AI-Driven Patient Recruitment Strategies for Healthcare Clinical Trials

Abstract

Patient recruitment has been one of the significant issues that have affected clinical research for the following reasons: Over 80% of the trials suffer delays in the recruitment of patients for the clinical trials and 30% of the trials do not recruit patients at all. The conventional methods of recruiting patients are from word of reference from doctors, newspaper and television advertisements, as well as manual searches through patient's records are not only time consuming but also have low return rates. Overall, recruitment has been one of the most significant challenges that companies have been facing, particularly due to the high costs involved in the process and time-consuming methods such as referrals and word-of-mouth recruitment. This paper discusses the concepts, approaches and practical use of AI technologies in patient recruitment platforms. These systems use Natural Language Processing (NLP), Machine Learning (ML) and Explainable AI (XAI) technologies, self-learn clinical trial eligibility criteria, search structured and unstructured Electronic Health Records (EHRs) and match eligible patients in real time. ACTES, for instance, generates significant increases in recruitment rate, precise matching and clinician satisfaction compared to TrialGPT. Similarly, incorporating AI with the Decentralized Clinical Trial (DCT) model opens participation to other underserved patients. However, issues like algorithmic bias, data privacy issues and generalized use across different healthcare systems exist. The paper highlights the future interests below, where federal learning, synthetic data generation and conformity with global policies should be attained to ensure AI applications' safety, ethics and value in clinical trial recruitment.

Keywords: Artificial Intelligence, Patient Recruitment, Clinical Trials, Machine Learning, Electronic Health Records (EHR), Natural Language Processing (NLP)

1. Introduction

A clinical trial is the main foundation for medical advancement. It is how new medications, treatments and diagnostic tools can be tested and approved by the general population. However, patient recruitment is still one of the most significant issues of clinical research across the country and the globe at large. Specifically, studies show that about 80% of clinical trials fail to meet the enrollment timelines and about 30% of trial locations do not enroll a single patient or enroll fewer than they should^1-3. These problems not only increase the costs and time needed for conducting any particular trial but also compromise the statistical relevance of the outcomes. These include word of mouth through the physicians, newspaper and magazine adverts and other clinical affiliations and are general in their approach and time consuming. More recently, AI has risen as a possible candidate to address these problems given its data-driven solutions in identifying better ways of recruiting patients. AI techniques, including machine learning, Natural Language Processing (NLP) and predictive analytics, can analyze large sets of structured and unstructured data from Electronic Health Records (EHRs), mHealth monitoring devices, insurance claims history and social media. These technologies make it possible to quickly identify patients targeted by the trials, with an opportunity to target clinical characteristics associated with conditions and behavioral, geographic and socio-demographic factors among patients.

AI systems may also promote intermittent contact with the targeted participants by using chatbots, personalized messages and changing interfaces that boost the responsiveness of their communication systems and minimize dropout rates while making them more inclusive. AI-based solutions can be enormously helpful in boosting the speed of patient recruitment at the same time as they ensure higher patient engagement and more diversification in patient recruitment with a focus on conditions beyond common ones or benefiting specific populations. Ethical, legal and operational considerations are also associated with procuring and implementing AI in the recruitment process. These risks include patients’ data protection, algorithmic risks and patient information concerns which should be well addressed to avoid regulatory and legal concerns. The purpose of this paper will, therefore, be to assess how Artificial Intelligence technologies are revolutionizing the approach to patient recruitment in clinical research. They address some of the main AI methodologies, powerful implementation examples, advantages and some possible disadvantages. Its purpose is to contribute to the existing knowledge for researchers, clinicians and policymakers to understand how to optimize clinical trials using the application of artificial intelligence.

2. Background and Related Work

2.1. Overview of patient recruitment in clinical research

Recruitment of patients is a core step in clinical research and is one of the most complex, time-consuming and resource-consuming activities of clinical trials. Recruitment in the past was accomplished through word of mouth from physicians, newspaper ads concerning the clinical trial, radio/TV advertising and face-to-face screening. While these approaches were somewhat successful on a small scale, they resulted in a whole wagonload of problems, such as delays, high operational costs and low recruitment rates. As per some analysis, it is seen that about 80% of clinical trials face some kind of delay due to problems in the recruitment of participants and about 10-30% of clinical trials fail only due to this reason^4-6. This is even more so when the trials are for rare diseases or highly specialized cancer research where the number of patients eligible for the trial is small and spread over a large geographical area. These are some of the reasons that have led to the rise of Decentralized Clinical Trials (DCTs) as a new approach that provides remote monitoring, telemedicine and home-based data collection. They increase the ease of access to the practitioners by the patients and vice versa, increasing recruitment and retention rates. However, eradicating unfavorable indicators is impossible today; certain critical issues still exist. Restriction criteria, practical troubles and regional disparities remain issues of concern, which restrict the diversities of clinical research subjects and their generalizability.

