A paper titled “Learning from Large-scale Wearable Device Data for Predicting Epidemics Trend of COVID-19“ was published on Discrete Dynamics in Nature and Society (DDNS) on May 5. In their paper, 12 researchers from world-leading wearable device maker Huami (NYSE:HMI) developed a framework to predict COVID-19 infection trends using data from wearable devices. The experiment on which the paper is based revealed that Huami’s prediction model can alert public officials in the early stages of an outbreak of COVID-19.
Using big data for epidemic forecasting is not a new idea. Though useful, many models use official epidemic statistics that are reported after an epidemic has passed and are therefore not applicable in real time. Real-time big data has been used to try to improve the timeliness of predictive models. Google launched its Flu Trends (GFT) model in 2008, which used people’s web searches about flu symptoms to predict outbreaks earlier than the Centers for Disease Control (CDC). However, GFT failed to achieve its goal because the search queries were affected by social hotspots and driven by users’ self-diagnoses. Indeed many people who have flu-like symptoms get tested and find they actually have other illnesses. The percentage of positive flu tests ranges from less than 1% to about 30%, demonstrating how difficult it is for people to self-diagnose. According to Time Magazine, GFT overestimated the prevalence of flu by 50% due to the fact that user data does not always reflect real trends.
In the study, Huami defined anomalous physiological signs using two key indicators: resting heart rate (RHR) and sleep duration, changes to which can suggest influenza-like illnesses. Each user’s data was compared with the average data, and anomalous readings were determined using standard deviations. Furthermore, to distinguish COVID-19 from other influenza-like illnesses, a heterogeneous neural network regression model was built and trained with data inputs including holiday activity, season, weather, historical physiological anomaly rate, active user density, and the official reported COVID-19 infection rate. The online learning model is constantly updated and trained with newly available official data which has allowed it to better predict the outbreak’s trends.
The paper highlighted findings in China and Southern Europe, including in Spain and Italy. Wuhan’s data showed the predicted physiological anomaly rate from official data inputs in 2020 fits well with the rate calculated by the anomaly detection algorithm. Researchers also analyzed data from 4 other cities in China and compared them with Wuhan. They found that each of the cities reached an outbreak peak which may correspond to the official confirmed numbers. Taking Wuhan as an example, the results also showed the proposed model could predict the outbreak of COVID-19 as early as 10 days in advance. On the contrary, the lab test and diagnosis results are often delayed because not all patients have access to professional opinions as soon as they develop symptoms.
Similar data trends were observed in Europe including Italy and Spain. The predicted infection rate in both countries coincided with the official outbreak numbers. The model found both countries’ outbreaks peaked at least one week ahead of the officially-reported peak. If the prediction model could be trained and applied appropriately, it could provide public health officials valuable lead time to help them better prepare for the next outbreak.
The study revealed promising results for using data from wearable devices and machine learning to predict COVID-19 outbreaks in advance. But the researchers also mentioned some limitations of their data, such as potential issues using information from holidaymakers, as drinking may lead to elevated heart rates which could skew the data. Furthermore, using large data inputs is important for ensuring the validity of results. If a particular age group’s wearable device user base is not large enough, it will be difficult to draw solid conclusions. However, the researchers believe that early warnings are useful for both individuals and healthcare professionals, and they aim to incorporate more data points such as BMI and gender into future research.
The wearable device market is expanding. According to Technavio, the wearable technology market will grow by USD 35.48 billion between 2020 and 2024, progressing at a CAGR (compound annual growth rate) of 13% over the forecast period. The more devices sold, the more data will be collected. On May 12, Huami announced its unaudited 2020 Q1 financial results which showed a 36.1% revenue increase compared with the first quarter of 2019. During the first three months of 2020, Huami shipped 7.6 million wearable devices, representing 35.7% growth from its 2019 Q1 figures. The report also showed the company spent more on R&D, as innovation is the company’s cornerstone.
On April 9, Huami announced a collaboration with Dr. Zhong Nanshan’s team to use data from wearable devices to monitor COVID-19 patients after they’re discharged from hospitals in China. The project will give researchers real-time data to identify abnormal signs that can predict lingering illness. With the development of algorithms and sensor technology, future wearable devices will enable researchers to explore new ways to use data to serve both users and society. When we discuss how big data can help us in the future, one must mention Huami’s contributions to building a healthcare ecosystem with its innovative technology and health data.
(Top photo from Pixabay)