Applied intelligence in clinical drug development: Potential benefits and emerging concerns

Abstract

The use of artificial intelligence (AI) technology and machine learning (ML) is growing exponentially and is moving from AI to applied intelligence. Pharma industry is actively exploring the potential use of AI tools in new product discovery and clinical development. Some of the practical applications of AI in clinical development are for improving the efficiency of enrollment, selection and stratification of participants, optimizing study treatment, enhancing compliance, data analysis, and pharmacovigilance. AI applications have been used for outcome prediction; covariate selection/confounding adjustment; anomaly detection; real-world data phenotyping; imaging, video, and voice analysis; endpoint assessment; and pharmacometric modeling in regulatory submissions. However, widespread applications of novel yet difficult-to-understand AI technology in clinical development would need balancing the benefits and risks and resolving issues of scientific validity, technical quality, and ethics. The article discusses the potential benefits and emerging concerns of applying AI in clinical drug development.

INTRODUCTION

The use of artificial intelligence (AI) technology and machine learning (ML) is growing exponentially and is rapidly progressing from promising potential to practical applications. For pharma industry, employing AI tools for identifying new drug targets and designing new molecules, improving the productivity of clinical trial conduct[1,2,3,4] and pharmacovigilance processes[4,5,6] is an attractive novel option. However, such widespread applications of novel yet difficult-to-understand AI technology are increasing concerns about scientific validity, technical quality, and ethical issues.[2,4] This brief review will discuss the potential benefits and emerging concerns of applying AI in clinical drug development.

POTENTIAL UTILITY OF ARTIFICIAL INTELLIGENCE

The two approaches – generative and predictive AI – have wide applications in planning and conduct of clinical trials.

Generative AI can create content, text, software code, audio, images, and video and translate data into different formats.[7] Generative AI models are trained on large datasets to create content that is similar, but not identical, to the original data and are useful for creative tasks or developing innovative solutions, e.g., drug discovery and medical writing.

Predictive AI can extract insights from historical data to make predictions, recommendations, and trends.[7] Predictive AI uses smaller, targeted datasets as input data and applies statistical algorithms to helpful decision-making and strategy formulation, e.g., healthcare diagnosis and prognosis.

We can consider the role of AI in different aspects of clinical trial conduct.

Selection of investigator sites

AI technology can support in identification of investigator sites, which have the greatest potential for a successful conduct and completion of clinical trial. AI algorithms, using data from clinical trials at a site, can help in evaluation of site performance and in determining high-risk sites which are likely to run behind schedule.[4,8]

Enrollment of participants

AI technology can mine vast amounts of data from clinical trial databases, trial announcements, social media, medical literature, clinical trial registries, and electronic medical records (EMRs) and facilitate matching of patients to clinical trials.[8] Recently, Jin et al. have developed TrialGPT, an end-to-end framework for matching patient to trial with large language models (LLMs).[9] TrialGPT model incorporates (1) TrialGPT-Retrieval, which does large-scale filtering to retrieve candidate trials, (2) TrialGPT-Matching, which predicts patient eligibility at level of selection criteria, and (3) TrialGPT-Ranking which generates trial-level scores.[9] TrialGPT was evaluated on three publically available cohorts of 183 synthetic patients with over 75,000 trial annotations. TrialGPT-Retrieval could recall over 90% of relevant trials using <6% of the initial collection. TrialGPT-Matching, during manual evaluations on 1015 patient criteria had an accuracy of 87.3% pairs, which was similar to the performance of medical expert. The TrialGPT-Ranking scores showed high correlation with human judgments and outperformed the best-competing models by 43.8% in ranking and excluding trials. TrialGPT could also decrease the screening time by 42.6% in patient enrollment. TrialGPT seems valuable in facilitating the process of matching eligible patients with clinical trials. There are also opportunities of using AI for improving informed consent process by tailoring delivery, language, cultural context, and understanding for the participants.[2]

Selection and stratification of trial participants

AI has been explored to predict the clinical outcome of participant based on baseline data, e.g., demographic information, imaging, laboratory data, genomic information, etc.[4] Such predictive strategies can improve the efficiency of clinical trial by identifying participants who may be at high risk or may be responders to treatment. Dercle et al. used radiomics signatures on computed tomography scans of lung cancer patients to predict tumor sensitivity to nivolumab, docetaxel, and gefitinib.[10] Radiomics features were based on quantitative analysis of early tumor changes from baseline to first on-treatment assessment. The use of such AI algorithms during screening process could reduce variability and increase study power. AI models could also help in the stratification of participants by predicting the probability of a serious adverse event (AE) before an investigational product is dispensed.[4]