2.2. Previous technological approaches

In addressing these drawbacks, early technologies strived to bring a new form of innovation to the recruitment approach. Patient registration in the web-based system facilitated expressing interest in the clinical trials and pre-screening the responses for basic exclusions. Electronic Health Records (EHRs) also helped track the candidates through diagnosis codes, medication and clinic history and enabled integration with automation. Social media accounts also played a significant role in configuring timely and relevant advertisements regarding a particular demography or disease group. Machine Learning (ML) stepped up with even more developments, allowing screening using large sets as indices such as genomics, imaging or even prodigious written notes to screen eligibility. For instance, there are large language models like TrialGPT, which can analyse a few thousand clinical trials and make a rational basis for matching the trials to the patients. Apart from increasing the accuracy of recommendations, these models also offer the ability to explain the recommendations they came up with. NLP has also enhanced the ease of understanding medical terms and translating the inclusion/exclusion criteria to make them easily understandable for potential participants. Moreover, AI-driven outreach platforms target the potential applicants in different channels like video ads or chatbot interaction depending on their behavior, which again increases the reach and the conversion rates.

2.3. Limitations of existing systems

The existing AI-based applications used in the recruitment process have certain drawbacks. There is also the concern about data validity and the possibility of some vital data sets not being included. This is because the records collected and analyzed by artificial intelligence could be siloed, a fraction of the overall picture or a portion of the data that skew general tendencies towards specific patients or groups. Such data issues are more devastating while forming new patients because it increases the difficulty of including minority or underrepresented patient populations, deepening health disparities. Ethical concerns also loom large. Given that AI systems are trained on data sets, they can replicate inherent bias in including particular participants. Interpretability of complex AI models and their decision-making mechanisms are subjected to major criticisms regarding the issue of transparency and accountability or fairness. Another critical issue is scalability. Most AI-based recruitment tools have previously been interrogated and validated only in a healthcare organization or among academic personnel. Their currently learnable algorithms may not generalize well for other geographic areas, different patient samples or other trials, thus restricting their versatility. Last but not least, it is important to note that, despite the possible automation of several activities in the recruitment process, human intervention is still needed to understand specific selection criteria and the final selection of patients for the trial. Such manual validation reduces the overall automation level and, therefore, the maximum level of efficiency that can be reached.

3. Methodology

3.1. Data collection and preprocessing

The role of data quality and source can help build foundational components of an AI-based system for patient recruitment^7-10. The main approach mentioned for this type of research is Merging Structured/Unstructured Data, extracting data from EHRs, clinical trial registries, open genetic databases and social media accounts. Structured data encompasses demographic data, diagnosis code (as ICD-10), prescription records, lab information and clinic visits. Unstructured data consists of doctor’s notes, x-rays, patient histories and stories posted on the World Wide Web by patients. Also, to meet the requirements of data protection principles, all data collected were scrubbed for patient identifiers to remove any identifiable information as favored by the HIPAA and GDPR principles. The pre-processing involved in the study involved managing missing values, conceptualising concepts using reference terminologies such as SNOMED CT and RxNorm and normalising text inputs. The raw text was lemmatized and analyzed by applying named entity recognition in order to obtain data such as symptoms, treatments and disease progression. They were instrumental in improving the quality of the data as well as making inputs suitable for feeding downstream AI models.

3.2. AI models used (e.g., NLP, ML, Deep Learning)

AI models were employed at various strata of patient recruitment, starting from early stages, mid-stages and later. Natural language processing models were used to process and interpret qualitative data, especially clinical notes and trial descriptions. Indeed, modern transformer-based models like BERT and BioBERT were used to derive semantic meaning from the medical texts and map them to the patient eligibility criteria for the trials. Consequently, major machine learning algorithms such as logistic regression, decision trees and random forests were used to determine the eligibility score for patients using past EHR patterns. Other deep learning methods like Convolution Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) were implemented for the strong correlations, especially if the network was related to patient records or imaging data over time. Areas such as ensemble learning techniques that enabled multiple models to generate outputs assisted in getting better all-rounded approaches to the predictions. To ensure interpretability and to give way to ethical control over these model decisions, different XAI methods like SHAP (Shapley Additive exPlanations) were incorporated.

