CSR.Gen

A Clinical Study Report (CSR) is a document that describes the methods, results, and conclusions of a clinical trial. The CSR is a highly structured document that follows the format outlined in ICH E3 Clinical Study Reports. One of the most time-intensive aspects of preparing a CSR is the review and description of the safety data from the clinical trial. To perform safety data review and description faster without compromising quality, Pfizer seeks a solution that automatically generates CSR text based on tabular data. In the envisioned workflow, CSR authoring teams will use the solution as a tool in crafting the CSR.

Pfizer challenges applicants to develop and demonstrate Generative AI solutions to automate the generation of the Clinical Study Report (CSR). Pfizer will provide training data in tabular format as input and corresponding output texts. This solution should utilize the tables to achieve a concise and efficient output that adheres to the Lean Writing Style. The solution should generate summary text for the sub-sections on Adverse Events, Laboratory Results, Vital Signs, Electrocardiograms and Physical Examination Findings within the Safety Evaluation Section of a CSR. 

Success will be assessed by comparing the accuracy of the generated content with human-generated content across each of the following sections: 

1. Safety
   
   1. 1. 1. Adverse Events
            
            1. Brief Summary of Adverse Events
            2. Deaths
            3. Serious Adverse Events
            4. Dose Modification and/or Discontinuations from Study Intervention or Study Due to Adverse Events
            5. Adverse Events of Special Interest
            6. Other Significant Adverse Events
         2. Clinical Laboratory Evaluation
         3. Other Safety Evaluations
            
            1. Vital Signs
            2. Electrocardiograms
            3. Physical Examination Findings
            4. Safety Observations Related to [Medical Device OR Combination Product]
            5. Other Observations Related to Safety

• Potential participants should submit a solution proposal from which a set of participants will be selected (see Selection Criteria for Success)
• CSR templates will be provided to the selected participants
• Final CSRs and corresponding tables and listings for 70% of the studies will be provided for training
• Challenge participants will be provided with the remaining 30% of studies’ tables later and asked to produce the CSR documentation section within 2 to 24 hours
• Pfizer SMEs will then evaluate the results
• Solutions should leverage large language models (LLMs) such as BLOOM, LLaMA. NeMo, PaLM or any custom models (see conditions in the next section).  An Azure GPT 3 service is available with restrictions.
• Participants will have an opportunity for a demonstration/presentation of the solution and discussion of the models and algorithms used

• Applicants are free to train and deploy models in the BCA environment setup by the Pfizer team and are fully responsible for any models deployed
• Data must remain in environment and cannot be downloaded
• All fine-tuned models developed during the course of the challenge cannot be used for purposes outside of this challenge
• Solution approach must be compliant with all applicable legal, privacy, and regulatory requirements, including Pfizer policies on AI
• Additional terms and conditions may apply

• Proposal addresses the challenge statement, and is designed to deliver the expected outcome at scale
• Explanation of planned approach for training, testing and evaluating model outputs and algorithms used
• Documented success with existing clients in similar clinical development use cases
• Experience working with Pfizer, and/or the pharmaceutical industry especially clinical development processes including generation of Clinical Study Report

1. Overall proposed solution
    Ability to generate objective Safety Evaluation from tables for Adverse Events, Laboratory Results, Vital Signs, Electrocardiograms and Physical Examination findings with:
   •    Accuracy: The generated text must be factually correct and consistent with the input data.
   •    Source: 100% of content should be derived from source documents
   •    Completeness: The generated text must be complete with all of the relevant information.
   •    Provenance of the relevant information.
   •    Quantitative assessment (e.g., BLEU, ROUGE, METEOR scores)

2. Ability to best meet Pfizer’s objectives and requirements for:

Safety Profile: Ability to detect patterns and establish a safety profile by identifying trends and corresponding patient Adverse Events, Laboratory Results, Vital Signs, Electrocardiograms and Physical  Examination Findings data, informing potential dosing strategy and prophylaxis

3. Technical capabilities

Ability to generate consistent and varied results based on inputs, prompts, models within a reasonable amount of time

4. Solution must be compliant with all applicable legal, privacy, and regulatory requirements

5. Solution should show potential for extensions to other protocols and therapeutic areas in clinical development programs

6. Implementation approach and timelines

Reduction in cycle time in CSR generation, specifically through the textual summarization of standard safety Tables.  

Opportunity for Collaboration on implementation 

To avoid any doubt, winning the challenge does not automatically leads to further legally binding contractual agreements with Pfizer. The detailed terms of possible future agreements between Pfizer and the winner (or any of the entrants) will be negotiated separately and will not be linked to the Term set herein

• Challenge posted – May 31st 2023
• Applicants to submit the proposal by June 14th 2023
• Pfizer to evaluate the responses based on selection criteria
• Finalize applicants for challenge – July 7th 2023
• Challenge start date: Aug 1 2023 (it is subject to legal and technical prerequisites for the challenge)
• Challenge end date: Sept 08, 2023 (may change depending on the challenge start date)

Questions? Contact [email protected]