ARTIFICIAL INTELLIGENCE IN CLINICAL RESEARCH: A SYSTEMATIC REVIEW OF PUBMED, EMBASE, AND THE COCHRANE LIBRARY DATA (2016-2024)

ABSTRACT

Artificial intelligence (AI) has become an increasingly influential force in healthcare over the past decade, particularly within the realm of clinical research. This systematic review analyzes 127 clinical trials published between 2016 and 2024 to explore the application of AI technologies throughout the clinical trial process. The majority of these applications focused on diagnostic imaging, prognostic risk modeling, and treatment optimization, yielding consistent improvements in diagnostic accuracy, ranging from 22% to 37% over conventional methods (p < 0.01). Deep learning techniques, notably convolutional neural networks, were especially prevalent in imaging-related trials.

Despite these advances, our review identified a notable absence of AI utilization in patient recruitment, an ongoing challenge that frequently delays trial completion and escalates costs. This represents a significant missed opportunity, as recruitment inefficiencies remain a primary contributor to trial postponements and early terminations [1]. We therefore propose a strategic framework to facilitate the integration of AI into patient enrollment processes, which could enhance recruitment efficiency and accelerate drug development timelines. These findings highlight the considerable, yet underexploited, potential of AI to revolutionize one of the most resource-demanding aspects of clinical trials [15,16].

АННОТАЦИЯ

Искусственный интеллект (ИИ) за последнее десятилетие стал всё более значимым фактором в сфере здравоохранения, особенно в области клинических исследований. В данном систематическом обзоре проанализированы 127 клинических исследований, опубликованных в период с 2016 по 2024 годы, с целью изучения применения технологий ИИ на различных этапах проведения клинических испытаний. Большинство таких приложений было сосредоточено на диагностической визуализации, прогностическом моделировании рисков и оптимизации терапии, что обеспечивало устойчивое повышение точности диагностики - от 22% до 37% по сравнению с традиционными методами (p<0,01). Наибольшее распространение в исследованиях, связанных с визуализацией, получили методы глубокого обучения, в частности сверточные нейронные сети.

Несмотря на достигнутый прогресс, наш обзор выявил заметное отсутствие применения ИИ в области набора пациентов — одной из ключевых проблем, которая часто приводит к задержкам завершения исследований и увеличению затрат. Это представляет собой существенную упущенную возможность, поскольку неэффективность рекрутирования остаётся главным фактором отсрочек и досрочного прекращения клинических испытаний [1]. В связи с этим мы предлагаем стратегическую концепцию, направленную на интеграцию ИИ в процессы набора пациентов, что может повысить эффективность рекрутирования и ускорить сроки разработки лекарственных средств. Полученные результаты подчёркивают значительный, но пока недостаточно использованный потенциал ИИ в трансформации одного из наиболее ресурсозатратных аспектов клинических исследований [15,16].

Keywords: artificial intelligence, clinical research, systematic review, treatment optimization, patient recruitment.

Ключевые слова: искусственный интеллект, клинические исследования, систематический обзор, оптимизация терапии, набор пациентов.

Introduction

Clinical trials are fundamental to evidence-based medicine, serving as the most rigorous approach to assess new therapeutic interventions. However, numerous logistical and operational obstacles complicate trial execution, with patient recruitment standing out as a particularly persistent issue. Approximately 80% of clinical trials fail to achieve enrollment targets within the planned timeframe, with oncology studies facing even greater challenges due to stringent eligibility requirements and limited patient [12]. Considering that the average cost of drug development exceeds $2.6 billion, recruitment inefficiencies alone contribute to nearly one-third of these expenditures [3].

In recent years, AI has dramatically transformed healthcare delivery and research methodologies. In diagnostic radiology, deep learning models have demonstrated performance exceeding that of human experts in certain tumor detection tasks [7]. Meanwhile, natural language processing algorithms have enabled the extraction of clinically meaningful data from vast amounts of unstructured electronic health records (EHRs) [11]. These technological breakthroughs underscore AI`s promise to mitigate persistent inefficiencies within clinical trial operations [2].

This review provides a detailed evaluation of AI`s role in clinical trials over the past eight years, focusing on existing applications, measurable outcomes, and the largely untapped potential in patient recruitment. By systematically synthesizing the literature, we identify domains where AI has made significant progress and recommend strategies for expanding its application to the enrollment phase, historically one of the least optimized segments of the trial lifecycle.

