The role of 'initial symptoms' in the quality of symptom checking tools

Did you know that something as small as a single initial symptom can determine the whole pre-medical interview? As an effect, it can also negatively influence the quality of digital symptom checkers.

In this article, we look at the 'initial evidence’ used in the symptom checking tools based on Infermedica API.

If you are planning to design or develop a symptom checking app, or a health chatbot, on your own - stop for a moment. Imagine yourself in the shoes of a physician meeting a new patient. This patient tells only about one symptom he's experiencing (e.g. fatigue) and the doctor asks him many more questions to understand his case. Unfortunately, there are too many possibilities related to this symptom, and the symptom itself is rather general.

“The majority of symptoms are nonspecific and can be present in multiple conditions” comments Otto Krawiec, GP and medical content editor at Infermedica. “With only one symptom, the physician would use a top-down approach, some questions may seem irrelevant for a patient but a physician is trying to pick up his trail.”

The investigation of multiple hypotheses would take a lot of time from the interview. In the time in which the diagnosis and treatment should be proposed, the physician would keep asking general questions and trying to find the right path.

Time is running, the patient is becoming impatient and maybe even thinks that the physician is not competent enough.

Would it be different with more initial symptoms given by this patient?

“Let us imagine another situation - a patient reports fatigue, weight gain, and constipation. A physician is already on the right path, as hypothyroidism has become one of the most probable conditions, while other diseases are considered as possibilities. Right from the start of the interview, he’ll ask more accurate questions. Our engine works the same way; by allowing and encouraging patients to input more initial symptoms, you will achieve more customized interviews”, adds Otto Krawiec.

In this regard, symptom checkers are very much like physicians. Of course, they are not going to replace doctors, but they use similar patterns to learn about the patient and lead them towards the best medical help.

By initial symptoms, we understand any signs, from headaches to bleeding, that the user is adding to a symptom-checking tool in the first steps of the interview. Complemented with risk factors, such as pregnancy, diabetes, or smoking cigarettes, they are used by the digital triage engine to investigate any probable health conditions. We should also remember that the digital triage engine checks multiple possibilities and learns about the user through evidence-collecting questions to better understand their health, and recommend the most probable evidence assessment.

Therefore, the more symptoms the engine knows about the user at the beginning of the interview, the more accurate the questions it presents during the interview, and the more exact the health assessment at the end is.

On one hand, the more symptoms given, the more accurate the initial health assessment is. On the other, in some cases a large number of general symptoms can make the interview longer, as the engine will investigate all of them. We see this number somewhere between two and four, but the ultimate goal is to look for the balance between the number of initial symptoms and the length of the interview, as it is the element that influences the user experience.

The most important role of initial symptoms is the quality of the interview and the resulting preliminary health and triage assessment. More initial symptoms give:

• a more unique dynamic interview

• less random questions

• a higher accuracy of the recommended well-being analysis

• a better chance to get the proper medical help

But for many users, this would also translate into a positive perception of the tool itself. A higher adjustment with the exact case of the patient will be seen as a smarter and better-designed tool, with higher attention to details, as well as being more credible and reliable.

No matter what the optimal number of initial symptoms might be, our goal is to listen to users. It is up to them how many symptoms they experience and want to share. Some people will feel comfortable with telling more about their signs, some others not. In some cases, users won’t be fully sure of what they’re feeling. For those users, our /suggest endpoint may be helpful as it states other symptoms related to a given initial symptom. In all other cases, questions generated by the inference engine will try to fill in the gap.

Our role is to open the tool to the user as much as possible and convenience them to disclose as many initial symptoms as they want to.

Although organizing the user journey is important, in the case of initial symptoms, it is better to leave as much freedom as possible.

The examples above are a clear way to push the user away from adding more symptoms. How then you can construct your symptom checking tool to encourage users to share as many initial symptoms as possible? Here are some tips divided by the type of tool:

From the user perspective, symptom checkers are probably the most flexible, easy-to-use, and fast tools to investigate symptoms on the Internet. With them, users don't have to go through the countless number of articles and verify them on their own. Users can choose how they want to use them. On a desktop or on a mobile, it is up to them. They also give the most advanced options to ask about initial symptoms.

Start with small things. On its multiple screens, always use a plural form of "symptoms", or "signs". This way, you are already building the awareness of more than one initial symptom.

Avoid constructing the interview by starting it with adding one chief complaint. Instead, give the users convenient tools to add their symptoms. There are various ways to do it: symptoms search, lists, or even body avatars. If possible, try to use two ways at once - users will pick their favorite ones. What's more, remind your users about the importance of adding more symptoms with additional communication after adding the first symptom, like: "Please try to add more than one symptom."

Check if all the ways to add symptoms work equally well on mobile and desktop devices.

Although the main principles are the same as in symptom checkers, chatbots are moving us towards more conversational forms. Do your best to build a friendly and honest dialogue between the bot and user. Start with the greeting and presenting the goals of this chat and, at the same time, look for something short like: “Hi! I'm an automatic symptom checker. I'll try to help you find the possible causes of your symptoms.”

Then move towards the symptoms themselves. Look for open questions asking your users about their symptoms. "Please describe your symptoms". "Tell me about your symptoms". "What symptoms are you experiencing today?". Use the plural form to suggest more than one symptom.

With each answer from the user, confirm if this is everything they wanted to list. A short "Do you want to add anything else?" will do it well.

Some user's answers might not be clear for the engine, as it doesn’t understand full sentences. However, our built in NLP will identify them and connect them with the respective medical terms. For example, “I have a headache” will be translated by our NLP to “headache” symptom.

Also, if there is no such answer in the engine, transparently communicate it with, e.g. “I'm sorry but I didn't recognize this symptom. Can you describe it with other words?".

Infermedica API enables you to also build tools dedicated to medical staff, like nurses or agents working at call centers. In this case, you can use communication similar to symptom checkers or medical chatbots, but via another perspective.

The right way to go is also to educate the team about the importance of initial symptoms and equip them in the sheets with possible scenarios or general questions, helping them to learn more about the patient.

Another idea is adding the NLP endpoint to your solutions, which enables the recognition of conversation and marking keywords connected with symptoms or risk factors. Natural Language Processing technology allows us to pick even patients' common language and translate it to the medical terms used in the inference engine. This way, the nurse or agent will be supplied with a list of initial symptoms to confirm and prioritize before guiding a patient to the right help.

The ‘Initial symptom’ is an indispensable element behind the symptom assessment capabilities of Infermedica’s intelligent engine. Learn more about indicating initial evidence.

Want to see more examples and ideas? Visit Symptomate and check its interface.

BL/EN/2021/04/14/2