What Are The Most Urgent AI Challenges In Clinical Trials (And How To Overcome Them)
Artificial intelligence (AI) is currently employed everywhere in drug development, and clinical trials are no exception. From speeding up patient recruitment to monitoring participants in real time, AI has made research faster, smarter, and less stressful. But let’s be honest: putting AI into practice is not as simple as plugging it in and pressing the “go” button.
This article explores the most pressing hurdles researchers face and how sponsors, CROs, and technology providers can address them.
1. Data Fragmentation and Quality Issues
Picture a trial participant. A trial participant’s smartwatch tracks their heart rate. The lab sends back blood test results. The hospital stores medical history in its electronic health record (EHR). Genomic data arrives from yet another system. None of these sources “speaks” the same language. For an algorithm, it’s like trying to follow five conversations at once. Confusing and unmanageable. Since AI models need clean, consistent data to learn, this kind of heterogeneity quickly becomes one of the biggest barriers to using AI in trials [1].
How to fix it:
Build in data standards and quality checks from the start. That means harmonization pipelines (with humans in the loop to catch oddities) and clinical data management tools flexible enough to integrate multiple formats without losing traceability.
2. Patient Recruitment and Diversity
AI promised to take the pain out of patient recruitment by automatically matching patient data with trial criteria. Sounds great. Until you realize that the limited inclusion of underrepresented populations in training datasets can bias algorithms. This further reinforces existing inequities in access to better medical treatment. For example, an AI trained mostly on white Caucasian male patient records might overlook tumor changes in groups like Black women patients.
Norori and colleagues curated a list of notable examples of impactful biases in AI training for medical purposes [2]. One of them is particularly noteworthy, as it has a significant influence on clinical trial results.
The majority of participants in clinical trials are often male, fall within a limited age range, and share similar ethnic backgrounds. Algorithms trained on such biased data may deepen bias in the recruitment process [3]. This lack of diversity in the participant pool can further lead to biased outcomes that may not be generalizable to the broader human population.
Figure 1. Sources of bias in AI systems used in clinical trials and pathways through which these biases propagate into model outputs.
How to fix it:
Be transparent and audit regularly. For the data collection step, invite communities that are frequently overlooked, such as Indigenous peoples, people with disabilities, the LGBTQ+ community, and immigrants.
This guarantees the models will work for them. It also allows scientists to validate whether the research questions they ask are relevant for the people traditionally underrepresented in science [2].
3. Regulatory and Ethical Uncertainty
Regulation around AI in the health domain still feels… unready. How much transparency is “good enough”? How do you balance patient privacy with data-hungry algorithms? Agencies are paying attention and developing regulations, but global rules aren’t aligned yet [4,5]. That uncertainty makes sponsors cautious.
Moreover, if the decision-making process of artificial intelligence is not transparent, it is difficult to determine what is really behind an incorrect prediction. Not knowing the reasons behind an AI’s wrong recommendation precludes making informed decisions or linking unintended effects to their causes. Without explainability, it’s challenging to attribute whether an error stemmed from the data, the model’s design, or a human oversight [6].
How to fix it:
Talk to regulators early, especially if you’re using AI to assess endpoints or design adaptive trials. Working with seasoned bioinformatics partners also helps keep everything in line with GCP, GDPR, and HIPAA. Moreover, explainability, as we discuss in the next point, can help demonstrate compliance and accountability.
4. Lack of Transparency and Explainability
Imagine an AI tool recommending that a patient should be excluded from a trial with zero explanation why. That’s a dealbreaker for clinicians and regulators. Black-box models might deliver strong performance, but they can’t build trust if no one understands their logic, especially when the stake is patient safety [7].
For instance, an AI might detect a spurious correlation in image data (e.g., a hospital’s specific scanner artifact) rather than actual pathological signs. Without explainability, the physician cannot verify the basis of the diagnosis. It is also challenging to identify and correct errors in AI logic or underlying data. This can lead to persistent flaws in the system that continue to negatively affect patient outcomes.
Another problem is that medical knowledge and disease patterns are constantly evolving. A non-transparent AI model might fail to adapt to new strains of viruses or emerging diseases because its internal logic cannot be easily updated or re-evaluated by human experts, potentially leading to outdated or incorrect recommendations.
How to fix it:
Opt for interpretability when developing or choosing AI software for clinical trials. Explainable frameworks are essential for making trial analytics credible. Hybrid approaches that balance performance with transparency (such as combining statistical methods with machine learning) can deliver accuracy without the opacity.
Explanations delivered by an AI model need to be context-aware, providing different levels of detail or types of information depending on the user and the clinical scenario. So these tools must be designed with end users (clinicians) involved in the process to ensure explanations are relevant, actionable, and presented in an understandable format.
Explainable AI models, like any medical intervention, require rigorous clinical validation to demonstrate their safety, efficacy, and utility in a healthcare setting. Partner with developers who implement restrictive validation strategies and have experience building tools that have passed this stringent sieve [7].
5. Integration With Existing Clinical Workflows
Here is another problem: many clinical trial IT systems were built long before AI models appeared. Trying to bolt on AI can feel like plugging a Tesla into a 1980s outlet – things don’t line up. Proprietary systems and interoperability gaps also hinder the deployment of AI models.
However, the COMPOSER sepsis prediction model, implemented in two emergency departments at UC San Diego Health, demonstrated that effective AI integration with existing EHR systems is possible [8].
How to fix it:
Choose modular, API-friendly tools. Collaborate with providers who understand the nuances of clinical IT infrastructure and can integrate AI without disrupting existing workflows.
COMPOSER Sepsis Prediction Model at UC San Diego Health Case Study
Background:
Sepsis leads to many patient deaths in hospitals worldwide. Early detection is critical, but in practice, it is often delayed. To address this, researchers at UC San Diego Health developed and tested COMPOSER (COnformal Multidimensional Prediction Of SEpsis Risk), a deep learning model trained on retrospective EHR data to predict sepsis risk in real time.
