, and interpretation framework
AI for Variant Pathogenicity Prediction: - Models trained on ClinVar (database of known pathogenic variants) - Predict whether novel variant is benign or pathogenic - Features: conservation across species, protein structure impact, population frequency - Examples: - PrimateAI: Deep-learning pathogenicity predictor whose performance depends on variant class, benchmark construction, and comparator - SpliceAI: Predicts impact on RNA splicing (high accuracy for splice variants) - AlphaMissense: DeepMind model predicts missense variant effects
Clinical Use: - AI assists geneticists in interpreting variants of uncertain significance (VUS) - Reduces time to diagnosis for rare diseases - Still requires human expert review: AI provides prediction, geneticist makes final call Topol, 2019
Polygenic Risk Scores (PRS)
Goal: Predict disease risk from genome-wide common variants
AI Approach: - Integrate hundreds to millions of variants - Weight each variant by effect size - Aggregate into risk score - Machine learning optimizes weighting and feature selection
Examples: - Coronary artery disease PRS: Can stratify relative risk in evaluated populations; effect estimates depend on score, ancestry, calibration, and baseline risk - Breast cancer PRS: Can refine risk stratification but should not be equated categorically with a pathogenic high-penetrance variant - Type 2 diabetes PRS: Can contribute to risk models, with incremental value requiring comparison against clinical predictors
Clinical Applications: - Screening (identify high-risk individuals for closer monitoring) - Prevention (statin therapy for high CAD PRS) - Clinical trials (enrich for high-risk participants)
Limitations: - Ancestry bias: PRS developed in European populations perform poorly in non-European populations - Predictive value varies: Explained variance, discrimination, calibration, and clinical utility depend on disease, score, population, and comparator - Ethical concerns: Risk of genetic discrimination (insurance, employment) Obermeyer et al., 2019
Multi-Omics Integration
Challenge: - Integrate genomics + transcriptomics + proteomics + metabolomics + imaging - Traditional statistical methods struggle with high-dimensional multi-omics data
AI Approach: - Deep learning integrates multiple data modalities - Identifies molecular signatures of disease - Predicts drug response based on multi-omics profile
Applications: - Cancer subtyping: Identify molecular subtypes beyond histology - Drug response prediction: Predict which cancer patients respond to immunotherapy - Disease mechanism discovery: Reveal pathways linking genetic variants to disease
Examples: - The Cancer Genome Atlas (TCGA): AI analysis identified novel cancer subtypes with distinct prognoses - Pharmacogenomics: Models can combine genetic and clinical variables for dose prediction; comparative and clinical value depends on the specific algorithm and population Schork, 2019
Methodological Rigor and Reporting Standards
The Reproducibility Crisis in AI Research
Problem: - Many AI studies cannot be reproduced - Reasons: - Code not shared - Data not available (privacy concerns) - Insufficient methodological detail - Overfitting (model works on training data, fails on new data) - Publication bias (only positive results published)
LLM Research Quality Crisis:
A systematic review of 137 studies evaluating LLM chatbots for health advice revealed pervasive methodological problems (Huo et al., JAMA Network Open, 2025):
- 99.3% assessed closed-source models without providing enough information to identify the specific LLM used
- 65.0% used subjective means as ground truth to define successful performance
- Less than one-third addressed ethical, regulatory, or patient safety implications
- Reporting variability was considerable across studies
Why this matters: If researchers cannot identify which LLM version was tested, studies cannot be replicated. The systematic review informed development of the CHART (Chatbot Assessment Reporting Tool) statement for standardized LLM health research reporting.
Scale does not solve the quality problem. Chen et al. found that 77.3% of clinical LLM studies evaluated non-real clinical data, including simulated scenarios, exam questions, and case vignettes; at least one-quarter had sample sizes below 30 (Chen et al., 2026). For research planning, this means another benchmark paper rarely changes practice. The field needs larger studies on real patient data, prospective workflow trials, and patient-centered outcomes.
TRIPOD+AI Guidelines
TRIPOD: Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis
TRIPOD+AI: A 27-item reporting guideline for clinical prediction models developed with regression or machine-learning methods (Collins et al., 2024). TRIPOD+AI improves reporting; it does not by itself judge risk of bias or validate a model.
Key Requirements:
1. Title and Abstract: - Clearly state AI/ML model is used - Report key performance metrics
2. Introduction: - Research question and rationale - Existing prediction models
3. Methods - Data: - Data source (EHR, registry, trial) - Eligibility criteria - Sample size - Missing data handling - Data preprocessing (normalization, imputation)
4. Methods - Model: - Model type (random forest, neural network, etc.) - Hyperparameters and tuning process - Training/validation/test split - Feature selection method - Software and version
5. Results - Performance: - Discrimination (AUROC, C-statistic) - Calibration (observed vs. predicted outcomes) - Performance by subgroups (age, sex, race) - Confidence intervals for all metrics
6. Results - Validation: - Internal validation (cross-validation, bootstrap) - External validation (independent dataset from different institution/time period) - Temporal validation (trained on old data, tested on new data)
7. Discussion: - Limitations (bias, generalizability, missing data) - Clinical implications - Comparison to existing models
8. Supplementary Materials: - Code availability (GitHub, Zenodo) - Model parameters (for reproducibility) - Data availability statement (de-identified data if possible)
PROBAST+AI and STARD-AI
PROBAST+AI evaluates risk of bias, quality of development, and applicability for prediction-model studies. It complements TRIPOD+AI rather than replacing or duplicating it.