3.3. Feature engineering and selection

Feature selection was critical for better model performance and giving useful results from a clinical perspective. The structured electronic data consisted of age, gender, laboratory, diagnostic code, comorbidity risk scores and medications and clinical encounter frequencies. The NLP pipelines extracted symptoms’ timeline, disease staging and treatment compliance markers from the unstructured data. To reduce overfitting and increase the efficiency of the computation, dimensionality was reduced using Principal Component Analysis (PCA) and recursive Feature Elimination (RFE). Consultation with clinical experts was also undertaken to confirm the relevance of features and rank variables closely related to trial inclusion criteria. Moreover, attention techniques within deep learning were used to scale the features according to their relative importance in a certain context, paying particular attention to the patient data in the time series format.

3.4. System architecture and workflow

Figure 1: AI-Driven Patient Recruitment Architecture.

The system's core is the AI Recruitment Engine which actively recruited both from the integrated healthcare systems and external sources. Integrate with clinical data like clinical data repositories, Electronic Health Records (EHRs), patient-derived data from mobile applications^11-14, census data, registries and population health data. This broad data intake provides a plethora of details important for each patient that is necessary to narrow the likelihood of the patient meeting the requirements for clinical trials. The principle of operation is in the AI Recruitment Engine, which triggers the processing from the Data Preprocessing Module. This module cleans inputs to remove noise and inconsistencies so as to improve compatibility and usability in the following modules. Secondly, using Natural Language Processing (NLP), the Eligibility Criteria Parser converts trial eligibility criteria typically stated in medical language into machine-readable formats. This parser maps eligibility rules defined in encoded form with patient profiles and uses feature extraction from structured and unstructured data. These structure features are then fed into a Machine Learning Classifier, with ML and DL models developed to predict trial enrolment eligibility.

The classifier produces scores for each patient in terms of the probability of satisfying the inclusion and exclusion criteria. They flow into the Patient Matching Engine, where they are sorted according to their priority to become the final list of patients. A Recruitment Optimization Module also categorizes patients and plans the timely assignments of each according to site and each investigator’s availability or workload as well as geographical location to enhance the overall recruitment process. The developed system's fairness, explainability and transparency are ensured by a specific Explainability Layer (XAI Module) and Transparency Layer. These components make the model’s outputs understandable by the clinicians and the trial coordinators, thus eliminating trust issues and ethical dilemmas. Also, the Clinician & Trial Coordinator Interface includes real-time recruitment templates and manual interventions for human control. The alerts raise concerns and create exemptions so that human intervention can fix things where automated processes go wrong in trials. Therefore, the Security & Governance Layer covers the architecture and ensures compliance with such concerns as HIPAA or GDPR, as well as controlling access to data through reliable control and logging. Moreover, it contributes to increased recruitment efficiency and compliance with ethical and legal requirements, making the approach integral to modern clinical trials.

3.5. Recruitment criteria optimization

Optimizing recruitment criteria is essential in increasing clinical trial enrolment and diversity in the same process. Usual criteria of inclusion are set very restrictive due to the interference of other factors, which affect the results and maintain an appreciable level of statistical analysis. However, such highly selective criteria lead to low recruitment, longer trials and lesser external validity of the study findings. Computer-based systems in the context of recruitment can create the powerful possibility to make these criteria dynamic and updated with the help of a database and prognosis, which enables a researcher to always find the golden middle ground between highly formulated research questions and the possibility of actually carrying them out in the real social practice. Optimization in artificially intelligent systems starts with evaluating other large datasets gathered from prior clinical trials, observational studies and Electronic Health Records (EHRs). Statistical analysis helps define which criteria really influence the results and which only distortions can hinder enrollment. For instance, this may uncover that excluding patients with minor co-morbid conditions does not significantly affect the trial results while reducing the number of patients available. They can then advance such inclusion and exclusion rules to reflect the necessary tradeoffs of data analysis in a more informed way.

