SMART Tools for FDA Regulatory Excellence: Integrating SmartSearch+ with FDA AI/ML Copilot for Compliance

Navigating the regulatory landscape in the pharmaceutical and life sciences industries requires precision, speed, and continuous monitoring. However, compliance is challenging because of outdated metadata-based search technologies and the overwhelming volume of FDA regulations, guidance documents, and Warning Letters. Pharmaceutical companies have new avenues to enhance their compliance processes with the rise of generative FDA AI/ML Copilot tools that leverage large language models such as ChatGPT-4 and proprietary technologies like SmartSearch+.  

This paper introduces a "duelling banjo" approach: combining AI/ML’s generative insights with human-led, bias-free full-text verification through SmartSearch+ on actual published FDA contents on FDA.Gov. Organizations can tackle complex regulatory requirements more accurately and efficiently by integrating these tools within a single eSTARHelper Microsoft AI/ML ChatGPT-4 app interface.  

Above is the screenshot of the eSTARHelper website featuring the FDA CoPilot. Depending on the topic, specific custom copilots can be created, such as the DSCSA AI/ML Copilot or a Microsoft AI/ML FDA Copilot for a Warning Letter Issued by the FDA. Custom Copilots can be grounded with external websites & internal enterprise data. 

Regulatory Data Sources and Challenges 

The FDA provides key regulatory documents essential for compliance:  

• 21 CFR: Enforceable regulations granting the FDA authority to impose legal sanctions.
• Guidance PDFs: Recommendations outlining the FDA’s thinking on drug lifecycle processes. 

Despite their value, the FDA’s search capabilities remain outdated and remain metadata search by design, often failing to surface critical documents. This creates significant hurdles for regulatory professionals seeking precise information, a challenge likened to "drinking from a firehose."  

The Role of AI/ML in Regulatory Compliance 

Generative AI/ML models like GPT-4 offer probabilistic insights that can kickstart the drafting of regulatory responses. However, these models alone cannot ensure the accuracy, completeness, or precision required. For example, FDA Warning Letters often highlight deficiencies but leave it to manufacturers to investigate ripple effects, such as those under 21 CFR 211.192. Human-led tools like SmartSearch+ on actual FDA-published regulatory contents validate AI outputs and ensure comprehensive responses.  

Integration of SmartSearch+ and FDA Copilot 

Key Features of SmartSearch+:  

• Full-Text Search: Indexes over All FDA guidance PDFs and 21 CFR regulations for precise retrieval published by CDER (Drugs), CBER (Biologics), and CDRH (Medical Devices), entire contents viewable online without downloads 

• Tangential Analysis: Identifies related regulations and ensures a thorough compliance strategy that allows two-level filters narrowing search result hits, namely by (a) only regulatory content data type, whether PDF Guidance or only 21 CFRs, and (b) filtering only those issued by CDER or by CBER or CDRH.  

• Human-Led Validation: Provides an additional layer of reliability for AI-generated insights. 

Key Features of FDA Copilot: 

• Generative Drafts: Produces initial response drafts aligned with FDA guidelines. • Retrieval-Augmented Generation (RAG): Integrates proprietary knowledge, enhancing context and precision. 

• Enterprise Integration: Connects seamlessly with secure authentication, enabling authorized access to Microsoft Office 365, SharePoint Online, and Teams data. 

There is no room for compromise regarding the accuracy, completeness, and precision required for FDA regulatory compliance. It’s standard practice for the FDA to include a disclaimer in any warning letter stating that the deficiencies and violations cited are not exhaustive. The FDA requires the pharmaceutical company to investigate in writing any related or ripple effects of the violated 21 CFR regulations, albeit not documented in the warning letter. For example, a deficiency related to 21 CFR 211.192 requires a thorough investigation of any unexplained discrepancies or batch failures, comprehensively addressed for compliance, and not relying on mere FDA Warning Letter written citations. Therefore, other imminent batch-related deficiencies must also be accounted for in the enterprise’s reply to the FDA Warning Letter. Within the fifteen-day reply clock, a very tall order for most. 

The Microsoft FDA OpenAI GPT-4 Copilot, integrated with eSTARHelper’s proprietary SmartSearch+, enhances regulatory intelligence by providing comprehensive insights into regulatory content. This innovative solution enables companies to effectively analyze regulatory reasoning and communication, such as FDA Warning Letters and the associated ripple effects of noted deficiencies. It empowers users to navigate FDA requirements precisely, reduces manual efforts, and ensures compliance with confidence. It does not require downloads and can run on laptops, tablets, and mobile phones. Figures 2A and 2B illustrate how these tools complement each other, enabling organizations to address regulatory challenges effectively. 

Figure 2A: SmartSearch+ Full-Text Follow-Up by You, the Human

Figure 2B: SmartSearch+ can provide specific and detailed information about sections of 21 CFR 211.192 view enabled online instantly by selecting from search result hits  

Enhancing Workflow with Custom AI/ML Applications 

RAG (Retrieval Augmented Generation) Integration  

By leveraging Retrieval-Augmented Generation, organizations can:  

• Upload proprietary data and train foundational models like ChatGPT-4.  

• Seamlessly integrate enterprise data, such as SOPs, into AI-generated drafts. 

RAG implemented on a custom Copilot allows for grounding existing proprietary knowledge and training foundational language models like ChatGPT-4. Data and messages stored in Microsoft Office 365, SharePoint Online, and Teams—often hidden within enterprise data repositories— can now be accessed to create a private, custom ChatGPT-4 model for AI and machine learning applications.  

Enterprise Integration

Microsoft’s ecosystem facilitates: 

• Summarize potentially discoverable deficiencies stored in SharePoint Online or Microsoft 365 documents, such as quality management policies, when compared side by side against actual FDA-published references 

• Incorporation of internal secure authorized access to SOPs for automating response workflows via Agentic AI or autonomous agent without ChatGPT4 conversations, e.g., via an email inquiry 

Authorized Copilot users can authenticate via SharePoint Online to seamlessly access proprietary enterprise data, such as Standard Operating Procedures (SOPs), whether they are stored in Microsoft Office 365 (in formats like PDFs, Word, PowerPoint, and Excel), in SharePoint document libraries, or hosted as SharePoint lists. This ensures that regulatory decisions are informed and timely. 

Consider the added value of using ChatGPT-4 to summarize the deficiencies shown in Figure 1 and your enterprise's private SOPs when drafting your response to the FDA regarding a Warning Letter. Refer to the Figure below.

The above figure demonstrates the value of combining the generative AI foundational model when grounded with enterprise data, showcasing improved efficiency and precision in regulatory responses. For example, a Copilot ChatGPT-4 conversation can include both FDA regulatory mandate summarizations and an ad-hoc comparative analysis against SOPs. 

Conclusion 

Combining generative AI and machine learning (ML) with SmartSearch+ allows organizations to achieve regulatory excellence. This "duelling banjo" approach ensures that AI insights are validated through thorough human-led analysis, resulting in a comprehensive and proactive strategy for navigating regulatory complexities. To showcase the practical applications of these tools, the companion paper titled SMART Practices for Responding to FDA Observational Warnings Using Generative AI places this discussion in the context of the U.S. FDA’s inspectional observations, Form 483, and Warning Letters.