Technical Summary: The Clinical Trial Pipeline Reveals the Next Wave of Artificial Intelligence in Healthcare

Problem Statement

While the prospective clinical evaluation of artificial intelligence (AI) in medicine has expanded rapidly, there is a lack of comprehensive, multidimensional analysis regarding the global AI clinical trial landscape. Existing discussions often rely on benchmark performance, retrospective validation, or high-profile demonstrations, which rarely translate to changes in clinical practice. Unlike static drug interventions, clinical AI systems are dynamic sociotechnical entities subject to continuous updates, recalibration, and workflow integration, making their prospective evaluation inherently complex. Consequently, there is an urgent need to understand which AI technologies are being tested, in which clinical contexts, at what level of autonomy, and with what degree of translational maturity to identify evidence gaps and direct future investment.

Methodology

The study conducted a systematic analysis of 8,532 AI-related clinical trials registered in ClinicalTrials.gov as of April 2026. The methodology involved a multi-stage filtering and classification process:

1. Identification and Filtering:

   • An inclusive keyword-based search strategy was applied to the entire ClinicalTrials.gov database, combining specific terms (e.g., "convolutional neural network") with broader terms (e.g., "model," "algorithm").
   • A few-shot learning approach using the Claude Sonnet 4.6 LLM was employed to classify trials. Trials with a probability of ≥95% were classified as positive; those ≤20% were negative. Ambiguous cases were reviewed by a second LLM (Gemini 3.1 Pro preview) and, if necessary, manual review.
   • Trials evaluating drugs discovered via AI but lacking explicit AI references in registry entries were excluded, acknowledging a limitation in registry-based identification.

2. Multidimensional Classification:
   Each identified trial was classified across seven complementary analytical dimensions:

   • Clinical Function: Diagnostic, prognostic, decision support, generative/language-based, agentic/semi-autonomous, and operational.
   • Data Modality: Imaging, physiological signals, lab data, omics, clinical text/NLP, structured EHR, wearables, behavioral/cognitive, patient-reported outcomes, administrative, and surgical data.
   • Clinical Specialty: Mapped to 17 domains (e.g., cardiology, oncology, rheumatology).
   • Integration and Autonomy Level: A four-level taxonomy ranging from Level 1 (silent/observational) to Level 4 (semi-autonomous/closed-loop systems capable of modifying treatment without mandatory human review).
   • Clinical Workflow Position: Prevention, screening, diagnosis, treatment selection, procedural guidance, monitoring, rehabilitation, and follow-up.
   • Translational Maturity Stage: A five-stage ladder from retrospective validation to semi-autonomous clinical operation.
   • AI Epistemic Role: Distinguishing between AI-in-the-loop (active intervention), AI-on-the-data (post-hoc analysis), and AI-out-of-the-loop (intervention design/discovery).

3. Data Extraction:
   Metadata (start dates, recruitment status) were extracted via API, while complex dimensions (specialties, modalities, autonomy) were inferred from free-text descriptions using few-shot learning guided by a predefined taxonomy.

Key Contributions

• Largest Multidimensional Analysis: This represents the largest analysis of AI-related clinical trials to date, encompassing 8,532 studies spanning over two decades.
• Novel Taxonomies: Introduction of a four-level autonomy taxonomy to characterize human oversight and a tripartite epistemic role framework to distinguish AI's function relative to the trial intervention.
• Structural Mapping: A comprehensive coordinate system allowing for the precise location of any AI trial and the identification of aggregate patterns, dominant modalities, and neglected specialties.

Key Results

Temporal Trends and Scale

• Acceleration: There has been a 3.6-fold acceleration in registrations from 2019 onward, with 80% of all trials registered in this period. The compound annual growth rate from 2015 to 2025 exceeds 20%.
• Study Design: 30.5% of trials employ a randomized controlled design. While observational studies overtook interventional ones post-2020 (likely due to the pandemic and EHR integration), 2025 shows a reversal, with interventional trials approaching parity with observational ones.
• Pipeline Status: 40.4% of the identified trials constitute an active evidence-generation portfolio, with a projected aggregate patient exposure approaching 5.3 million individuals.

Clinical Function and Task Shift

• Prognostic Shift: Prognostic AI (4,324 trials) has narrowly exceeded diagnostic AI (3,828 trials), indicating a field-wide shift from disease detection toward risk stratification and trajectory prediction.
• Treatment Gap: Treatment recommendation AI remains the least developed category (768 trials, 9%), highlighting a gap in prescriptive AI.
• Generative AI: Clinical text and NLP-based trials grew 7-fold between 2018 and 2025, coinciding with the public availability of large language models.

Data Modalities and Multimodality

• Dominance of Imaging: Imaging-based AI constitutes the largest single modality (29%).
• Multimodal Integration: 33.6% of trials integrate two or more data streams. The proportion of multimodal trials rose from 27.8% in 2020 to 36.4% in 2025. Dominant fusion architectures include imaging-physiological signals and imaging-omics.

Translational Maturity and Autonomy

• Pre-Deployment Concentration: The majority of the pipeline (38% retrospective, 21% silent prospective) is still generating algorithmic rather than clinical evidence.
• Level 4 Autonomy: Only 184 trials are classified as Level 4 (semi-autonomous/closed-loop). 68% of these are concentrated in closed-loop glucose management (artificial pancreas). Other clusters include closed-loop anesthesia, respiratory management, vasopressor control, and neurostimulation.
• Notable Exception: Trial NCT07132151 employs reinforcement learning for fully autonomous control of anesthetic drug dosing, representing a frontier beyond traditional model-predictive control.

Geographic and Specialty Distribution

• Geographic Imbalance: West Europe, North America, and East Asia account for 86% of geolocated trials. Africa, South America, and South Asia combined contribute under 5%.
• Specialty Underrepresentation: Severe underrepresentation exists in rheumatology (1.2%), hematology (1.7%), infectious diseases (1.5%), and medical genetics (0.4%) relative to global disease burden.
• Sponsorship: 83% of trials are sponsored by academic or non-profit institutions. The median enrollment is 193 participants, with 35% enrolling fewer than 100.

Significance and Claims

The paper asserts that clinical AI has completed its first major transition from retrospective algorithmic development to large-scale prospective evaluation. However, it has not yet completed the more consequential transition toward adequately powered, geographically representative, long-term outcome trials capable of supporting guideline-level recommendations.

The authors claim that the defining transitions visible in the pipeline are:

1. The near-exclusive concentration of semi-autonomous closed-loop AI in glucose management.
2. The rapid emergence of language-based multimodal systems.
3. The growing separation between AI-as-clinical-agent and AI-as-analytical-instrument.

The study concludes that addressing structural imbalances in specialty coverage, geographic representation, trial scale, and translational maturity is critical to determining whether the breadth of ongoing AI clinical investigation translates into equitable and demonstrable benefit for patients. The current landscape is characterized by a "moving target" nature of AI interventions, requiring prospective evaluation to capture robustness, workflow compatibility, and safety that retrospective studies cannot provide.