Implementation:
The integration followed a three-stage pathway:
- Exploratory model development – trained on retrospective EHR data.
- Silent trial – tested prospectively in real time without showing results to clinicians.
- Clinical deployment – fully integrated into the EHR at two emergency departments.
Integration with EHR workflow:
COMPOSER predictions were embedded directly into the Epic EHR system.
A nurse-facing Best Practice Advisory (BPA) alert displayed the sepsis risk score plus the top predictive features.
Nurses were given clear response options: indicate no suspicion of infection, confirm ongoing treatment, or notify a physician.
The advisory improved nurse-physician communication and reduced time-to-antibiotics, a key factor in sepsis survival
Results (5-month period):
- 17% relative reduction in in-hospital sepsis mortality.
- 10% relative increase in sepsis bundle compliance (timely steps such as cultures, lactate measurement, fluid administration).
- High adoption: only 5.9% of alerts were dismissed.
Supporting ecosystem:
- Explainability: COMPOSER displayed relevance scores for top predictive features, building clinician trust.
- Continuous monitoring: a dashboard tracked input data quality and model performance, with retraining protocols in place if accuracy declined.
- Stakeholder engagement: multidisciplinary planning, staff education, and feedback loops to secure end-user buy-in.
Key takeaway:
The success of COMPOSER was attributed to its holistic ecosystem, which included thoughtful EHR integration, explainability, clear clinician action pathways, continuous monitoring, and organizational buy-in. This case demonstrates that AI in healthcare can be effectively integrated into clinical workflows, supported by both infrastructure and human trust.
6. Scalability And Generalization
AI models often perform well in the specific trial where they were trained, but struggle elsewhere. A tool designed for oncology patients in Europe may not be effective for patients in Asia. Dataset shift and contextual differences frequently erode performance when models leave their original environment. Scaling is a dynamic process of adapting innovation, and without broad validation, results don’t generalize [9].
How to fix it:
Effective scaling requires two moves: external validation across diverse datasets and local customization. While interoperability standards and shared benchmarks provide a common foundation, every deployment must be adapted to local realities. This might mean tailoring interfaces for different clinical teams, adjusting data inputs to regional standards, or aligning the tool with on-site resources. Bioinformatics partners with access to global, harmonized data can help adapt models beyond their “home turf.”
How Ardigen Supports AI in Clinical Trials
Real progress comes from pairing smart algorithms with strong data practices, regulatory dialogue, and inclusive design. Done right, AI makes clinical research more reliable and more human-centered.
At Ardigen, we help our partners develop scientifically grounded, modular solutions. We use our expertise in clinical data management, multimodal integration, and interpretable model development to build AI technologies that enhance, not complicate, trial execution.
We build tools that support:
- Protocol feasibility and site selection.
- Biomarker discovery and patient stratification.
- Real-time monitoring and predictive analytics.
- Multimodal clinical data harmonization.
By combining domain knowledge with scalable bioinformatics infrastructure, Ardigen effectively supports pharma and CROs in confident and compliant AI adoption.
Learn more about how we can support AI implementation in your clinical trials.
https://ardigen.com/bioinformatics/
Author: Martyna Piotrowska
Technical editing: Ardigen expert: Anna Sanecka Duin, PhD
Bibliography
- Hong L, Luo M, Wang R, Lu P, Lu W, Lu L. Big Data in Health Care: Applications and Challenges. Data and Information Management. Data Inf Man. 2018. Dec 31;2(3):175–97. https://doi.org/10.2478/dim-2018-0014
- Norori N, Hu Q, Aellen FM, Faraci FD, Tzovara A. Addressing bias in big data and AI for health care: A call for open science. Patterns (N Y). 2021. Oct 8;2(10):100347. https://doi.org/10.1016/j.patter.2021.100347
- Lu X, Yang C, Liang L, Hu G, Zhong Z, Jiang Z. Artificial intelligence for optimizing recruitment and retention in clinical trials: a scoping review. J Am Med Inform Assoc. 2024 Nov 1;31(11):2749-2759. https://doi.org/10.1093/jamia/ocae243
Erratum in: J Am Med Inform Assoc. 2025 Jan 1;32(1):260. https://doi.org/10.1093/jamia/ocae283 - European Medicines Agency. Reflection paper on the use of artificial intelligence (AI) in the medicinal product lifecycle [Internet]. 2024 [cited 2025 Sep 16]. Available from: https://www.ema.europa.eu/system/files/documents/scientific-guideline/reflection-paper-use-artificial-intelligence-ai-medicinal-product-lifecycle-en.pdf
- U.S. Food and Drug Administration. Artificial Intelligence and Machine Learning in Software as a Medical Device [Internet]. 2021 [cited 2025 Sep 16]. Available from: https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-software-medical-device
- Doshi-Velez F, Kim B. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:170208608. 2017. https://doi.org/10.48550/arXiv.1702.08608
- Holzinger A, Biemann C, Pattichis CS, Kell DB. What do we need to build explainable AI systems for the medical domain? arXiv preprint arXiv:171209923. 2017. https://doi.org/10.48550/arXiv.1712.09923
- Kwong JCC, Nickel GC, Wang SCY, et al. Integrating artificial intelligence into healthcare systems: more than just the algorithm. npj Digit Med. 2024. 7, 52. https://doi.org/10.1038/s41746-024-01066-z
- Lukkien D, Nap H, Peine A, et al. Responsible scaling of artificial intelligence in healthcare: standardization meets customization. Ethics Inf Technol. 2025. 27, 34. https://doi.org/10.1007/s10676-025-09842-5
