Fitbeat: COVID-19 estimation based on wristband heart rate using a contrastive convolutional auto-encoder

Abstract

This study proposes a contrastive convolutional auto-encoder (contrastive CAE), a combined architecture of an auto-encoder and contrastive loss, to identify individuals with suspected COVID-19 infection using heart-rate data from participants with multiple sclerosis (MS) in the ongoing RADAR-CNS mHealth research project. Heart-rate data was remotely collected using a Fitbit wristband. COVID-19 infection was either confirmed through a positive swab test, or inferred through a self-reported set of recognised symptoms of the virus. The contrastive CAE outperforms a conventional convolutional neural network (CNN), a long short-term memory (LSTM) model, and a convolutional auto-encoder without contrastive loss (CAE). On a test set of 19 participants with MS with reported symptoms of COVID-19, each one paired with a participant with MS with no COVID-19 symptoms, the contrastive CAE achieves an unweighted average recall of $95.3\%$ , a sensitivity of $100\%$ and a specificity of $90.6\%$ , an area under the receiver operating characteristic curve (AUC-ROC) of 0.944, indicating a maximum successful detection of symptoms in the given heart rate measurement period, whilst at the same time keeping a low false alarm rate.

Keywords: Anomaly detection; COVID-19; Contrastive learning; Convolutional auto-encoder; Respiratory tract infection.

Fig. 1

Segmentation and pre-processing of heart rate data of a participant with reported COVID-19-like symptoms. Top: Heart rate data recorded 24-hours-a-day/7-days-a-week from 21 February to 20 May 2020 (total 90 days). Onset (black vertical bar) indicates 0 o’clock at the reported symptom onset date. Red rectangle – 7 days heart rate data before and after symptom onset representing a symptomatic segment; green rectangle – asymptomatic segment. Middle: Symptomatic segment. Blue curve – unprocessed heart rate trajectory of the red rectangle above; red curve – heart rate trajectory averaged over 5-minutes intervals. Bottom: Representation of the symptomatic segment as $24\times 168$ sized image of 5-minutes heart rate data related pixels. Each column represents an interval of 2 h, the 168 columns sum up to 14 days.

Fig. 2

The convolutional auto-encoder (CAE) architecture with 4 encoder layers and 4 decoder layers as an example. An encoder layer is a sequence of convolution – batch-normalisation – PReLU – max-pooling. A decoder layer is a sequence of transposed convolution – batch-normalisation – PReLU – transposed max-pooling. The distance between the original and reconstructed image represents the reconstruction error.

Fig. 3

Training and testing curves illustrated by the reconstruction errors when using different margin sizes.

Fig. 4

Reconstruction errors for continuous binary COVID-19 yes/no classification on 14-days heart rate windows of an exemplary individual (the same as in Fig. 1, top).