STARD-AI extends diagnostic-accuracy reporting for studies that use AI. It is appropriate when a study estimates diagnostic accuracy, not merely whenever AI appears in the methods.
Choose the framework from the study question and design. A long checklist inventory does not improve rigor if the applicable items are not actually used.
TRIPOD-LLM Guidelines
TRIPOD-LLM: Extension specifically for studies using large language models (Gallifant et al., Nature Medicine, 2025)
Developed by an 18-member steering group and expert panel through a Delphi process, TRIPOD-LLM addresses LLM-specific reporting gaps that TRIPOD+AI does not cover. Its checklist has 19 main items divided into 50 subitems. Gary S. Collins, first author of TRIPOD+AI, is among its authors.
LLM-Specific Requirements:
1. Model Identification: - Exact model name and version (e.g., “GPT-4-0613,” not just “GPT-4”) - API version and access date - Whether model was fine-tuned or used out-of-box
2. Prompt Documentation: - Full prompts used (system prompts, user prompts) - Prompt engineering iterations and rationale - Temperature and other inference parameters
3. Output Handling: - How non-deterministic outputs were handled (multiple runs, aggregation) - Post-processing of model outputs - Human review protocols
4. Reproducibility Challenges: - Acknowledgment that closed-source models may change without notice - Version locking where possible - Strategies for reproducibility despite model opacity
Why a separate guideline? LLMs differ fundamentally from traditional ML models:
- Closed-source models change silently (GPT-4 in March ≠ GPT-4 in December)
- Prompts critically affect outputs but are rarely documented
- Non-deterministic generation requires specific handling
- Traditional metrics (AUC, sensitivity) may not capture LLM failure modes
For healthcare studies that develop, tune, or evaluate an LLM, use TRIPOD-LLM together with the reporting guidance required by the study design. TRIPOD-LLM is modular: it does not replace the trial guidance in Clinical Trials and AI for randomized evaluations.
CONSORT-AI Guidelines
CONSORT-AI extends randomized-trial reporting for AI interventions (Liu et al., 2020). Use it with CONSORT 2025. The canonical explanation and framework comparison now appear in Clinical Trials and AI.
SPIRIT-AI Guidelines
SPIRIT-AI extends protocol reporting for trials involving AI interventions (Cruz Rivera et al., 2020). Use it with SPIRIT 2025. See Clinical Trials and AI for the current protocol-to-reporting sequence.
DECIDE-AI Guidelines
DECIDE-AI addresses early clinical evaluation of AI decision-support systems, including feasibility, safety, workflow, and human factors (Vasey et al., 2022). It is a reporting guideline, not evidence of clinical effectiveness. See Clinical Trials and AI.
See the Validation Accords Resources note for the consensus generative-AI validation framework call.
CONSORT-EHEALTH Guidelines
CONSORT-EHEALTH addresses web-based and mobile interventions, including access, delivery, engagement, and platform details (Eysenbach, 2011). It complements rather than replaces the applicable core and AI-specific trial guidance.
Best Practices for Reproducible AI Research
Code Sharing: - Publish code on GitHub, Zenodo, or similar platform - Include dependencies, environment specifications - Document code clearly - Provide example data (de-identified or synthetic)
Validation matched to the claim: - Use independent institutional, temporal, or population data to assess transportability when feasible - Use prospective workflow evaluation to assess behavior in practice - Use comparative patient-relevant endpoints before claiming clinical benefit - Report prespecified subgroup performance with uncertainty Nagendran et al., 2020
Preregistration: - Register study protocol before analysis (ClinicalTrials.gov, OSF) - Prespecify analysis plan, outcomes, subgroups - Reduces selective analysis and makes deviations from the planned protocol visible
Transparent Limitations: - Acknowledge bias (selection bias, measurement bias, missing data) - Discuss generalizability limits (which populations, settings) - Describe failure modes (when does model perform poorly?) - Avoid overstating clinical utility
Ethical Review: - IRB approval for human subjects research - Data use agreements - Address privacy and consent - Equity and fairness analysis (performance by demographics)
Clinical Bottom Line
Continue Reading
- Clinical Trials and AI for protocol design, matching, recruitment, endpoints, outcome prediction, and trials of AI interventions
- Evaluating Clinical AI for appraisal frameworks and endpoint boundaries
- Clinical Data and AI for provenance, missingness, leakage, and transportability
- AI in the Life Sciences for evidence-calibrated emerging research capabilities
- Vendor Evaluation Framework for due diligence on model and simulation claims