AI answer engine for doctors and patients

Solution

Itransition implemented a PoC of a web application that interacts with the Davinci AI model to give answers to visitors’ questions.

Itransition’s team began by testing the feasibility of the AI answer engine idea by looking for an AI model that:

• efficiently processes language queries (requests for medical definitions, abbreviation explanations, descriptions, advantages and disadvantages of procedures, etc.)
• is accurate enough to be used in the surgery niche of the healthcare sector

We suggested OpenAI since, in addition to its automation and time-saving features, it is efficient and secure and could be trained by the customer. Instead of AI models, we considered integrating Azure services into the solution’s architecture since Azure offers native integrations between OpenAI and Cognitive Search, provides indexing and querying capabilities for data, and has great AI model deployment possibilities. We also tried to uncover and address all possible AI implementation challenges to implement the best-fitting solution.

OpenAI also offers a fine-tuning service which was a relevant capability for the customer. That is why we tested the four models with fine-tuning capabilities (Davinci, Curie, Baddage, and Ada) available at the time of the PoC creation. In the end, we chose Davinci for its size, high accuracy, capability to perform complex language tasks, and ability to be trained on larger, diverse datasets.

The web app serves as an admin space for the developed AI answer engine and offers the following capabilities:

• Add, edit, and create question-answer lists
• Train the AI model
• Track incorrect answer history
• View question time stamps, questions, answers, and model versions
• Monitor the status of question-answer pairs (applied, registered, in progress, etc.)

The AI model is fine-tuned using the knowledge base built by the customer that includes more than 100 question-answer pairs. To ensure the model gives accurate answers to very different user questions, we paraphrased the most frequent questions and fed them to the AI model. The PoC showed that certain answers were on par with the required quality level and others were not. The answer-question pairs were improved by the customer to contain only proven information and produce well-grounded answers.

The AI answer engine was developed by Itransition to continuously expand the knowledge base with the help of fine-tuning happening overnight and overseen by the customer. The Davinci model uses the previous model as the foundation for the next one during the fine-tuning process. On top of that, the PoC developed by Itransition’s team ensures easy scalability of the future solution, meaning that admins can extend it with as many question-answer pairs as needed and increase the number of useful features.

Based on our experience in other healthcare industry projects where we handled legal, ethical, and regulatory requirements, Itransition recommended adding a disclaimer feature to the solution. The disclaimer precedes all answers, warning users about the limitations of AI and highlighting the importance of qualified professional advice for accurate medical guidance. Itransition also implemented a method of garnering answers from scientific literature quotations while observing the author’s copyright.

Having completed the PoC, Itransition suggested creating a product development plan that included the cost and labor estimates for winning future investment opportunities. Additionally, we delivered estimates for further development enumerated below:

MVP quote. Itransition put together an MVP documentation package with details about team composition, characterizing project roles for software and QA engineers, a project manager and business analyst, and a UX/UI designer. The MVP quote describes a browser-friendly responsive web app and includes monetization options like monthly subscriptions, pay-per-question, pay-per-bundle of questions, etc. We also suggested adding the ability to embed the AI answer engine as a chat widget to the admin panel, as well as hosting the solution in the cloud environment.

Integration estimates. Itransition’s team calculated estimates for developing the solution further and integrating it with third-party apps. For instance, we estimated the integration of ChatPDF AI with the developed AI answer engine. Azure OpenAI Service and Azure Cognitive Search were utilized for embedding vectors as a numerical representation of data that captures semantic relationships and similitudes. Text data from PDFs was vectorized by finding text extracts that fit semantically to produce an answer. Additionally, Itransition generated an estimate and quote for integrating the AI answer engine into the customer’s website using various approaches.