NLP technologies can parse through trial descriptions and match them to patients by demographics and geographic dispersion. This shows discrepancies between the eligibility language and the reality prevailing within society. Therefore, recruitment systems can suggest changes or develop other wording that is more likely to align with how these clinical conditions are documented in real-life practice to help achieve high match rates without sacrificing study quality. The optimization process should also be a part of the patient matching engine and provide feedback sent back to the optimization process. If many potentially eligible patients are not selected then notification is done to the trial coordinators and other clinicians from the dashboard. This makes it possible for an eligibility threshold for PNP qualification upgrades to be adjusted in the continuous learning process. It means that they can be implemented in the course of the trial. This characteristic is significant since adaptive trial designs are about flexibility and patient involvement. Optimizing recruitment criteria is a vital component when it comes to achieving early and faster recruitment, as well as the achievement of good sample diversities and, ultimately, studies’ external validity. When integrated into this process, the recruitment engine is no longer only a screening mechanism but a powerful AI ally within the trial design process that makes decisions that impact the effectiveness and applicability of clinical research to society.

4. Implementation and Case Study

Figure 2: Real-Time Integration with EHR Systems.

4.1. System design and tools

The modern AI systems embedded in patient recruitment comprise various tools, including Machine Learning (ML) and Natural Language Processing (NLP). These systems have been developed to compare large amounts of patient data and accurately identify patient matches to specific clinical trials^15-18. Tools such as TrialGPT use these LLMs to read past trials, cross-check the eligibility criteria with patient records and develop a list of suitable patients, plus the reasoning process behind each step. Other platforms, including myTrialsConnect and Infiuss Health’s AI-driven Study Finder, amplify the recruitment process even for prescreening, outreach and scheduling. These tools have the potential to relieve stress on clinical staff and increase trial durations without a negative impact on the quality of trials.

4.2. Integration with EHR and clinical systems

Integration with Electronic Health Record (EHR) systems requires integration with the various artificial intelligence recruitment approaches and tools. AI algorithms can check for eligibility on their own as they are constantly stimulated by seamless access to real-time EHR data. These algorithms are able to extract data from structured fields like lab values and medication lists as well as unstructured clinical notes to get a full picture of a patient. One notable example is the ACTES system put in place at the Cincinnati Children’s Hospital. Implemented as a module directly linked to the actual activity of a pediatric emergency department, ACTES incorporates modules for real-time patient identification using the analysis of EHR data in real-time. This sort of integration not only eliminates chart review but also makes sure that recruitment chances are not removed due to lack of time or maybe due to inattention.

4.3. Real-world case study

The use-case of the ACTES system is an attractive example of AI implementation for clinical recruitment in practice. In a 12-month evaluation period, the system was credited with a screening time reduction of 34% and an 80% usability rating among the clinical staff. The outcomes are the testimony of how effective the system has been in achieving high operational efficiency and satisfaction among users. In a related case, AI was employed to help conduct a Phase II clinical trial on Chronic-Induced Corneal Pain (CICP) in the case of laser eye surgery. Initial recruitment activities through traditional methods found filling up recruitment targets challenging. Researchers could find underutilized high-volume clinical sites and previously unidentified patient populations by including AI and selected real-world data. This change in recruitment strategy gave rise to a successful enrollment period and underscored the usefulness of AI in adaptive trial performance and candidate identification.

4.4. Evaluation metrics and benchmarks

Many parameters are tracked to measure the efficiency of AI-based recruitment systems. The most important is reduced screening time, where systems like ACTES report a 34% improvement versus manual processes. Usability scores are also important indicators of staff satisfaction and system adoption; for instance, ACTES had an 80% usability rating. The speed of recruitment is another important measure; AI-based approaches typically achieve recruitment rates three times quicker than the traditional approaches. Enrollment and trial completion rates are also ultimately the concluding measures of the success of recruitment systems since these directly affect the feasibility and statistical power of clinical trials. Combined, these measures provide an overall view of how AI is reshaping the operational terrain of clinical trial recruitment.

5. Results and Discussion

5.1. Performance of AI model(s)

High accuracy and efficiency of the AI models used in the recruitment engine in detecting eligible clinical trial participants were demonstrated. The system reported precision scores exceeding 85% when matching patients against trial criteria from structured and unstructured EHR data by employing a mixture of machine learning classifiers and NLP-driven eligibility parsers. Moreover, feedback loops and human-in-the-loop control facilitated ongoing improvements in the algorithms, heightening recall rates in long-term tests. The AI demonstrated great adaptability during the processing of various source data that included clinical notes and feature genomics to demographic databases such that the eligibility matching was contextually relevant and all inclusive (Table 1).

