Ä¢¹½tv

What challenges are pharmaceutical, medical device, and in vitro diagnostic companies facing in realizing the benefits of AI/ML applications?

Artificial intelligence (AI) and machine learning (ML) have proven to be game changers for the pharmaceutical, in vitro diagnostic (IVD), and medical device industries, where AI/ML-powered algorithms can rapidly analyze huge amounts of data, identify patterns, uncover correlations, and generate actionable insights. 

As AI/ML applications in the pharmaceutical space have grown, the number of submissions to the Food and Drug Administration that include AI/ML components has significantly increased. FDA received  in 2021, an approximately 10-fold increase over 2020. As of February 2025, FDA had authorized more than , with more than 160 authorized since January 2024.

While AI/ML systems have found widespread use in unregulated domains such as consumer electronics and marketing, adoption in life sciences has been comparatively slower because of concerns about performance in high-risk systems and deployment without a full understanding of how the system arrived at its conclusions. 

Besides explainability, other hurdles with AI/ML systems include the challenge of training AI/ML algorithms when high-quality real-world data (RWD) is not available, as models trained on lab data or limited RWD may not generalize well across the population. Despite these issues, the applications of AI/ML in pharmaceutical manufacturing, in vitro diagnostics, and medical devices are evolving as practitioners identify new approaches to addressing challenges.

AI/ML can help analyze, synthesize, and aggregate RWD into RWE to improve regulatory submissions, expand product indications, and accelerate pharmaceutical manufacturing, IVD, and medical device development. 

Impact of AI/ML on pharmaceutical manufacturing

Continuous manufacturing

Advanced pharmaceutical manufacturing using continuous processing can help larger pharmaceutical companies lower costs and increase efficiencies by replacing batch manufacturing and batch testing with continuous manufacturing and continuous testing supported by AI/ML-based systems. Using sensors and AI/ML to collect and analyze data, continuous processing employs real-time monitoring and adjustments to assess product quality, decrease waste, and optimize asset utilization while providing:

• Real-time assurance: AI/ML can provide immediate insights by predicting optimal conditions for drug synthesis; optimizing manufacturing parameters; and identifying patterns and signatures of effective processes, which can help in replicating successful manufacturing outcomes.
• High-quality standards: AI/ML tools can identify potential issues and make necessary adjustments in real time, helping final products meet manufacturing specification requirements while identifying drug impurities to improve quality and safety. 

Predictive capabilities

By analyzing large datasets, AI/ML models can provide insights into manufacturing processes by identifying trends and predicting process issues, such as equipment failures or deviations in the production process, which can help in:

• Optimizing process design and process control to make production both efficient and scalable
• Increasing equipment reliability and throughput
• Monitoring early warnings or signals that the manufacturing process is not in a state of control
• Detecting recurring problem clusters
  • Regulatory authorization: Regulatory agencies are risk-averse and require thorough validation of AI/ML models to ensure they are making accurate and reliable decisions.
  • Validation: Performing traditional software validation and gaining regulatory confidence in AI/ML's predictive capabilities is a significant hurdle for widespread adoption in the pharmaceutical industry. This is because there are  and, unlike device development or other applications, the lead times for pharmaceutical process development are typically much longer, so compelling use cases have to be designed from the beginning of pharmaceutical development to make it economically viable.

AI/ML can also improve operational efficiency by predicting supply chain disruptions and helping manufacturers plan accordingly. This includes forecasting demand, managing inventory, mitigating the impact of potential disruptions, and predicting raw material shortages to ensure that raw materials are available when needed, reducing the risk of production delays. AI/ML can be used in this predictive capacity for supply chain bottlenecks relatively quickly, as regulatory approval is not required to improve operations. 

AI/ML challenges in pharma manufacturing 

Although pharma manufacturing has been transitioning from batch processing to continuous processing over the past 20 years,  remain despite its benefits. For example, FDA regulatory review and validation through data submission are required for any major change to an existing approved pharmaceutical manufacturing process.

Overcoming these hurdles requires defining appropriate test cases, understanding the nuances of different risk profiles, and conducting risk evaluations. It also involves engaging early with FDA to understand the requirements for AI implementation, discuss AI models, and seek guidance on validation and approval processes.

AI/ML-enabled in vitro diagnostics and other medical devices 

The IVD space has been flooded by integrated biosensors — fueled by recent innovations in cloud computing and embedded electronics — that gather data for analysis by AI/ML models to improve testing processes and in vitro diagnostic wearables. AI/ML has shown the potential to improve "diagnostic stewardship" by streamlining and , increasing efficiency, and reducing costs while avoiding duplicative tests, reducing errors, improving throughput, and enhancing the use of laboratory resources.

Advantages of AI/ML in IVD wearable devices

In vitro diagnostic wearables, such as continuous glucose monitors and health-related monitoring devices, can be worn or partially inserted into the body to evaluate biomarkers or provide information for diagnostic purposes. These devices analyze biomarkers using software components such as AI/ML models, which analyze data to predict the likelihood of events and provide better clinical decision-making tools.