Study treatment

AI can be used in the prediction of pharmacokinetic profiles after drug administration and to investigate the relationship between drug exposure and therapeutic response.[4] Rakaee et al. conducted a cohort study in 958 patients with non-small cell lung cancer treated with immune checkpoint inhibitors and demonstrated that deep learning prediction scores were associated with response rate, progression-free survival, and overall survival.[11] These types of models can be used to optimize the dose regimen selection for a clinical trial, especially in special populations, e.g., rare disease, pediatric patients, and pregnant populations.

Compliance

AI can be used to monitor and improve compliance during clinical trial through tools such as smartphone alerts, eTracking of drugs by smart pillboxes, applications using digital biomarkers, e.g., facial expressivity, to monitor adherence remotely, and eTracking of missed clinical visits.[4]

Retention of participants

AI, enabled by Chatbots, can help in reducing the burden for participants by providing access to important clinical trial information from data generated during clinical practice or by study activities.[4] Participant’s data from digital health technology (DHT) can be used to predict dropouts and AEs, which can improve participant retention.

Clinical trial data collection, management, and report

Participant’s physiological data collected using DHTs, such as wireless and smartphone-connected products and wearables, enable the use of AI algorithms to predict the status of a chronic disorder and response to therapy or to identify novel characteristics of an underlying condition. AI can be used in analysis of large and diverse DHT data generated from the continuous monitoring.[4] Stehlik et al. monitored 100 patients using a disposable multisensor patch and showed that ML analytics on such data can accurately predict hospitalization for heart failure exacerbation.[12] AI, by detection of duplicate participants and imputation of missing data values, can potentially enhance data accuracy and improve the speed of organizing data for statistical analyses.[4] AI can analyze high volumes of diverse and complex data extracted from EMRs, medical claims, and disease registries.[4] AI-based simulation of participants and creation of digital twins with a variety of demographic traits could predict clinical trial outcomes before conducting human trials.[4] AI-based automated reporting can help in integrating information from all sources and in disseminating clinical trial results to researchers and participants.[2]

Pharmacovigilance

AI automation can help individual case safety report (ICSR) processing from spontaneous reports, clinical trials, EMRs, social media, phone calls, emails, literature, patient registries, claims data, and postapproval safety studies.[4] AI has potential usefulness for case validity, case prioritization, duplicate check, coding, and quality control and can automate reporting rules for regulatory submission of ICSRs.[4,5] AI has been applied for causality assessment, for classification of AEs by expectedness, deciding seriousness, and identifying a potential safety signal from clusters of signs and symptoms in real time.[4,13] Routray et al. used a stratified random sample of 22,932 AE cases from a data set of spontaneous or postmarketing, solicited, and medical literature and developed a deep learning model for evaluation of AE and case seriousness classification.[13]

ARTIFICIAL INTELLIGENCE IN REGULATORY SUBMISSIONS

Widespread expansion of applications of AI has become an important approach in regulatory submissions. Liu et al. from the Center for Drug Evaluation and Research, US Food and Drug Administration (FDA), have analyzed regulatory submissions of drugs and biological products from 2016 to 2021 which included AI/ML.[14] In 2021, the number of submissions increased approximately 10-fold to 132 as compared to 14 in 2020. AI/ML was part of submission for a variety of projects [Tables 1 and 2] as drug discovery/repurposing, improving clinical trial design elements, dose optimization, adherence to drug regimen, endpoint/biomarker assessment, and postmarketing surveillance.[14]

Table 1.

Artificial intelligence/machine learning analysis in the Food and Drug Administration submissions

Theme	AI/ML use
Outcome prediction	Prediction of clinical outcome, disease prognosis, treatment response
Covariate selection/confounding adjustment	Decision tree-based algorithms to screen through baseline information
Anomaly detection	Identifying potential outliers or anomalies
Real-world data phenotyping	Natural language processing to support phenotyping
Imaging, video, and voice analysis	Deep learning, for the analyses of imaging data, videos, or voices
Pharmacometric modeling	Algorithms for pharmacokinetic–pharmacodynamic modeling

AI=Artificial intelligence, ML=machine learning

Table 2.

