Health Analytics on COVID-19 Data with Few-Shot Learning

Although huge volumes of valuable data can be generated and collected at a rapid velocity from a wide variety of rich data sources, their availability may vary due to various factors. For example, in the competitive business world, huge volumes of transactional shopper market data may not be made available partially due to proprietary concerns. As another example, in the healthcare domain, huge volumes of medical data may not be made available partially due to privacy concerns. Sometimes, only limited volumes of privacy-preserving data are made available for research and/or other purposes. Embedded in these data is implicit, previously unknown and useful information and knowledge. Analyzing these data by data science models can be for social and economic good. For instance, health analytics of medical data and disease reports can lead to the discovery of useful information and knowledge about diseases such as the coronavirus disease 2019 (COVID-19). This knowledge helps users to get a better understanding of the disease, and thus to take parts in preventing, controlling and/or combating the disease. However, many existing data science models often require lots of historical data for training. To deal with the challenges of limited data, we present in this paper a data science system for health analytics on COVID-19 data. With few-shot learning, our system requires only limited data for training. Evaluation on real-life COVID-19 data—specifically, routine blood test results from potential Brazilian COVID-19 patients—demonstrates the effectiveness of our system in predicting and classifying of COVID-19 cases based on a small subset of features in limited volumes of COVID-19 data for health analytics. © 2021, Springer Nature Switzerland AG.