#8 Use of AI as a Knowledge Management Tool in Medical Products Development

The medical products lifecycle is a very complex, lengthy and expensive process. During the discovery, development and marketing of medical products tremendous amounts of information are generated. In the past, the information in company repositories were accessed through key words searches. Now with the availability of Artificial Intelligence (AI)/ Machine Learning (ML) platforms this information can be mined, utilized and projected for new products under development. Recognizing this transition from traditional approaches to AI, FDA is requesting feedback on its discussion paper titled “Using Artificial Intelligence & Machine Learning in the Development of Drug & Biological Products” and EMA has released a reflection paper titled “Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle”. [Note: The term AI in this write-up encompasses all aspects of AI and ML]

FDA’s and EMA’s papers provide a good overview of the use of AI/ML and current thinking on how AI can be used within the regulated industry in all areas of medical product life cycle (discovery, non-clinical research, clinical research, post-market safety surveillance and advance pharmaceutical manufacturing). In addition, it is notable that it has also begun to explore the use of AI and has launched an agency-wide AI tool named Elsa, which stands for "Electronic Language System Assistant". Elsa is a large language model-powered tool intended to assist FDA staff with a variety of tasks such as summarizing information, drafting code, and reviewing documents to improve efficiency. From health authorities view point the regulatory agency is open to the use of AI in all aspects of the medical products lifecycle provided the assessment are conducted within a specific context of use that is defined and risk-based approaches are integrated in such evaluations and management of knowledge/information.

Accelerating Discovery, Research, and Manufacturing

The highest potential for the use of AI is in drug discovery and identification of potential molecules for development. There is a wealth of information available in the literature on small molecule chemical structure-activity relationships and biological targets that can alleviate symptoms or even cure a disease. The area where AI would be very helpful is in cell-gene therapy where molecules (genomic, proteomic, transcriptomic, etc.) have complex structures as well as new research information being generated every day. The vision is to use AI/ML to mine and analyze these large datasets for potential structure and function of biological targets thereby predicting their role in a disease pathway (Vamathevan et al., 2019; Weissler et al., 2021). The hypothesis generated using AI/ML can then be verified by laboratory work to validate science. In theory, the traditional iterative process that would have taken decades to identify a target molecule for a novel biological target can now be demonstrated in months. In addition, with the use of AI, currently available molecules can be repurposed for new therapeutic uses or design new molecules that can be used therapeutically. Insilico Medicine used AI in the generation of fibrosis drug (INS018_055) which successfully entered Phase I clinical trials. Insilico used a generative AI platform called Pharma.AI to identify a novel biological target for idiopathic pulmonary fibrosis (IPF) and designed a new small-molecule inhibitor using generative chemistry. Use of AI reduced the timeline to 18 months from target discovery to preclinical studies by helping in compound selection and optimization using target prediction models. This is the first AI-designed drug and AI-identified target to reach human trials.

Once the molecule and target are identified, the molecule must be tested extensively in animals to understand the safety profile and the pharmacokinetic/pharmacodynamic (PK/PD) profile. This area of pre-clinical or non-clinical research is resource intensive and utilizes a large number of animals to demonstrate science. Globally, this area of research on animals is controversial and there is a push to reduce or eliminate the use of animals. With the availability of AI and Insilico models there is great expectation for the potential to reduce the utilization of animals. FDA issued a press release in April 2025 on a plan to phase out animal testing requirement for monoclonal antibodies and other drugs leading the way for more future reductions in animal use for testing. Novartis collaborated with MIT and IBM Watson to use ML models to predict liver toxicity, cardiotoxicity and metabolic instability which led to reduced animal studies and improved preclinical decision-making.

In addition, to understanding the safety of the molecule, there is a lot of work done to scale-up the production of the molecule from few grams to kilos. AI can help in the manufacturing development process to scale-up production molecules by predicting critical parameters in the manufacturing process and how to control them. The wealth of available knowledge/ information in the manufacturing process development area together with AI can also support the dosage form (e.g., tablet, capsule) which will speed molecules from discovery to transition into clinical development. Finally, with the help of AI/ML strategies, selection of a dose and dosing regimen to start a clinical development program can now be streamlined and risk largely mitigated compared to pre-AI days. One case study is the use of AI-supported mRNA sequencing to optimize mRNA sequence design, Immunogenicity prediction and dose selection in early clinical development which helped accelerate the COVID-19 mRNA-1273 vaccine development timeline.

Transforming Development, Real-World Evidence, and Pharmacovigilance

Clinical research and development of medical products is the most expensive process in the life cycle of a medical product. Only 1 in 10 drugs that enter clinical trials are approved. With the help of AI, it is anticipated the success rate from development to approval of new medical products to increase. All aspects of clinical development are anticipated to be supported by AI. Designing clinical trials with appropriate trial outcomes to demonstrate the intended use of medical products is a very iterative process pre-AI where clinical trials may have to be repeated as the appropriate primary variable may have not been used or the appropriate statistical methodology identified a priori. It is anticipated that AI/ML tools will help in designing trials using predictive modeling and counterfactual simulation that are more efficient as AI will be capable of analyzing and integrating all available information from clinical trial databases and real-world data (RWD) from electronic health records (EHR) to inform the appropriate clinical trial design.

