Israeli AI model predicts COVID-19 cases among general population using self-reported symptoms

Trained and tested on roughly 100,000 RT-PCR records collected by the Israeli Ministry of Health, the algorithm looks at eight different features including self-reported symptoms and age to make its predictions.
By Dave Muoio
11:03 am

Israeli researchers have recently detailed a new machine learning-based COVID-19 model that predicts a positive diagnosis from symptoms and – notably – is designed to be implemented among the general population.

Published in npj Digital Medicine, the tool was developed and tested on roughly 100,000 reverse transcriptase polymerase chain reaction (RT-PCR) testing records recorded from Israeli residents during the early months of the pandemic.

By collecting eight straightforward clinical signs and symptoms and running them through the model, the researchers said that their framework could help public health officials prioritize testing among the public.

"Improving clinical priorities may lower the burden currently faced by health systems, by facilitating optimized management of healthcare resources during future waves of the SARS-Cov-2 pandemic," the Tel Aviv University researchers wrote. "This is especially important in developing countries with limited resources."


The researchers wrote that their model was highly accurate, and demonstrated an area under the receiver operating characteristic curve (auROC) of 0.90 for predictions with a prospective test set (95% CI: 0.892 - 0.905). This translated to possible accuracy of 87.3% sensitivity and 71.98 specificity, or 85.76% sensitivity and 79.18% specificity. For the model's positive predictive value of a COVID-19 diagnosis against sensitivity, area under the precision-recall curve (auPRC) was 0.66 (95% CI: 0.647 - 0.678).

Among the eight symptoms and clinical signs considered by the model, fever, cough and close contact with a confirmed case were major predictors of COVID-19 contraction.

The researchers did note that their dataset included certain limitations and biases, including more comprehensive symptom reporting among positive cases than negative. To offset these and other misreporting of symptoms. Adjusting the model to include filters for these led to a minor drop in the auROC to 0.862.


The researchers built their tool to consider a handful of binary characteristics about a presenting subject: basic information regarding their sex or whether they were aged 60 years or older; self-reported symptoms including cough, fever, sore throat, shortness of breath and headache; and whether or not there was known contact with another confirmed COVID-19 case.

These features were pulled from a training-validation set of records from 51,831 RT-PCR-tested individuals (4,769 of whom were confirmed with COVID-19), and a testing set of 47,401 individuals (3,624 of whom were positive). The first set of data was came from tests conducted between March 22 and March 31, 2020, with the second from the following week.

All of the records were collected from ones publicly released by the Israeli Ministry of Health, and, alongside test dates and results, included relevant clinical signs and symptoms used for the model. All of the individuals met the ministry's indications for testing, with the exception of "a small minority who were tested under surveys of healthcare workers."


The researchers' algorithm adds to a growing library of COVID prediction models. These feature a myriad of designs. The models could incorporate individuals' symptoms, CT scans and lab tests, or may be focused on different clinical outcomes, such as hospital admission or mortality. The intended population can also vary from model to model, they continued, and several have been built using data collected from patients who have already been hospitalized.

With that being said, there have been examples of researchers using digital technologies to predict cases among individuals. The first analysis of data from the COVID Symptom Study app, downloaded by millions, included a rough predictive model to estimate the portion of users likely having COVID-19.

The Scripps Research Translational Institute's DETECT study, meanwhile, is one of several programs that combined symptoms with wearable sensor data to inform prediction algorithms.


"Based on nationwide data reported by the Israeli Ministry of Health, we developed a model for predicting COVID-19 diagnosis by asking eight basic questions. Our framework can be used, among other considerations, to prioritize testing for COVID-19 when testing resources are limited.

"In addition, the methodology presented in this study may benefit the health system response to future epidemic waves of this disease and of other respiratory viruses in general," the researchers wrote.

More regional news


(Image: "1 US Bank Note"/geralt via Pixabay, licensed under Creative Commons Zero)

Health funding

(Photo by Kwanchai Lerttanpunyapoorn/Getty Images) 



The latest news in digital health delivered daily to your inbox.

Thank you for subscribing!
Error! Something went wrong!