Table 1: Comparative Metrics AI-Driven vs. Traditional Patient Recruitment.

5.2. Comparison with traditional recruitment

The AI-driven system is compared to the traditional recruitment procedures; the system provided substantial performance blessings on several metrics. Manual recruitment normally entails physical review by physicians or coordinators of patient charts, which are costly in labor and slow throughput. Instead, AI recruitment works in real time and scans thousands of records with a negligible delay. Furthermore, with the help of AI, AI-enabled systems enormously reduce the bias and fatigue of the human, which are prevalent in traditional approaches (Table 2).

Table 2: Comparison of Traditional vs AI-Driven Recruitment.

5.3. Ethical, Regulatory and privacy doubts discussion

Despite the many benefits of AI, its flavor brings some ethical and regulatory issues to the table. Algorithmic bias is undoubtedly a severe problem and this can come about due to the lack of a representative training dataset. In the event that historical EHR data have disparate reflections of populations, the AI may again unintentionally prefer or exclude a particular demographic group, thereby reducing equity in trial access. Transparency is another issue, as many AI models operate as black boxes, making it hard for clinicians to understand or trust the automated determination. Approaching this problem, the system incorporates XAI layers that provide rationale for each match and enhance interpretability and accountability. US guidelines like HIPAA and EU’s guidelines of GDPR must be adhered to even in handling sensitive health data, by AI recruitment tools. These regulations require strong measures around data access, usage and consent. Access control, logging and continuous checks for monitoring compliance provide a Security & Governance Layer in the proposed architecture. However, IRB ethical surveillance is still necessary in practice, even with algorithmic adaptation of eligibility criteria.

5.4. Limitations and trade-offs

There are limitations to the system. One of the pertinent tradeoffs is that generalizability models trained on data of a particular hospital or a region may not generalize equally well across different world healthcare systems because of differences in documentation style, coding practices or patient demographics. Data incompleteness is also a long-term problem; a lot of EHRs have missing or outdated information that can influence eligibility assessment. There is also an element of human validation. Even though the system can automate much of the screening process, in the end, final decisions on a high-stakes trial are almost always made based on a clinician’s independent review. However, This human-in-the-loop approach is advantageous regarding safety but partly restricts full automation's scalability and speed benefits. Cost and training requirements for deploying and maintaining such systems in smaller clinics or developing regions can be prohibitive until a cost-effective approach has been found, which will be a barrier to adopting such global systems.

6. Future Directions

6.1. Enhancing explainability in AI recruitment

AI systems are becoming more important in clinical trial recruitment and improvement of explainability becomes significant for establishing trust among clinicians, regulators and participates. Despite the use of explainable AI (XAI) modules in current systems such as TrialGPT, these XAI modules are still developing. Future improvement should value intuitive, human-readable explanations for why a patient qualifies (or does not) for a trial. This might include some visual dashboards, natural language justifications or even transparency scores for each recommendation presented. Improved explainability will also allow for greater monitoring of Institutional Review Boards (IRBs), thereby validating ethical recruitment procedures and that AI powered decisions are harmonious with clinical reasoning.

6.2. Integration with decentralized trials

As Decentralized Clinical Trials (DCTs) that use remote monitoring, wearable devices and telemedicine increase, the next wave of AI recruitment systems needs evolution to support the virtual workflow. Unlike traditional trials that rely on hospital-based screenings, DCTs call for AI models to interpret various data inputs, including mobile health records, home diagnoses and patient-reported outcomes. Future systems will have to be able to connect easily with telehealth platforms and remote consent modules to allow Continuous eligibility monitoring and real-time patient engagement. This change increases geographic inclusion and inclusivity for populations with little access to traditional healthcare facilities.

6.3 Use of federated learning and synthetic data

To mitigate the privacy and data representation issues, new recruitment models will be informed by federated learning and synthetic data generation. Algorithms of federated learning make it possible to train AI models at healthcare institutions without sharing raw data while protecting patients' confidential data and increasing generalizability across various populations. At the same time, fake data is artificially created. Still, statistically accurate patient records can augment training data sets to address areas inadequately represented in the demographics or underrepresented cohorts of rare diseases. These technologies will help to build stronger and less biased recruitment algorithms while conforming to the world’s privacy regulations, such as HIPAA and GDPR.