Artificial intelligence/machine learning objectives in the Food and Drug Administration submissions

Theme	AI/ML use
Enrichment design	Patient characteristics to select a population in which detection of a drug effect more likely
Patient risk stratification and management	Prediction for a specific severe AE based on baseline information, and need of patient monitoring
Dose optimization	Based on patient characteristics
Adherence to regimen	Monitoring platforms to aid and confirm adherence
Synthetic control	Digital twins to predict placebo response
Endpoint/biomarker assessment	Digital health/audiovisual/radiographic biomarkers for outcome assessment
PMS	RWD for PMS requirement pregnancy outcomes study
Drug discovery	Selection of the therapeutic targets or drug candidates
Drug toxicity prediction	Based on drug structure/physiochemical properties/affinity for targets

AE=Adverse event, RWD=Real-world data, PMS=Postmarketing surveillance, AI=Artificial intelligence, ML=Machine learning

EMERGING CONCERNS

As the scope of utilizing AI is moving from artificial to applied intelligence, concerns [Table 3] are emerging about diverse scientific, technical, ethical, legal, and social issues and challenges of exploiting AI for product development.[1,2,4]

Table 3.

Challenges and issues in employing artificial intelligence

Themes	Issues
Data characteristics	Accuracy, bias, completeness, consistency, integrity, provenance, quality, reliability, representativeness, relevance, replicability, reproducibility
Synthetic data	Sensitive private information in training data
Model development	Bias, consistency, explainability, generalizability, monitoring, quality, relevance, reliability, transparency, validation
Bias	Human, systemic, statistical/computational
Ethics	Accountability, privacy, safety, security, transparency
Governance	Human-led governance, accountability, transparency

To address these concerns, academic institutions, government organizations, and health authorities have published discussion papers and guidelines.[4,15,16,17,18,19] Some of these are:

• US FDA Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products 2025

• Using artificial intelligence and machine learning in the development of drug and biological products discussion paper and request for feedback 2023

• US FDA Good machine learning practice for medical device development: guiding principles 2021

• European Medicines Agency guiding principles on the use of large language models in regulatory science and for medicines regulatory activities 2024

• European Commission living guidelines on the responsible use of generative AI in research

• SPIRIT-AI and CONSORT-AI Working Group. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI Extension.

INDIAN MILIEU FOR ARTIFICIAL INTELLIGENCE IN RESEARCH

The potential utility of AI in health care is the theme in review articles in Indian journals and news media. A recent scoping review of randomized controlled trials (RCTs) on AI in clinical practice showed 86 unique RCTs focused on deep learning systems for medical imaging.[20] Over 60% of RCTs were reported from the USA and China. However, there were no RCTs from India.[20] In the US registry, ClinicalTrials.gov., 30 AI-based Indian studies in healthcare diagnostics were registered. Research on AI modeling for health care in India seems to be in early stages.

Indian Council of Medical Research, anticipating growing interest in application of AI in biomedical research and health care, has issued ethical guidelines.[21] These guidelines cover AI applications for (1) diagnostics and screening, (2) therapeutics, drug discovery, and development, (3) clinical care, (4) epidemiology and prevention of disease, (5) behavioral and mental health care, and (6) management systems for health, clinic, and hospital. These guidelines describe ethical principles, guiding principles for development, validation, deployment of AI, ethical review of protocols, informed consent process, and governance of AI technology use. The ethical principles lay emphasis on autonomy, safety and risk minimization, trustworthiness, data privacy, accountability and liability, optimization of data quality, accessibility, equity and inclusiveness, collaboration, nondiscrimination and fairness principles, and validity.

CONCLUSIONS

Integration of AI models in drug development has enormous potential to improve the efficiency of clinical trial conduct processes from planning and initiation to performing, recording, analysis, and reporting. Despite the potential utility of AI technology, critical technical and ethical concerns remain for application of AI in clinical development. Hopefully, current initiatives of regulatory authorities in providing systematic framework for developing high-quality validated AI models which are ethical, effective, and safe would give a tremendous boost to incorporation of AI in drug development.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

Nil.