Implementing Generative AI for Medical Standard Response Document Creation
Clarkston Consulting recently partnered with a biopharma client on a data and analytics project in which the team used generative AI for medical standard response documentation. Read a synopsis of the project below or download the full case study.
Download the Generative AI for Medical Standard Response Documentation Case Study Here
Clarkston recently partnered with a leading biopharmaceutical company to streamline the creation of standard response documents (SRD) to medical information requests (MIRs), which are pre-made answers to a variety of medical questions from doctors, ranging in topic from dosage to adverse events. The client wanted to accelerate the creation process of the SRDs so that the collection of documents could be quickly augmented as new questions were raised by healthcare professionals (HCPs). As a relatively less-known pharmaceutical company, this was a key opportunity to create a great experience for HCPs in an interaction point with the company and brand. Clarkston designed a framework where Large Language Models (LLMs) provide a first draft and the medical team provides only validation, vastly reducing effort.
Clarkston’s data and analytics team created a custom chatbot with a knowledge set consisting of several research publications and a prescribing information packet. Considering the limitations of current LLMs, the Clarkston team set targets to determine what would constitute sufficient similarity to the content and style of existing human-written SRDs. To be useful to the medical information team, the product needed to function well for the entire set of SRD topic areas that could range from adverse events to dosage instructions. The team used a supervised learning approach to execute the prompt engineering process. By creating a feedback loop, the team accelerated the prompt engineering process to create a sophisticated set of instructions for the model, leading to consistent and accurate SRD creation for specific topic areas.
By clustering the topics of the existing SRDs, the team determined several topic areas. Each of these areas pulled in a unique set of information from the knowledge base and contained topic-specific formatting. This breakup allowed the team to engineer prompts tailored for each topic area, which improved the content created by a single prompt. However, doing so required an extension of the chat model’s pipeline so that the LLM could be queried multiple times to first assess the topic area and then create the SRD with that context.
The team used the OpenAI Enterprise GPT functionality to create an LLM that could be accessed by all business users. This tool provides a straightforward configuration interface and includes a user interface for ease of business use. The team also used Python and the OpenAI Assistants platform to refine the prompts, self-query, and modify response formatting. The Assistants platform, while needing a separately developed user interface, allows for more rapid development through programmatic access via an API call.
Thanks to Clarkston’s efforts, the client had a roadmap of several options using their data to quickly draft medical SRDs. The Clarkston team advised and upskilled the client for continued internal capabilities leveraging LLMs.
Download the Generative AI for Medical Standard Response Documentation case study, and learn more about our Data & Analytics Consulting and Life Sciences Sales & Marketing Consulting Services by contacting us below.
Contributions from Xavier Brumwell