Using available information from clinical trial databases, the AI tools can also enable drug sponsors to identify appropriate study sites and investigators to run the clinical trials. All clinical trials involve recruitment of patients which can be challenging and rate-limiting for the conduct of trials. If AI is allowed access to medical records and hospital records, to identify a list of potential patients with the disease or symptom quickly and if needed the patients pool can be further refined using the inclusion and exclusion criteria for the specific clinical trial. Then AI could also play role in connecting these patients to potential investigational trial sites, thereby maintaining privacy and minimizing bias. Once the patients are enrolled into the study AI assisted applications can help patients adhere to the dose/dosing regime, capture any side effects and keep them motivated to remain in the trial for completion. GSK and Viant AI used AI to analyze EHR’s and historical clinical data to predict ideal clinical trial sites, improve patient recruitment and Identify dropout risk factors that led to 15–20% reduction in enrollment timelines. Janssen used ML models to predict high-performing sites with reduced risk of protocol deviations and more efficient recruitment patterns which led to improved trial operational efficiency and resource allocation.

During the clinical trial process, there is no evaluation to see if the medical product is working as intended, except for monitoring for safety signals or if there is a pre-planned interim analysis. Looking at data in real time and in an unbiased manner to see if the medical product is working as per the intended use or if there are any safety signals with the use of AI. Drug sponsors can monitor the patients in the trial and watch for alerts or safety signals. These sophisticated tools can be used by institutional review boards and data monitoring boards to monitor study progress across sites and make any decisions to stop the trials or put trials on hold. If the product is working as intended from the clinical endpoint assessment and the regulatory authority can review data as it is generated, maybe trials could be stopped earlier than intended and the medical product could be submitted for approval.

Digital health devices such as wireless and smartphone-connected products, wearables, implantables, and ingestibles, are being widely used to collect objective, quantifiable, longitudinal, and continuous physiological data in clinical trials. These digital health devices have transformed and revolutionized clinical trial data collection and management. The embedded algorithm within these digital health devices have been used to predict the status of a chronic disease and its response to treatment (Stehlik et al., 2020). AI can help analyze the large and diverse data generated from the continuous monitoring of patients using these technologies and aid in the evaluation of multimodal data and composite measures (Cohoon & Bhavnani, 2020).

AI can be used for data cleaning and curation purposes such as imputation of missing data values (Zhang, Yan, Gao, Malin, & Chen, 2020) to prepare the data for analysis. FDA paper purports that AI can be used to enhance data integration efforts by using supervised and unsupervised learning by integrating data submitted in various formats and perform data quality assessments across studies. This data curation via masking and de-identification of personal identifiable information, metadata creation, and search and retrieval of stored data will be valuable for regulatory authorities in the review process and drug companies in the development process. Overall, it is anticipated that AI applications can potentially increase data accuracy and improve the speed at which data are prepared for analyses. (FDA discussion paper).

The biggest expectation for the use of AI will be its ability to analyze high volumes of diverse and complex RWD extracted from EHRs, medical claims, and disease registries. Clinical endpoint assessment is a key part of evaluating safety and efficacy of medical interventions in clinical trials. The success of the trial relies on the clinical endpoint assessment to demonstrate if a medical product performed as intended. The capability of AI in data analysis as the trial is ongoing (i.e. real-time) may be of immense value to patients, investigators, companies and regulatory authorities in the future. This will allow medical products to reach market faster.

Drug and biological product applications submitted to FDA that include AI/ML have increased over the last few years to more than 100 submissions in 2021 (Q. Liu et al., 2022). FDA is using AI for pharmacovigilance (FAERS data mining) to detect adverse event safety signals, prioritize cases for human review and Identify drug-drug interaction risks. This has improved both signal detection speed and sensitivity. Many pharmaceutical companies are also using AI-assisted systems to process adverse event reports and for pharmacovigilance safety signal detection. AI is here to stay and expansion in the scope of its involvement in all aspects of medical products development leading to shortening the time and cost of medical product development is inevitable. These AI/ML tools are great assistants (or an intern that gathers all the pertinent information) to humans to speed up the medical products development process but ultimately humans will have to review the information and be the ultimate decision maker!

References

FDA Discussion Paper

EMA Reflection Paper

Cohoon, T. J., & Bhavnani, S. P. (2020). Toward precision health: applying artificial intelligence to digital health biometric datasets. Per Med, 17(4), 307-316. doi:10.2217/pme-2019-0113

Liu, Q., Huang, R., Hsieh, J., Zhu, H., Tiwari, M., Liu, G., . . . Huang, S. M. (2022). Landscape analysis of the application of artificial intelligence and machine learning in regulatory submissions for drug development from 2016 to 2021. Clin Pharmacol Ther.doi:10.1002/cpt.2668

Stehlik, J., Schmalfuss, C., Bozkurt, B., Nativi-Nicolau, J., Wohlfahrt, P., Wegerich, S., Pham, M. (2020). Continuous Wearable Monitoring Analytics Predict Heart Failure Hospitalization: The LINK-HF Multicenter Study. Circ Heart Fail, 13(3), e006513. doi:10.1161/circheartfailure.119.006513

Vamathevan, J., Clark, D., Czodrowski, P., Dunham, I., Ferran, E., Lee, G., . . . Zhao, S. (2019). Applications of machine learning in drug discovery and development. Nat Rev DrugDiscov, 18(6), 463-477. doi:10.1038/s41573-019-0024-5

Weissler, E. H., Naumann, T., Andersson, T., Ranganath, R., Elemento, O., Luo, Y., Ghassemi, M. (2021). The role of machine learning in clinical research: transforming the future of evidence generation. Trials, 22(1), 537. doi:10.1186/s13063-021-05489-x

Zhang, X., Yan, C., Gao, C., Malin, B. A., & Chen, Y. (2020). Predicting Missing Values in Medical Data via XGBoost Regression. J Healthc Inform Res, 4(4), 383-394. doi:10.1007/s41666-020-00077-1