DXC’s FirstPoint RIM.AI Medical document writing solution powered by Amazon Bedrock

By Hiranya Bharadwaj — DXC Technology
ByLucy Bailey — DXC Technology
By Akash Subramani — DXC Technology
By Dhiraj Thakur — AWS

DXC Technology

Global Pharma companies need to prepare extensive documentation to seek regulatory approvals to market their drugs in the country. For each new drug application, thousands of documents must be drafted and reviewed for quality. This manual process poses a high risk of errors late in the drug development cycle, making it a top target for automation using artificial intelligence (AI) and machine learning (ML).

DXC Technology and AWS collaborated to jointly develop a cloud-native AI framework that addresses multiple use cases to help pharma companies, de-risk their Regulatory Information Management [RIM] workflows, such as drafting technical summaries, checking document quality, and translating documents, with the intention to add other use cases over time.

This post discusses how DXC Technology and AWS jointly developed the cloud native AI framework that enables pharmaceutical companies to realize compounding cost savings. This framework allows a pharma company to onboard multiple automation use-cases over time.

DXC is an AWS Premier Tier Services Partner and Managed Cloud Service Provider (MSP) that understands the complexities of migrating workloads to AWS in large-scale environments, and the skills needed for success.

Solution architecture 
The following diagram illustrates solution architecture of DXC’s FirstPoint RIM.AI framework on AWS.

Figure1 – Architecture diagram

The architecture workflow is described below.

1. FirstPoint RIM.AI uses REST API to accept requests from/ send response to business applications

• • RIM.AI receives the request from the user with references to electronic Common Technical Document [eCTD] template documents.
    • Requests are processed based on business logic defined/ prompt templates available.
  • RIM.AI pulls the related documents and metadata from the repository to a temporary area, compiled into a domain of data.
  • RIM.AI reads the template for prompts and then, for each prompt found:
    •  A request is made to the specified AI engine to summarize or pull out the relevant sections from the source documents.
    • The prompt embedded in the document is then replaced with the AI output, inheriting the embedded styles in the document.
  • The generated document is then returned to the repository and, if configured, a workflow can be initiated on the document.

2. Requests that require data aggregation are executed by first creating a domain of data relevant for the request and associated content files

1. • The AWS Lambda function uploads and erases both domain of data and content files in Amazon S3 for Amazon Bedrock.
   • Amazon S3 Bucket is the temporary storage for the domain of data and content files during an active session.
   • All Event Notifications are stationed in Amazon Simple Queue Service (Amazon SQS), which triggers the Lambda function to initiate action and later purge the queue once session is over.

These three forms a basic component unit required for all additional use cases to exercise control on data/documents for security and regulatory compliance standards practiced in the industry

3. The execution path for the request can be sequential/parallel based on discrete-workflow and applicable AI use case [which are themselves modular].

• • For document generation:

• • • The Amazon Kendra is used as Vector Database and Search Engine / Index.
    • In Lambda function, the LangChain Framework provides modules to integrate with Models and Orchestration tools.
    • Amazon Bedrock offers a choice of Foundational Models to design Retrieval Augmented Generation (RAG) approach.
    • The domain of data sets the context perimeter for the Anthropic LLM model to execute the request against.
    • The output is routed back to FirstPoint, put in the required document format, then sent to the end user.

• • For document classification:
    • The Lambda function runs asynchronous Jobs for custom analysis of input document against test documents and/or data models like DIA reference models.

• • • Amazon Comprehend then outputs a Classifier score of the document.
    • Comprehend returns a classifier score to FirstPoint user.

4. Application Load Balancer increases response time and reduces network latency.
5. AWS Key Management Service (AWS KMS) is a managed service that creates and controls cryptographic keys for data protection.
6. AWS Identity and Access Management(IAM) enforces the organization’s existing policies.
7. Amazon CloudWatch to monitor and optimize system performance and operational health
8. AWS CloudTrail to enable operational and risk auditing, governance, and compliance of AWS system.
9. AWS Security Hub automates security checks, aggregates alerts, and enables automated remediation for your AWS system
10. Rest API is used to communicate with data & document sources and business applications. This makes the framework modular and scalable, and additional automation use cases can be built and bolted on for long-term return on investment. Here are some examples:

• • Document Validation: Automate your document rendition and quality checks with suggestions and automatic operations at a click.
  • eCTD Submission Assembly: Assemble your documents into a submission tree structure automatically.
  • Regulatory Intelligence: automated regulatory intelligence feed based on therapeutic focus area.
  • Tracker: Automatically enable end-to-end submission planning based on your existing data.