6.4. Policy and standardization efforts

The regulatory framework and standardized protocols should be rolled out for AI to be well deployed in clinical recruitment. There is a shortage of approved verification procedures, transparency criteria and auditability standards for AI-powered recruitment tools. Future policy efforts must state specific benchmarks to measure algorithm performance, fairness and interpretability. Regulatory bodies like the FDA EMA may put out new guidance for ICT-based recruitment, while international consortia can introduce interoperability standards so that tools work the same way across EHR platforms and regions of the world. Such initiatives will enhance adoption and ensure ethical, equitable and clinically valid AI recruitment.

7. Conclusion

AI-driven patient recruitment constitutes an unprecedented revolution in clinical trial activities. It presents scalable, efficient and increasingly intelligent solutions to a once manual process that has been dismally in error. Using machine learning, natural language processing and real-time data from the EHRs, these systems reduce recruitment time dramatically, improve match accuracy and offload operational burden to the research teams. TrialGPT and ACTES, among other tools, have already proved inducible benefits in actual world settings, such as increased enrollment rates, faster eligibility screens and increased user satisfaction. However, butting into play all the potential of AI in recruitment means continuous investment in explainability, fairness and data governance. In order for AI recommendations to be transparent, include all stakeholders and be compliant with patient privacy law, ethical and regulatory protection should keep growing in line with the technological capabilities. Other upcoming inventions (differentiated with decentralized trial integration, federated learning and global standardization) will also increase the outreach and dependability of AI recruitment systems. The combination of AI and clinical research is likely to speed up the drug development process and make clinical trials more affordable and acceptable to the diverse patient populations of the world.

8. References

1.   Nashwan AJ, Hani SB. Transforming cancer clinical trials: the integral role of artificial intelligence in electronic health records for efficient patient recruitment. Contemporary Clinical Trials Communications, 2023;36: 101223.
2. Ismail A, Al-Zoubi T, El Naqa I, et al. The role of artificial intelligence in hastening time to recruitment in clinical trials. BJR| Open, 2023;5: 20220023.
3. Thoma A, Farrokhyar F, McKnight L, et al. How to optimize patient recruitment. Canadian Journal of Surgery, 2010;53: 205.
4. Sood A, Prasad K, Chhatwani L, et al. Patients' attitudes and preferences about participation and recruitment strategies in clinical trials. In Mayo Clinic Proceedings, 2009;84: 243-247.
5. https://www.avenga.com/magazine/how-ai-advances-patient-recruitment-in-clinical-trials/
6. Visanji EC, Oldham JA. Patient recruitment in clinical trials: A review of the literature. Physical therapy reviews, 2001;6: 141-150.
7. Brøgger-Mikkelsen M, Ali Z, Zibert JR, et al. Online patient recruitment in clinical trials: systematic review and meta-analysis. Journal of medical Internet research, 2020;22: 22179.
8. Krieger JL, Neil JM. Communication and recruitment to clinical research studies. In Oxford Research Encyclopedia of Communication, 2016.
9. Suryanarayanan P, Epstein EA, Malvankar A, et al. Timely and efficient AI insights on EHR: system design. In AMIA Annual Symposium Proceedings, 2021;2020: 1180.
10. Lu X, Chen M, Lu Z, et al. Artificial intelligence tools for optimising recruitment and retention in clinical trials: a scoping review protocol. BMJ open, 2024;14: 080032.
11. Yina W. Application of EHR in health care. In 2010 Second, International Conference on Multimedia and Information Technology, 2010;1: 60-63.
12. Anuyah S, Singh MK, Nyavor H. Advancing clinical trial outcomes using deep learning and predictive modelling: Bridging precision medicine and patient-centered care, 2024.
13. Hutson M. How AI is being used to accelerate clinical trials. Nature, 2024;627: 2-5.
14. Goetz J, Tewari A. Federated learning via synthetic data, 2020.
15. https://www.clinicalleader.com/doc/areas-where-ai-could-revolutionize-patient-recruitment-and-retention-0001
16. Snell L, Tallett S, Haist S, et al. A review of the evaluation of clinical teaching: new perspectives and challenges. Medical education, 2000;34: 862-870.
17. Baker T, Gerdin M. The clinical usefulness of prognostic prediction models in critical illness. European journal of internal medicine, 2017;45: 37-40.
18. McKinley RK, Fraser RC, Baker R. Model for directly assessing and improving clinical competence and performance in revalidation of clinicians. Bmj, 2001;322: 712-715.