Combatting Medical Misinformation

Problem Background

Around the world, public health administrators are overwhelmed by false medical information and its impacts on their communities. As demonstrated throughout the COVID-19 pandemic, community members’ acceptance of dangerous medical rumors can have severe impacts on their health and livelihood, leading them to  eschew professional diagnoses in favor of scientifically unfounded and potentially dangerous treatment paths. In many cases, this negatively impacts patients’ health and recovery and, in extreme cases, causes loss of life. Currently, public health administrators do not have the ability to rapidly identify and respond to these false medical information trends.

The Spectrum of False Medical Information

False medical information can fall into three primary large classification types:

  • Misinformation: False information that spreads regardless of intent to mislead others.
  • Disinformation: Deliberately created to harm, manipulate, or mislead.
  • Malinformation: Based on fact but taken out of context to mislead, harm, or manipulate.

Each type of false medical information has unique response strategies that can be utilized by public health administrators to effectively combat the threat.

Proposed Solution

Project Heal will be an open-source toolkit that can be easily deployed on Amazon Web Services by entities across the globe.

Developed by a partnership of several research institutions, the toolkit uses cloud-based AI services that allow public health administrators worldwide to be rapidly informed of emerging false medical information threats, obtain insights into them, and generate tailored response communications for impacted communities.

Healthcare Innovation CIC Takes on Initial Designs

In late 2022, the Healthcare Innovation CIC began collaborating with two leading Universities on the aforementioned health misinformation engagement. In May 2023, the project team and other subject matter experts gathered in Washington, DC for an Amazon-style Working Backwards Workshop, where they dove into the customer persona and problem statement. The outputs included a fictional press release/list of frequently asked questions (PR-FAQ) and storyboard, along with a list of features – or user stories – to be developed into a prototype and UI/UX mockup.


The following eight-image panels convey the basic premise of Project Heal. The storyboard tells the story of a child who is sick with “Stomach Virus X” in the hospital; the parents of the child tried to treat them with bleach based on false medical information spread through the media. In parallel, we see a Public Health Administrator, Mary, overwhelmed with the efforts to understand how this piece of misinformation is spreading. She is introduced to Project Heal, allowing her to dive deep into misinformation associated with “Stomach Virus X”.  Ultimately, Project Heal allows her to get ahead of a new threat, “Stomach Virus Z”. Mary’s proactive approach helped prevent the parents of the same child from believing Motor Oil was an appropriate treatment of “Stomach Virus Z”.

Conceptual Architecture for Prototype

From napkin sketch to a stronger conceptual architecture, our team proposed a multitude of services and third party providers to help support the foundations of Project Heal.

Prototype Phase One: Mockups of North Star

Project Heal has an ambitious vision. UI mockups were developed to help demonstrate core functionality of the system to allow customers to have a north star for development. The mockups are tailored towards the Public Health Administrator experience on the platform. Prioritization and having a clear roadmap for development will be critical for customers to create prior to development phases of Project Heal.

North Star Screenshots

Prototype Phase Two: Comm Generation

This prototype focused on experimenting and developing a potential technical approach to the communication generation component of Project Heal.

The prototype utilized Generative AI through Amazon Bedrock to help generate messaging combatting false claims.

Comm Generation Screenshots

Prototype Phase Three: Search Engine

This prototype focused on experimenting and developing a potential technical approach to the search portal of Project Heal.

The prototype utilized pre-staged information to feed into search capabilities. The high-level flow is as follows:

  1. False medical information “documents” (news articles, blog posts, tweets, etc.) are manually collected.
    • EX: “Drinking bleach cures MPOX #bleach #cure”
  2. Documents are run through a process that:
    • Identifies the disease in the piece of incorrect information
      • EX: “Only gay men can get MPox #besafe” would be identified as Mpox
    • Identifies the topic of the false claim
      • EX: “The Mpox vaccine is new and unresearched” would have a topic label of Vaccines
    • Groups ingested statements based on similarity thresholds
      • EX: “Drinking bleach cures COVID”, “Bleach, the cure for covid!”, “If you have COVID, house hold cleaners like bleach will cure you!” should be grouped together. 

The following shows a conceptual view of the prototype:

Classification processes could utilize a mix of generative AI capabilities and services such as Amazon Medical Comprehend.

Search Engine Screenshots

Search Engine Future Enhancements

Below describes how the built prototype could be iterated upon and enhanced.

General Processing Flow

Currently, we are limiting scope to “one sentence” tag lines. Ideally, we should have a workflow as follows:


Given the following three documents:

Doc A (Ingested already to AOSS)

Covid causes autism.

Expected Extracted Misinfo:

  1. Covid causes autism.

Doc B (Newly received doc)

Covid causes Autism. You want to avoid windmills too because they cause cancer.

Expected Extracted Misinfo:

  1. Covid causes autism.
  2. You want to avoid windmills too because they cause cancer

Doc C (Newly received doc)

Covid causes Autism. Parkinson’s is not a real disease.

Expected Extracted Misinfo:

  1. Covid causes autism
  2. Parkinson’s is not a real disease.

Consideration/Design Choice to Make:

“Covid causes autism” would find an exact match in AOSS and skip processing. Generally, this is ok, but we will lose the capability to map/update a reference to a source document that also contained that statement (DOC A/DOC B).  Consider updating logic so that when we have an exact match (IE score == 1), we still map the doc.

Consideration/Design Choice to Make:

One must think through how to structure the document.  A statement can be in multiple documents. Right now, our prototype has a field called similar statement where we capture metadata for that similar statement (which I would imagine now includes the SHA512 mapping back to the source document of that similar statement). Think through if there is a need at the top level of the document for another field entry such as a field called similar-documents that captures the situation above (IE a score==1 as that would not flow into the similar-statement field).

Call to Action

Interested in integrating the concepts of this challenge into your production workloads? Contact us and we can discuss how AWS Cloud and the AWS Partner Network can help accelerate your goals.


PRFAQFictional press release for the CloudPRO vision
StoryboardVisual journey of the concept
Go To Production (GTP) WriteupChallenge documentation and thoughts on how to take to production & scale
MockupsClickable wireframe of the concept
Prototype 2 Backend (Comm Generation)IAC & Backend code for the prototype
Prototype 2 Frontend (Comm Generation)IAC & Frontend code for the prototype
Prototype 3 Frontend (Search Engine)IAC & Backend code for the prototype
Prototype 3 Backend (Search Engine)IAC & Frontend code for the prototype