Machine learning-based optimization of pre-symptomatic COVID-19 detection through smartwatch

Abstract

Patients with weak or no symptoms accelerate the spread of COVID-19 through various mutations and require more aggressive and active means of validating the COVID-19 infection. More than 30% of patients are reported as asymptomatic infection after the delta mutation spread in Korea. It means that there is a need for a means to more actively and accurately validate the infection of the epidemic via pre-symptomatic detection, besides confirming the infection via the symptoms. Mishara et al. (Nat Biomed Eng 4, 1208-1220, 2020) reported that physiological data collected from smartwatches could be an indicator to suspect COVID-19 infection. It shows that it is possible to identify an abnormal state suspected of COVID-19 by applying an anomaly detection method for the smartwatch's physiological data and identifying the subject's abnormal state to be observed. This paper proposes to apply the One Class-Support Vector Machine (OC-SVM) for pre-symptomatic COVID-19 detection. We show that OC-SVM can provide better performance than the Mahalanobis distance-based method used by Mishara et al. (Nat Biomed Eng 4, 1208-1220, 2020) in three aspects: earlier (23.5-40% earlier) and more detection (13.2-19.1% relative better) and fewer false positives. As a result, we could conclude that OC-SVM using Resting Heart Rate (RHR) with 350 and 300 moving average size is the most recommended technique for COVID-19 pre-symptomatic detection based on physiological data from the smartwatch.

Figure 1

(a) A confusion matrix of the relative accuracy rates of the anomaly detection techniques. (b) The average the relative accuracy rate of each technique w.r.t the other techniques.

Figure 2

Variations of anomaly detection performances by moving averages.

Figure 3

(a) A confusion matrix of the relative accuracy rate of anomaly detection for each combination of models, methods, and the moving average sizes. (b) The average of the relative accuracy rate of each technique w.r.t. the other techniques.

Figure 4

(a) The total number of outliers produced by each techniques (b) The minimum training periods of each technique.

Figure 5

The total number of outliers from COVID-19 negative participants.

Figure 6

An example of anomaly detection using MD and OC-SVM (Wine recognition).