DXC’s FirstPoint RIM.AI is the orchestration layer developed with decades of DXC’s experience and best practices in building validated software for the pharma industry to manage regulatory information. RIM.AI intelligently scans all relevant documents, aggregates the content in a domain of data that can be packaged together with a prompt for the AI engine to summarize.

DXC has also devised a methodology to ensure that use cases are defined with proper metrics for business impact and return on investment calculation. This ensures that business aligned use cases are scaled up to compound productivity gains and savings while not compromising on any risks.

Implementation, scale up, and ML-Ops
DXC has developed the framework to baseline the current process, evaluate automation potential, and measure productivity gains for any automation use case with AI to bring better quality lifesaving drugs faster and cheaper to patients.

To be able to measure success of any initiative, it needs to be measurable. DXC’s recommended approach is rooted in this aspect, as the first step in AI adoption journey is to baseline current process, measure AI productivity impact at POC level and continuously after as the organization scales up the use case.

Here is a simple agile exercise that ensures any AI use case is foundationally built for success.

•  Week 1: Shortlist three workflows to focus; define business metrics for success
• Week 2: Setup Pilot system; Map Value Stream of focus workflows
• Week 3-4: Weekly sprint to review and refine AI enabled workflow
• Week 5-6: Review final output and business alignment

The framework is fundamentally scalable while using resources economically based on workflows for optimal performance. DXC recommends a matrixed support model for operations to ensure continuous business alignment. The matrix support model constitutes technology as horizontal; industry knowledge as vertical; and validation as diagonal element. All three axis have to be harmonized for the use case to be scaled up successfully over time to compound the productivity gains and savings.

 The solution can be applied to several industry use cases, including:
 Life Sciences industry use cases

• Document classification: Classify your electronic Trial Master File [eTMF] documents automatically when imported into your electronic Document Management System [eDMS].

• Document generation/ translation: Generate your first draft of submission documents from your clinical trial/existing submission data and translating them into various languagesLS-RIM.AI is a headless service that takes the following from a source DMS when requested
  • A template with AI instructions and a configuration set defining how the AI should be controlled.
  • Either one or both:
    • A set of source content documents
    • One of more sets of metadata from other relational database or data sources compiles the content and data into a domain of data and pushes this to a target AI system.

Once a response is returned to LS-RIM.AI, the service composites the template with the responses to create an output document and publish this document back to the source DMS with optional post activities defined.

Use cases for Other Industries
 While this was developed with a specific industry in focus, i.e. Life Sciences; at a basic level the workflow automation is scalable across industries, for example  Aggregate documents from variety of sources and package them into domain of data along with the smart prompt enabled templates that is input to AI engine for processing output at scale.

• Consumer Retail industry: create localized/individualized promotional materials in multiple languages at scale and content quality control.

• Supply Chain departments: get visibility, enforce, and manage compliance/renewal of legal contracts

• Media organizations: quality check content and generate out promotional messaging in multiple languages across variety of formats

DXC FirstPoint RIM.AI Key Features

• This framework applies the blinders on the AI engine to focus on the domain of data and reduce risk of hallucinations in the output.
• Takes the AI engine output and processes into the applicable template and pushes out to the output location like file share or eDMS.
• Enforces all security parameters on the domain of data, like deleting the data from all instances after the AI request is processed.
• Enable validation compliance of the ecosystem.
• Acts as an orchestration layer to execute additional/ multiple use cases over time to scale the system for better return on investments.

Customer benefits
Pharma companies can expect the following benefits to begin with, and the benefits grow as more use cases are onboarded:

• 40% faster regulatory submissions.
• 50% improvement in cost efficiency across regulatory organizations.
• 2x reduction in quality issues.

Conclusion 
A single AI use case is unlikely to justify the investment for most organizations. To realize a compelling return on investment, companies need a framework that allows them to rapidly deploy multiple AI use cases over time. DXC Technology and AWS jointly developed the cloud-native AI framework for regulatory information management, taking this scalable approach. This framework includes AI engine that can be trained to align with an organization’s business focus, and the agility to evolve to continue delivering productive gains and compound business benefits.

By adopting this framework, companies can start with critical AI use cases like automating submission document drafting. They can then expand to other AI use cases for document quality checks, translation, risk assessment, and more.

.

.

DXC Technology – AWS Partner Spotlight

DXC Technology is an AWS Premier Tier Services Partner that understands the complexities of migrating workloads to AWS in large-scale environments, and the skills needed for success.

Contact DXC Technology | Partner Overview | AWS Marketplace | Case Studies