3 steps to successfully designing and implementing the technology for a med affairs AI solution
This post is part of a series describing how to implement AI solutions to extract value from med affairs notes. Through our collaborative projects with clients, we apply a systematic approach to design fit-for-purpose solutions that meet our clients’ needs. In this series, we introduce the common challenges we see life science companies encountering and how these can be successfully overcome.
In our previous post, we discussed the important concept of organizational pull-through for successful adoption. This blog post continues the conversation by expanding on the next 3 steps for designing and implementing the technology for a med affairs AI solution, as introduced in our first post:
- Obtain organizational buy-in
- Optimize your data foundation
- Process the data to generate accurate, relevant insights
- Visualize and report the insights for all stakeholders
Step 2: Optimize your data foundation
Insights gathered from medical notes are only as good as the data they are based on, as the saying goes: “Garbage in, garbage out.” Therefore, the first step in any data-driven technology process is to make sure the content being ingested by the system is robust. For med affairs notes, this means that the information is tagged and categorized appropriately and accurately for near real-time summarizing, analyzing, and reporting of field trends and patterns.
How this is achieved using NLP and GenAI
When working with med affairs notes, natural language processing (NLP) or GenAI processes automatically tag each medical insight for the content or overall topic. This creates the foundation for further data processing and analysis through:
- Process-driven tagging, which eliminates human bias and rationale to improve the accuracy and consistency
- Faster tagging that can be performed at any frequency, such as immediately upon entry of new data, once an hour, or once a day
- Multi-modal processes that can ingest and infer tagging from any input such as text, images, and html files
- Natural language instructions that do not require code, which makes it easy for the user to just provide tagging instructions rather than having to know computer code
Examples of automated tagging of notes from a CRM
Example entries in the notes | Tags assigned by the system |
|---|---|
“Asset 1 has the highest safety and tolerability for XXX indication. However, there are issues with YYY.” | Asset, Efficacy & Safety, Indication |
“Asset 2 is likely going to produce better clinical results than comp 1. It will be interesting to assess which regime will have the best efficacy profile for patients and how it will impact corresponding access restrictions.” | Asset, Clinical Biomarker, Competitor, Efficacy & Safety, Indication, Access |
“Asset 1 might not be a strong drug for XXX indication, Asset 2 hepatotoxicity is not an issue and has a similar clinical profile as Asset 3” | Asset (1, 2, and 3), Efficacy & Safety, Indication, Side Effects |
Roadblocks that often delay or derail AI solution implementation
Many general off-the-shelf systems struggle to complete this step well because medical affairs content is highly specific and technical.
Common challenges | Recommended solutions |
|---|---|
The system misclassifies or misses information, requiring a lot of human input and correction for accurate tagging. | Find a solution designed for the life sciences industry that can quickly and accurately interpret specific terms and how they are used. |
Teams lack confidence in the outputs of an AI solution. | Use best practices for data preparation, validation, integration, and management; insert human experts as auditors in the process; and ensure transparency throughout the system build. |
Insights are delayed because the frequency at which new information is processed is once a quarter or once every 6 months. | Develop a more frequent feedback loop that allows near real-time ingestion of new data. |
There is a risk of commonly available systems such as ChatGPT adding sensitive data to its public domain. | Implement appropriate data security processes within a system to prevent data from being shared outside the company. |
The solution quickly becomes out of date within the constantly evolving clinical and medical landscape. | Design highly configurable processes that can accommodate changing medical strategies or the identification of a new signal. |
Step 3: Process the data to generate accurate, relevant insights
Using the data foundation established in the first step, the tags can now be analyzed to consolidate the findings and assign notes categories. This step forms the basis of understanding, visualizing, and reporting the trends, topics, and sentiments of the information.
AI streamlines this process and ensures greater consistency in the outputs, allowing med affairs teams to quickly derive meaningful insights from the vast amounts of unstructured data in their CRM.
How this is achieved using NLP and GenAI
There are several processes that happen simultaneously to dive deeper into the information captured in the notes. In one of these processes, the structured insights from the previous step are fed into large language models (LLMs) to perform sentiment analysis, which reveals the underlying emotions and attitudes expressed in the notes so the context around the information can be better understood.