Benefits of AI/ML in other medical devices 

Similar to AI/ML-enabled IVDs, software as a medical device (SaMD) and software in a medical device (SiMD) based on AI/ML may be trained on new, continuously collected data enabling routine software updates in minimal time. AI/ML-enabled medical devices trained on personalized data are helping create a more customized, data-driven healthcare approach that can improve patient outcomes and lower treatment costs.

For example, remote patient monitoring that leverages adaptive AI/ML technology can reduce in-person visits and identify opportunities for early treatment interventions. In medical imaging, AI/ML is automating time-consuming, high-volume repetitive tasks, and, in the future, customized therapies may use AI/ML medical devices to deliver treatments based on individual needs. 

AI/ML hurdles 

As IVD wearables and other medical devices consider novel biomarkers and use cases, stakeholders face the challenges of using AI/ML, which include: 

• Developing unbiased algorithms that are interpretable and explainable, allowing healthcare professionals to understand the rationale behind AI-driven decisions.
• Validating learning algorithms to meet FDA requirements and ensure clinical utility; applying AI/ML algorithms to healthcare data; and demonstrating that AI/ML models perform accurately and reliably across different patient populations and in real-world settings, which remains difficult for exploratory algorithms and novel use cases.
• Keeping data secure, private, and in compliance with the Health Insurance Portability and Accountability Act of 1996 (HIPAA).

Pre-determined change control plans

FDA introduced "pre-determined change control plans" (PCCPS) in a draft version of the now finalized December 2024 guidance document ." PCCPs describe how foreseeable modifications can be made to a medical device without requiring a new marketing submission. By outlining protocols on how modifications will be implemented and validated, a PCCP is intended to provide reasonable assurance of a product's safety and effectiveness. This enables manufacturers using AI/ML in medical devices to update trained models without requiring a new marketing submission, thus allowing them to bring the next generation of their product to market faster. 

While the ability to learn from data and integrate iterative improvements can lead to the development of more advanced technologies and precise diagnoses, it also carries risks, such as introducing bugs or reducing model performance on outliers. PCCPs intend to mitigate some of these challenges by defining what modifications will be implemented, how they will be implemented, and what impact assessments will be executed before they are deployed. The table below lists four examples of authorized AI/ML-enabled medical devices spanning a variety of applications that also include an authorized PCCP.

FDA's latest draft guidance, dated August 2024, "" also discusses software changes to improve a device that are consistent with its intended use. This change allows vendors to consider PCCPs for cybersecurity, a critical component in AI/ML-based medical devices that also connect to a network. Manufacturers may provide software updates to address cybersecurity issues without requiring new FDA authorizations, provided there is a validation and verification process in place. This flexibility is crucial for maintaining medical device safety and effectiveness over time. 

Four examples of AI/ML-enabled medical devices with an authorized PCCP from FDA 

RWD/RWE in pharmaceutical manufacturing, IVD, and medical devices

AI/ML can help analyze, synthesize, and aggregate RWD into RWE to improve regulatory submissions, expand product indications, and accelerate pharmaceutical manufacturing and IVD and medical device development.

AI/ML algorithms can clean and analyze "messy" RWD by removing noise and extracting useful information from large datasets, making the data more valuable for generating RWE. In this way, AI/ML can:

• Improve regulatory submissions: AI/ML can provide robust RWE to support the effectiveness of drugs and medical devices and strengthen the case for regulatory authorization.
• Expand product indications: Stakeholders can use RWE to identify new uses and benefits for their drugs or devices and to grow the indications for their products, leading to expanded labeling and market opportunities.
  • Variable quality: RWD typically varies in quality and consistency because it includes data collected from a variety of real-world sources: a patient's electronic health record in a hospital, decentralized trials, product and disease registries, self-generated patient data (e.g., home-use settings or mobile devices or digital health tools, wearables, etc.), a manufacturing site, etc.
  • Interpretation, repeatability, and reliability: RWD often lacks the context provided in controlled clinical settings, making it difficult to analyze; it can also vary significantly due to natural patient behavior, making it difficult to replicate results consistently.
  • Data security: Managing large volumes of data and maintaining compliance with data protection regulations in real-world settings is more difficult than in more controlled clinical settings.

Challenges in generating RWD

AI/ML algorithms are only as good as the data they are trained on, meaning producing high-quality RWD is essential. However, collecting and analyzing real-world data compared to controlled clinical settings creates its own obstacles: 

Stakeholders can surmount these challenges by working with experts who can:

• Run clinical validation studies, design verification studies, and create protocols to test and validate AI/ML algorithms
• Define the appropriate test cases and conduct risk evaluations
• Develop fit-for-purpose tools and protocols to tailor the tools and processes to specific populations and to specific diseases to garner the most meaningful data
• Apply an in-depth knowledge of data architecture, infrastructure, and management to processes
• Implement good clinical practice requirements of health data to keep data secure and compliant