The challenge with COVID-19 symptom assessment platforms, in general, is that they have high false positive and false negative rates. This could lead to a false sense of safety or unnecessary panic.
There is a 1 – 14 day incubation period of COVID-19, so symptoms can largely delay. There are asymptomatic (no symptoms) cases. Using the app and being told you are at low-risk even though you might have come in contact with someone who tested positive for COVID-19 is misleading and risky. Those are some of the false-negative instances.
The symptoms of COVID-19 are very much seasonal flu-like. You might have the seasonal flu or just poorly assessing your symptoms. Out of fear of becoming infected, your brain exaggerates the symptoms till they are very much real to you and you’re convinced you to have all the symptoms of COVID-19. This dynamic is called “medical students’ disease” (or nosophobia). Note, I have personally had two instances where I was very much convinced I had COVID-19 with all the symptoms, only for me to be back in full health after taking a much-needed nap (it was stress and sleep deprivation, not COVID-19 lol). A number of friends have had similar experiences too. These are some instances that could lead to false-positive risk assessment.
𝗣𝗼𝘀𝘀𝗶𝗯𝗹𝗲 𝗜𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁𝘀 (𝗥𝗲𝗰𝗼𝗺𝗺𝗲𝗻𝗱𝗮𝘁𝗶𝗼𝗻𝘀)
1. Rather than collect binary responses (‘yes’ or ‘no’), use response scales (eg. Cough: None, Barely, Mild, Moderate, Severe). This would allow you to differentiate between COVID-19 and the seasonal flu a bit better.
2. Periodically alert users to take the symptom assessment test. Rather than a flat one instance response, you can map the progression of symptoms in a user over a period of days and weeks. This should help you further differentiate between potential COVID-19 cases and the seasonal flu or nosophobia, as progressions would vastly differ.
3. Use an algorithm that provides different weights to each question and symptom. That way, symptoms, as well as a combination of symptoms, that have stronger correlations with COVID-19, can be identified. This can be based on studies done comparing COVID-19 symptoms and that of other conditions including the seasonal flu. You could instead collect that data from confirmed positive and negative cases.
4. An extension of the third point could be using that data to build a machine learning model that can identify the right weights to assign each symptom and question. This would be much helpful if there’s a complex multidimensional pattern.
Bonus: Don’t just ask about COVID-19 symptoms, ask questions and collect symptoms that are strongly unrelated to COVID-19. It could help in isolating true potential cases, though it might also be possible the person simply has COVID-19 and another condition.
PS: The challenge however with points 3 and 4, if the data on symptoms and questions collected from confirmed positive and negative cases is noisy, which I suspect it’d be, you can’t trust any model or algorithm you develop out of it.
Author – Darlington Akogo