Extracting Interpretable Knowledge from the Remote Monitoring of COVID-19 Patients

Abstract

Apart from providing user-friendly applications that support digitized healthcare routines, the use of wearable devices has proven to increase the independence of patients in a healthcare setting. By applying machine learning techniques to real health-related data, important conclusions can be drawn for unsolved issues related to disease prognosis. In this paper, various machine learning techniques are examined and analyzed for the provision of personalized care to COVID-19 patients with mild symptoms based on individual characteristics and the comorbidities they have, while the connection between the stimuli and predictive results are utilized for the evaluation of the system’s transparency. The results, jointly analyzing wearable and electronic health record data for the prediction of a daily dyspnea grade and the duration of fever, are promising in terms of evaluation metrics even in a specified stratum of patients. The interpretability scheme provides useful insight concerning factors that greatly influenced the results. Moreover, it is demonstrated that the use of wearable devices for remote monitoring through cloud platforms is feasible while providing awareness of a patient’s condition, leading to the early detection of undesired changes and reduced visits for patient screening.

Keywords

• SARS-CoV-2
• modified medical research council dyspnea grade
• machine learning
• interpretability
• ensemble classifiers
• time-to-event analysis