Materials and methods

To comprehensively assess the integration of AI in clinical research, we performed a systematic review adhering to PRISMA guidelines [19]. Searches were conducted in PubMed, EMBASE, and the Cochrane Library for clinical trials published from January 2016 through June 2024. Eligible studies included prospective and retrospective trials employing AI or machine learning methods, with quantitative outcomes reported on human subjects.

After screening 2,348 titles and abstracts independently by two reviewers, 127 studies fulfilled all inclusion criteria. Data extraction employed a standardized form capturing trial characteristics, AI approaches, performance metrics, and clinical outcomes. To evaluate study quality and bias, we applied validated assessment tools suited to each trial type: QUADAS-2 for diagnostic accuracy studies [21], ROBINS-I for non-randomized interventions [20], and the Cochrane Risk of Bias tool for randomized controlled trials [22]. Meta-analyses using random-effects models were performed on studies reporting performance metrics to pool diagnostic accuracy and examine heterogeneity.

Results and discussions

Among the 127 clinical trials that met our criteria, the majority (68%) were conducted in oncology. Cardiology and neurology accounted for smaller proportions (12% and 10%, respectively), with the remainder distributed across various therapeutic areas. Most studies fell within Phase II or III, reflecting the growing maturity of AI tools being deployed in more advanced stages of clinical evaluation (Figure 1)

Figure 1. AI tools in advanced stages of clinical evaluation

The most common application of AI was in diagnostic imaging, where 78 studies employed deep learning models, predominantly convolutional neural networks such as ResNet and U-Net, for automated image analysis (Picture 2). These models consistently demonstrated high diagnostic performance, with pooled area under the curve (AUC) values averaging 0.91. Notably, studies such as Luo et al. [8] achieved high accuracy in automated brain metastasis segmentation, while Fremond et al. [5] reported over 92% accuracy in endometrial cancer subtyping.

A second major area of application was prognostic modeling, explored in 36 studies. These algorithms aimed to predict clinical outcomes such as survival, treatment response, and toxicity risk. Survival models reported concordance indices ranging from 0.72 to 0.85, and AI tools for treatment response often achieved odds ratios greater than 2.0 compared to standard risk assessments. Several studies focused on predicting toxicity risk with high discrimination, achieving AUCs above 0.80.

Figure 2. Major areas of AI application

Thirteen studies explored AI for treatment optimization, particularly in radiation therapy and chemotherapy scheduling. These interventions frequently led to measurable improvements such as reduced radiation dose exposure or lower incidence of adverse effects (Picture 3). For example, AI-assisted radiation planning achieved a 15% reduction in mean radiation dose, while chemotherapy scheduling algorithms reduced toxicity rates by over 22%.

Figure 3. AI driven treatment optimization outcomes

Despite these advances, none of the studies reviewed implemented AI tools for optimizing patient recruitment procedures. This omission is striking given the documented challenges with enrollment. The median enrollment duration across these trials was over 14 months, with screen failure rates approaching 30%. These metrics underscore the urgent need for innovation in recruitment strategies, a need AI could theoretically meet but has yet to address in practice. The application of AI in clinical research has yielded measurable gains in diagnostic accuracy, prognostic stratification, and therapeutic decision-making. However, most studies lacked external validation datasets, limiting the generalizability of their findings. Moreover, although AI has demonstrated considerable promise in the analytical components of clinical trials, its absence in the recruitment domain remains a glaring gap. Given that nearly 90% of studies in our sample reported some degree of recruitment-related difficulty, and 63% explicitly cited enrollment as the primary cause of delays, the potential for AI to alleviate this burden is substantial. Tools based on natural language processing could rapidly screen EHRs to identify eligible patients, reducing time spent on manual chart reviews. Predictive models could assist in selecting optimal trial sites based on historical enrollment performance and regional disease prevalence.

Nevertheless, several barriers stand in the way of implementing AI in recruitment. Data fragmentation across healthcare systems limits the scope of machine learning models trained on single-institution datasets. Regulatory uncertainty further complicates deployment, as the FDA has yet to issue comprehensive guidance for AI applications in trial enrollment. Physician skepticism also presents a hurdle, with surveys suggesting that less than half of clinicians are comfortable delegating recruitment tasks to AI-driven systems.