The structured data are also automatically categorized to guide how they can be used. For example, they could be categorized as:
- Actionable insights, or a list for the med affairs team to build strategies around
- Competitive intelligence to better understand the role of other companies within the space
- No action, or items that can be “archived” and resurfaced again later if needed
Note classification based on the assigned tags
Example entries in the notes | Assigned tags | Notes category |
|---|---|---|
“Asset 1 has the highest safety and tolerability for XXX indication. However, there are issues with YYY.” | Asset, Efficacy & Safety, Indication | Insight |
“Asset 2 is likely going to produce better clinical results than comp 1. It will be interesting to assess which regime will have the best efficacy profile for patients and how it will impact corresponding access restrictions.” | Asset, Clinical Biomarker, Competitor, Efficacy & Safety, Indication, Access | Insight and intel |
“Asset 1 might not be a strong drug for XXX indication, Asset 2 hepatotoxicity is not an issue and has a similar clinical profile as Asset 3” | Asset (1, 2, and 3), Efficacy & Safety, Indication, Side Effects | Insight |
Topic modeling also assigns tagged insights to predefined topics for the industry or to newly created topics, if needed. These can include indirect references that could be considered as being in a “gray area” but that GenAI can interpret as belonging to a specific topic.
Domain-specific topic modeling for key themes and concepts
Roadblocks that often delay or derail AI solution implementation
Widely available models are not trained on the medical literature, scientific concepts, and regulatory-standard documentation, resulting in outputs that are not relevant or do not instill confidence in its users.
Common challenges | Recommended solutions |
|---|---|
The system misses the nuances of industry-specific terms, acronyms and abbreviations, drug names, and manufacturer names. | Start with a platform trained on life science language models that easily understands the terms as well as the context, which will require fewer iterations and therefore shorten start-up times. |
The insights provided by an off-the-shelf solution are not immediately usable, and it takes a lot of time to train the model to work as intended. | Use domain-specific ontologies that enable topic modeling to accurately identify and categorize key themes and concepts within call notes. |
Hallucinations and inaccurate insights are delivered because GenAI models are allowed to be the final decision-maker amid a changing landscape, resulting in model drifts and drops in concordance. | Include a human expert in the feedback loop during development for better model training, and use hypercare setups for periodic monitoring of the outputs by humans who know what “good” looks like and can tweak the processes if necessary. |
Step 4: Visualize and report the insights for all stakeholders
Creating visualizations and reports of the data allows stakeholders across the organization, from the med affairs team to executive leadership, to easily understand and use the insights. The outputs can range from a quick overview of trends for leadership to more granular insights to help field reps tailor their interactions with HCPs.
How this is achieved using NLP and GenAI
GenAI generates summaries of the structured, categorized information from the previous two steps. Because those steps process information in near real-time, these summaries reflect the latest information available and enable timely decision-making. These can be tailored to the user’s role (e.g., executive summaries vs granular insights) and refined using filters.
Q&A functionality allows users to ask questions about the content and receive contextual responses. When the responses include references to the content that contributed to the answers, users can read further or verify that the content is relevant to the question.
Roadblocks that often delay or derail AI implementation
The purpose of categorizing and synthesizing med affairs notes is to be better informed about the landscape, and easy-to-use, understandable visualizations and role-specific reporting are fundamental components of that process.
Common challenges | Recommended solutions |
|---|---|
Different users require different data access. | Build user or persona-based permissions into the process pipeline. |
Versions are not managed properly, causing confusion and duplication in the data shown. | Implement a governance process to ensure only the latest version is used. |
The system needs to provide the information in different formats and levels of detail for different stakeholders (i.e., key highlights for the executive leadership but detailed HCP input for the medical science liaison). | Create different views, such as detailed report with multiple filtering options, include chatbots for users to retrieve answers quickly, and use GenAI-driven summaries. |
Users are uncertain about the recency and timing of information being shown. | Generate the visualizations and reports in real time so the information is always the latest. |
Key takeaways
When successful, introducing AI in med affairs, for which AI is highly appropriate, within the organization’s AI road map could be an early use case for later, larger-scale adoption for other use cases throughout the company. The following should be considered when embarking on the design of a new solution:
- Ensure the technology can be updated to incorporate of the latest advancements
- Implement a system with industry-specific ontologies, libraries, and models, for time savings and greater accuracy
- Adopt a co-development approach with a technology partner that results in fit-for-purpose, adaptive models
- Incorporate change management starting at project start-up, to ensure both digital and organizational transformation
Shift med affairs teams’ time to more valuable work
NLP and GenAI can augment the expertise of med affairs teams by quickly consolidating, understanding, and summarizing call notes and other information collected by the reps. Companies using these technologies benefit from using fewer human resources for these tasks, which frees their time for more value-add work, and reduced human errors during manual tagging, which enhances the accuracy of the insights and reporting used for decision-making.
Download our case study to read how the approach in this blog series helped a biotech company’s medical analytics team deliver real-time intel and decrease the manual effort and error. Reach out to us if you’d like to discuss how to implement AI-enabled solutions for your med affairs team.
