Building a quality digital medical knowledge base
A carefully curated medical information database lies at the very heart of AI-powered triage solutions offered by Infermedica. In this article, we unveil how we develop and maintain our medical information database to the highest standard.
What is the Medical Knowledge Base?
The Medical Knowledge Base used at Infermedica is a massive database of thousands of symptoms, conditions, risk factors and lab tests intersecting with each other in a web-like manner. Gathered and organized in Metabase—our internal database system—they link with multiple symptoms. Thanks to this collection and its instant data capability, available through Symptom Checker or API-based apps, patients receive a reliable symptom analysis. Access to medically-approved information reassures patients and directs them towards appropriate treatment.
Moreover, the result delivered by the Infermedica inference engine is an important part of triage. Medical facilities, insurance companies, among others, use health and triage information technology to recommend an appropriate level of help (from self-care to emergency room) and route patients to suitable medical centers (which is often crucial for insurance companies). This process leads to the mutually-desired result—the best possible medical support available for the patient, with Metabase at its foundation.
Patient satisfaction and business constancy from many companies is attributed in large part to the quality of this database. That's why we concentrate the same care and focus on the design and development of our medical database as do doctors in the diagnosis of their patients.
“Our Medical Knowledge Base is the very foundation of all the products we offer", explains Jakub Jaszczak, Medical Doctor & Product Manager at Infermedica. "It directly reflects how effective and accurate our evidence assessments are. It is kept thorough and up-to-date, just like the one of an actual doctor” .
What makes our Medical Knowledge Base outstanding
1. The Medical Content Team is composed of seasoned professionals
The medical information database that we maintain was developed in 2012 by over forty physicians. They continuously add new conditions, symptoms, risk factors and lab tests to the medical information database and otherwise enhance its encyclopedic collection.
Our Medical Content Team consists of experts representing different fields and seniority who work collaboratively and collegially building our content with the comprehensive approach to patient health. They also represent different countries and, in some cases, support us to adjust our solutions to be legally compliant in different locations. Each of them has experience as a practicing doctor and is focused on getting the best possible treatment for the patients.
Using their practical professional experience as a backdrop, they seek out the most reliable, evidence-based medical sources and propose new concepts to enhance the database. The concepts of review and verification are the fundaments to the database. Each medical concept is verified by at least two colleagues, as well as proofread by a native language editor. During the peer reviews, they evaluate and correct pieces of information on conditions and symptoms before publication. In parallel, the peer reviews are the time to signal possibly threatening interactions with other concepts in the Metabase. To date, they have contributed over 34,000 hours building and validating the highest standard of medical content.
2. Medical content in the patient’s language
As we construct our database of medical information, we place an emphasis on how patients will be using it and how they communicate their symptoms. Alongside professional medical content, we include common names and phrases of symptoms typically used by patients. This is the key to the database’s effectiveness; laymen terms linked to medical vocabulary. With each new concept, we connect a patient's language with medical terminology in an understandable and unambiguous manner.
"Communication is at the very base of the doctor/patient relationship. If AI applications want a place in medicine, they have to get it right", says Daphnée Dubouchet-Olsheski, Medical Content Editor, Infermedica. "This is where Infermedica excels, with their dialectically-accurate symptom assessment application. Using simple words to explain medical nuances in the app increases patient satisfaction and reduces frustration on the part of the practitioner", she adds.
Our solutions are currently available in seventeen languages. Each of them has its specificities and regional variations. To accurately communicate these regional particularities to app users, we have doctors from different countries on board our medical team. They identify all nuances and, again, make our medical content clear to patients in each area we serve.
3. Trusted medical information sources
Our medical team is deeply rooted in evidence-based medicine. As our content editors initiate new concepts for our medical information database, they rely on respected, proven, reliable sources, such as well-established medical journals (e.g. Best Practice by British Medical Journal, New England Journal of Medicine, Lancet, Clinical Key, Up to Date), guidelines and publications from specialized agencies and organizations operating globally (e.g. WHO, CDC) as well as numerous medical papers with a high impact factor (e.g. NEJM, Lancet).
“Our medical content has been growing over the years. We began with the most common everyday diseases like the common cold or migraines. Later on, we expanded our base with skin conditions, mental health issues and surgical complications. Lately, we focused on assessing injuries and, finally, pediatric conditions", adds Jakub Jaszczak.
How to improve Emergency Departments with AI?
Faith in the ER is being restored. Thanks to Artificial Intelligence predicting patient flow and avoiding unnecessary hospital visits.
Our medical information database also contains concepts that provide information about conditions, symptoms, risk factors and their probabilistic relations for our triage tools, which are regularly updated based on current triage protocols.
Every data item entered into the medical database is meticulously verified and published only after a series of tests.
Explore our Acceptance Test Cases to learn more →
4. Proven process of medical content creation and verification
Our specialty has been that of developing medical content for nearly a decade. During that time we have established a content development process that prevents mistakes and maintains the highest quality of publishable medical information. Our content publication process is innately complex, but its key stages include defining a scope of enhancement to existing content and structure, eliciting expert knowledge, specifying test criteria, performing tests and deploying procedures. Our Medical Content Team adheres to this guidance daily in performing their duties.
“With the database's steady growth, the maintenance of such a plethora of information has become a challenge. Thanks to clear, written processes and meticulous peer review, we ensure that our algorithms are “fed” only quality information, leaving out all unnecessary background noise”, says Anna Nowicka, Head of Medical Content.
5. Continuous verification and update process
Knowledge in all disciplines is rapidly growing but medicine especially is ever-changing. New research, updated guidelines, improved recommendations encourage up-to-speed medical practice every day. Consequently, our medical information database expands with each release. As such, our team is driven to update medical cases that have integrated revised best practices of day-to-day clinical practice scenarios.
Simultaneously, our doctors extend the magnitude and capacity of the probabilistic model that links all medical concepts in the medical information database to associated symptoms, conditions, case studies, etc.
“Both the mathematical inference model and the medical database are complementary elements of our solutions and cannot exist without one another", underlines Anna Nowicka. "We put the emphasis on both researching enhancements of our probabilistic algorithms and keeping our medical content consistently rooted in evidence-based medicine.”
6. Complex testing methodology
Testing of every single piece of data that we enter into our medical information database is our guarantee of a consistent level of superior performance of the entire AI system.
In medicine, error is unforgivable. With patients' health at stake, our applications are run through a battery of tests, including all medical and technical aspects:
acceptance tests - new concepts are tested against model cases in several cycles of manual and automated tests;
expert review - concept is verified by medical professionals from the Infermedica expert panel;
technical review - data scientists verify that new concepts are in accordance with Infermedica internal guidelines and its structure of medical content;
regression testing - a diagnostic model of the new concept is built and its interactions with the medical information database are observed;
manual testing - the final step during which our in-house medical team manually tests conditions and symptoms of the newly introduced concept.
Building a knowledge base for medical purposes – summary
Care and precision are the key elements that allow us to build a highly effective medical information database. The work entailed reaps the reward it delivers - satisfied patients and healthcare practitioners. The medical information database used at Infermedica has proven a 93% efficiency rate in acceptance tests; a laudable feat in an industry overburdened by demand supported by an unequal supply of healthcare providers.
To date, the Infermedica Medical Knowledge Base comprises over 1260 symptoms, 680 conditions and 133 risk factors that interconnect with one another and allow accurate symptom assessment of patients and support triage, having already assessed the symptoms 6 million patients and seeing that number grow every day.
You can test it out with Infermedica API →