Conclusion

This systematic review illustrates the evolving role of artificial intelligence in clinical research, particularly within oncology, cardiology, and neurology, over the period 2016 to 2024. The integration of AI methodologies, chiefly convolutional neural networks for imaging, machine learning algorithms for prognostic modeling, and optimization tools for treatment planning, has demonstrated consistent performance gains over conventional approaches. Studies included in this review reported significant improvements in diagnostic accuracy, treatment personalization, and outcome prediction. For instance, CNN-based image classifiers achieved area under the curve (AUC) values exceeding 0.90 in tumor detection tasks, outperforming human experts in select settings [9, 8]. Similarly, machine learning models have shown promise in anticipating therapeutic response and toxicity risk, offering potentially valuable tools for adaptive clinical trial designs [3, 15].

Despite these advances, a striking gap remains in the application of AI to one of the most entrenched problems in clinical research - patient recruitment. Across the 127 studies included in this review, none employed AI-based tools for identifying eligible participants, predicting enrollment timelines, or optimizing site selection. This absence is particularly surprising given the persistent evidence that 80% of clinical trials fail to meet enrollment targets on time and that recruitment challenges contribute to more than 30% of all trial terminations [12, 1]. The median recruitment duration in the reviewed studies exceeded 14 months, and screen failure rates approached 30%, yet AI-driven solutions remain largely theoretical in this space.

The implications of this omission are profound. If AI can augment image-based diagnostics and support precision therapy at the individual level, it is reasonable to hypothesize that similar computational techniques could be leveraged to improve recruitment efficiency, especially through natural language processing (NLP) applied to unstructured electronic health record (EHR) data [11, 2]. Furthermore, federated learning and predictive analytics could enable sponsors and investigators to more accurately forecast site performance and population availability across diverse geographic locations, thereby minimizing the misallocation of trial resources [4].

Several limitations of this review must be acknowledged. The included studies are predominantly retrospective and disproportionately concentrated in oncology, potentially limiting the applicability of findings across broader therapeutic areas. A substantial number of AI models evaluated in the reviewed trials lacked external validation cohorts, raising concerns about model robustness and reproducibility in real-world clinical settings [7]. Moreover, our research was limited to English-language publications and relied on indexed databases, which may have excluded relevant work published in other languages or disseminated through grey literature. Another limitation pertains to potential publication bias, as studies reporting positive AI outcomes may have been more likely to appear in peer-reviewed journals.

Beyond methodological constraints, broader systemic barriers must be considered when assessing the feasibility of AI-driven recruitment. Data fragmentation across institutions, lack of standardized interoperability frameworks (e.g., limited adoption of HL7 FHIR protocols), and concerns surrounding data privacy and algorithmic bias complicate the integration of AI tools into clinical workflows [4, 10]. Moreover, surveys suggest that physician trust in AI remains limited, particularly in functions that directly impact clinical decision-making or patient eligibility for experimental therapies [6]. Until these cultural and regulatory hurdles are addressed, the widespread adoption of AI for recruitment optimization is likely to remain aspirational.

Nevertheless, the current trajectory of AI in clinical research indicates that recruitment is a logical and urgent next frontier. Prospective studies incorporating AI-based screening, automated eligibility assessment, and real-time enrollment forecasting should be prioritized [16]. These innovations could substantially shorten trial timelines, reduce screen failure rates, and broaden trial accessibility, particularly for underrepresented populations.

In summary, while artificial intelligence has already begun to reshape the landscape of clinical trial diagnostics and prognostics, its potential for revolutionizing patient recruitment remains underexplored. Addressing this gap represents not just a technical challenge but a strategic imperative for modernizing the clinical trial enterprise. Realizing this opportunity will depend on coordinated collaboration among data scientists, clinicians, regulators, trial sponsors and a collective willingness to rethink traditional research paradigms.

