Combining self-reported symptoms with Artificial Intelligence can predict the early symptoms of Covid-19, according to research led by scientists at Kings College London. Previous studies have predicted whether people will develop Covid using symptoms from the peak of viral infection, which can be less relevant over time — fever is common during later phases, for instance.
The new study reveals which symptoms of infection can be used for early detection of the disease. Published in the journal The Lancet Digital Health, the research used data collected via the ZOE COVID Symptom Study smartphone app. Each app user logged any symptoms that they experienced over the first 3 days, plus the result of a subsequent PCR test for Coronavirus and personal information like age and sex.
Researchers used those self-reported data from the app to assess three models for predicting Covid in advance, which involved using one dataset to train a given model before its performance was tested on another set. The training set included almost 183,000 people who reported symptoms from 16 October to 30 November 2020, while the test dataset consisted of more than 15,000 participants with data between 16 October and 30 November.
The three models were: 1) a statistical method called logical regression; 2) a National Health Service (NHS) algorithm, and; 3) an Artificial Intelligence (AI) approach known as a ‘hierarchical Gaussian process’. Of the three prediction models, the AI approach performed the best, so it was then used to identify patterns in the data. The AI prediction model was sensitive enough to find which symptoms were most relevant in various groups of people.
The subgroups were occupation (healthcare professional versus non-healthcare), age group (16-39, 40-59, 60-79, 80+ years old), sex (male or female), Body-Mass Index (BMI as underweight, normal, overweight/obese) and several well-known health conditions. According to results produced by the AI model, loss of smell was the most relevant early symptom among both healthcare and non-healthcare workers, and the two groups also reported chest pain and a persistent cough.
The symptoms varied among age groups: loss of smell had less relevance to people over 60 years old, for instance, and seemed irrelevant to those over 80 — highlighting age as a key factor in early Covid detection. There was no big difference between sexes for their reported symptoms, but shortness of breath, fatigue and chills/shivers were more relevant signs for men than for women.
No particular patterns were found in BMI subgroups either and, in terms of health conditions, heart disease was most relevant for predicting Covid. As the study’s symptoms were from 2020, its results might only apply to the original strain of the SARS-CoV-2 virus and Alpha variant – the two variants with highest prevalence in the UK that year.
The predictions wouldn’t have been possible without the self-reported data from the ZOE COVID Symptom Study project, a non-profit collaboration between scientists and personalized health company ZOE, which was co-founded by genetic epidemiologist Tim Spector of Kings College London.
The project’s website keeps an up-to-date ranking of the top 5 Covid symptoms reported by British people who are now fully vaccinated (with a Pfizer or AstraZeneca vaccine), have so far received one of the two doses, or are still unvaccinated. Those top 5 symptoms provide a useful resource if you want to know which signs are common for the most prevalent variant circulating in a population — currently Delta – as distinct variants can be associated with different symptoms.
When a new variant emerges in future, you could pass some personal information (such as age) to the AI prediction model so it shows the early symptoms most relevant to you — and, if you developed those symptoms, take a Covid test and perhaps self-isolate before you transmit the virus to other people. As the new study concludes, such steps would help alleviate stress on public health services:
“Early detection of SARS-CoV-2-infected individuals is crucial to contain the spread of the COVID-19 pandemic and efficiently allocate medical resources.” Follow me on Twitter or LinkedIn. Check out my website or some of my other work here.
I’m a science communicator and award-winning journalist with a PhD in evolutionary biology. I specialize in explaining scientific concepts that appear in popular culture and mainly write about health, nature and technology. I spent several years at BBC Science Focus magazine, running the features section and writing about everything from gay genes and internet memes to the science of death and origin of life. I’ve also contributed to Scientific American and Men’s Health. My latest book is ’50 Biology Ideas You Really Need to Know’.
Healthcare providers and researchers are faced with an exponentially increasing volume of information about COVID-19, which makes it difficult to derive insights that can inform treatment. In response, AWS launched CORD-19 Search, a new search website powered by machine learning, that can help researchers quickly and easily search for research papers and documents and answer questions like “When is the salivary viral load highest for COVID-19?”
Built on the Allen Institute for AI’s CORD-19 open research dataset of more than 128,000 research papers and other materials, this machine learning solution can extract relevant medical information from unstructured text and delivers robust natural-language query capabilities, helping to accelerate the pace of discovery.
In the field of medical imaging, meanwhile, researchers are using machine learning to help recognize patterns in images, enhancing the ability of radiologists to indicate the probability of disease and diagnose it earlier.
UC San Diego Health has engineered a new method to diagnose pneumonia earlier, a condition associated with severe COVID-19. This early detection helps doctors quickly triage patients to the appropriate level of care even before a COVID-19 diagnosis is confirmed. Trained with 22,000 notations by human radiologists, the machine learning algorithm overlays x-rays with colour-coded maps that indicate pneumonia probability. With credits donated from the AWS Diagnostic Development Initiative, these methods have now been deployed to every chest x-ray and CT scan throughout UC San Diego Health in a clinical research study.