How Healthily understands your symptoms – our technology explained

18 February 2022 in Tech

When it comes to supporting health with digital technology, the goal is to find the best next steps to every user’s problem.

At Healthily, we’ve worked hard to reach those quality next steps by developing a self-assessment tool that improves continuously with the help of AI. The result: a tool that gives users everything they need to review symptoms, find relevant health information and advice on the best next steps

But how did we get here?

We’ve used three types of logic to understand your needs. These might be used alone, or in combination, it depends on the problem at hand. When designing the Healthily self-assessment tool, it became clear that each of these types of logic has benefits and limitations, so understanding each one has been key.

Logic type 1: Rules

Conditional ‘if-then’ logic has a clear advantage when you need to ensure something happens 100% of the time. But this certainty comes with a major downside: lack of flexibility and scalability. Complex chains of ‘if-this-happens-then-that-happens’ quickly become unmanageable and end up creating logic loopholes.

So we use these rules when we need certainty – and to ensure our recommendations are safe. For example, if a user mentions ‘blood in vomit’ this triggers a rule that the user will be told to visit A&E in their assessment report.

Logic type 2: Bayesian inference

This is a method of statistical inference, named after its developer Reverend Thomas Bayes. He died in 1761 before he could publish his work, but it’s now widely recognised as an approach to complex problems where new information is constantly changing the probability of the outcome.

We use Bayesian logic to estimate the probability of a condition being the cause of the symptoms reported. The logic can also acknowledge personal factors like a history of smoking or diabetes.

The strength of this approach is in its mathematical foundations. Unlike digital software, the human brain is not well-versed in the language of statistical analysis - tricks of perception will often lead to misinterpretation. While we absolutely do not believe the tool can replace a doctor, we are confident that we can triage users (i.e. advise on the best next steps) consistently and effectively.

Another advantage of Bayesian Inference is that it requires a clearly defined set of data in order to reach decisions, which means these can be traced and explained. To get this data, we enlisted the help of doctors to create symptom and condition profiles based on medical evidence. This makes Healthily an augmented intelligence app rather than an artificial intelligence (AI) one.

Early experiments using AI to learn from data sets created unsafe results because the full list of symptoms were not present in the data. The AI would miss the signs of a heart attack. over a certain age with chest pain all had heart attacks and people with pain down their left arm had a muscle strain. We concluded it is much safer - as we know what the symptoms of diseases are - to tell the AI rather than let it guess.

Thanks to this approach it is always possible to replicate the results of an outcome because we know why the AI is asking the questions and what condition data sets it is attempting to narrow down its focus on. It’s very important to be able to justify why one option was chosen over another. In other words, the tool is not a ‘black box’ –we understand why decisions have been made.

Logic type 3: Machine Learning (ML)

The last piece of our logic puzzle is ML - what some describe as pure AI.

If you have a lot of unstructured ‘fuzzy’ data, ML is a very powerful tool for labelling information as it’s very good at recognising situations and patterns. Kenneth Cukier, author of Big Data, explains that ML can tell you the ‘what’ but not the ‘why’.

You may have noticed that Healthily is the only self-assessment tool that doesn’t force the user to pick a symptom from a list. Instead we have a free text box which allows you to tell our chatbot DOTTM what the problem is. That’s great from a user perspective, but it means Healthily needs to understand what you’re talking about. This can be tricky, as you might be typing in a medical problem or simply just chatting.

To tackle this, we use ML to extract concepts and understand if they are medical or not. First, we spellcheck your answers using a checker that has been enhanced by our medical team. Then we examine the colloquialisms and turn them into medical concepts, so something like ‘I pee every hour’ becomes ‘frequent urination’. We then use our own ontology system to classify medical concepts found in the input.

We also use ML to determine user intent: do they want information about a condition or do they want an assessment of their symptoms? We have to be able to understand the difference between ‘I have a headache’ and ‘can I use painkillers for my headache if I am pregnant?’ These are clearly two very different cases which must be dealt with in different ways. Once we determine intent we can use our search engine to find the best information from our medical library, or start asking questions about the user's symptoms using our Bayesian Inference engine.

Augmented intelligence - what it means in practise

We describe what we do as augmented intelligence because we have created an algorithm which is neither exclusively ML nor Bayesian, but which uses the principles of Information Theory, developed by Claude Shannon, to mathematically evaluate which questions would provide the most useful information.

If you know the children’s game “guess the animal” you will know that your brain will start with questions that give the highest amount of differentiating information: asking if the animal is a mammal is more useful and more likely to receive a positive answer than asking if the animal lives in Madagascar.

Our question selection algorithm also flags if any of the symptoms, symptom combinations or possible conditions described indicate an emergency. The emphasis on “triage” – understanding what the users best next steps could be – rather than the most probable conditions sets us apart from our competition.

It also makes us safer. In an independent study by Imperial College in association with the Royal College of GPs our advice to the user was found to be safe 99% of the time.

Healthily has been pioneering the personalisation of health information with AI since 2015. Our approach is very pragmatic: we identify a problem and then look for the best technologies to solve it.

Do your customers need help understanding their health conditions using smart AI? Get in touch and we'll arrange a call with an expert

*Your MD is the previous name for Healthily

Important: Our website provides useful information but is not a substitute for medical advice. You should always seek the advice of your doctor when making decisions about your health.

The self-care app
Check your health, manage 100s of conditions, track your progress and much more. Download today.