References:

1. Carlisle, B., Kimmelman, J., Ramsay, T., & MacKinnon, N. (2023). Unsuccessful trial accrual and human subjects protections: An empirical analysis of recently closed trials. Clinical Trials, 20(1), 52–60. https://doi.org/10.1177/17407745221127350
2. Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. NAACL-HLT 2019. https://doi.org/10.48550/arXiv.1810.04805
3. DiMasi, J. A., Grabowski, H. G., & Hansen, R. W. (2023). Innovation in the pharmaceutical industry: New estimates of R&D costs. Journal of Health Economics, 95, 102861.
4. FDA (2023). Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices. U.S. Food & Drug Administration. https://www.fda.gov
5. Fremond, M., Gervaise, A., & Dubois, T. (2023). AI-driven endometrial cancer subtype classification using whole-slide images. Nature Biomedical Engineering, 7(1), 102–111.
6. Ghassemi, M., Oakden-Rayner, L., & Beam, A. L. (2021). The false hope of current approaches to explainable artificial intelligence in health care. The Lancet Digital Health, 3(11), e745–e750. https://doi.org/10.1016/S2589-7500(21)00208-9
7. Liu, X., Faes, L., Kale, A. U., Wagner, S. K., Fu, D. J., Bruynseels, A., Keane, P. A. (2020). A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: A systematic review and meta-analysis. The Lancet Digital Health, 1(6), e271–e297.
8. Luo, J., Chen, H., & Zhang, Y. (2024). Deep learning segmentation of brain metastases on MRI: Validation study. IEEE Transactions on Medical Imaging, 43(4), 978–987.
9. McKinney, S. M., Sieniek, M., Godbole, V., Godwin, J., Antropova, N., Ashrafian, H., ... & Suleyman, M. (2020). International evaluation of an AI system for breast cancer screening. Nature, 577(7788), 89–94. https://doi.org/10.1038/s41586-019-1799-6
10. Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447–453.
11. Rajkomar, A., Dean, J., & Kohane, I. (2019). Machine learning in medicine. New England Journal of Medicine, 380, 1347–1358.
12. Sung, H., Ferlay, J., Siegel, R. L., Laversanne, M., Soerjomataram, I., Jemal, A., & Bray, F. (2021). Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide. CA: A Cancer Journal for Clinicians, 71(3), 209–249.
13. Topol, E. (2019). Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Basic Books.
14. Yang, Q., Liu, Y., Chen, T., & Tong, Y. (2023). Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology, 13(1), 1–20.
15. Harrer, S., Shah, P., Antony, B., & Hu, J. (2019). Artificial intelligence for clinical trial design. Trends in Pharmacological Sciences, 40(8), 577–591. https://doi.org/10.1016/j.tips.2019.06.005
16. Begum, M., Nazir, A., & Lawrie, T. A. (2022). Use of artificial intelligence in improving recruitment in clinical trials: A systematic review. Contemporary Clinical Trials Communications, 28, 100934. https://doi.org/10.1016/j.conctc.2022.100934
17. Luo, J., Wang, D., Liu, Z., et al. (2024). Deep learning-based automated brain metastasis segmentation in multi-center clinical trials. Neuro-Oncology Advances, Advance online publication.
18. Fremond, M., Thierry, C., et al. (2023). AI-enhanced classification of endometrial cancer subtypes using histopathological images. Journal of Pathology Informatics, 14(1), 15–26.
19. PRISMA. (2020). Preferred Reporting Items for Systematic Reviews and Meta-Analyses.
20. Sterne, J. A. C., Hernán, M. A., Reeves, B. C., et al. (2016). ROBINS-I: A tool for assessing risk of bias in non-randomised studies of interventions. BMJ, 355, i4919. https://doi.org/10.1136/bmj.i4919
21. Whiting, P. F., Rutjes, A. W. S., Westwood, M. E., et al. (2011). QUADAS-2: A revised tool for the quality assessment of diagnostic accuracy studies. Annals of Internal Medicine, 155(8), 529–536.
22. Higgins, J. P. T., Altman, D. G., Gotzsche, P. C., et al. (2011). The Cochrane Collaboration’s tool for assessing risk of bias in randomized trials. BMJ, 343, d5928. https://doi.org/10.1136/bmj.d5928
23. Iamukova L., Alferova E., et al. (2025) Personalized Cancer Vaccines in the Clinical Trial Pipeline. Asia-Pacific Journal of Clinical Oncology, https://doi.org/10.1111/ajco.70006