Utilizing Wearable Device Data for Syndromic Surveillance: A Fever Detection Approach

Abstract

Commercially available wearable devices (wearables) show promise for continuous physiological monitoring. Previous works have demonstrated that wearables can be used to detect the onset of acute infectious diseases, particularly those characterized by fever. We aimed to evaluate whether these devices could be used for the more general task of syndromic surveillance. We obtained wearable device data (Oura Ring) from 63,153 participants. We constructed a dataset using participants' wearable device data and participants' responses to daily online questionnaires. We included days from the participants if they (1) completed the questionnaire, (2) reported not experiencing fever and reported a self-collected body temperature below 38 °C (negative class), or reported experiencing fever and reported a self-collected body temperature at or above 38 °C (positive class), and (3) wore the wearable device the nights before and after that day. We used wearable device data (i.e., skin temperature, heart rate, and sleep) from the nights before and after participants' fever day to train a tree-based classifier to detect self-reported fevers. We evaluated the performance of our model using a five-fold cross-validation scheme. Sixteen thousand, seven hundred, and ninety-four participants provided at least one valid ground truth day; there were a total of 724 fever days (positive class examples) from 463 participants and 342,430 non-fever days (negative class examples) from 16,687 participants. Our model exhibited an area under the receiver operating characteristic curve (AUROC) of 0.85 and an average precision (AP) of 0.25. At a sensitivity of 0.50, our calibrated model had a false positive rate of 0.8%. Our results suggest that it might be possible to leverage data from these devices at a public health level for live fever surveillance. Implementing these models could increase our ability to detect disease prevalence and spread in real-time during infectious disease outbreaks.

Keywords: illness detection; syndromic surveillance; wearables.

Figure 1

Instance selection and normalization procedure. At least 7 out of the 14 days in the range of −28 to −14 relative to the ground truth day were retrievable. The mean (μ) and standard deviation (σ) from these days were used to normalize z-score wearable device metrics. We depict an example of a valid instance with its baseline period (−28 → −14) with retrievable data from 9 out of 14 nights (nights without retrievable data are indicated by a white cross). This instance is based on sleep summary features from the night before (night −1) and the night after (night 0) relative to the ground truth day.

Figure 2

Self-reported body temperatures from non-fever examples are in blue and fever examples are in orange.

Figure 3

Z-score-normalized wearable metrics from individuals, aligned by self-reported fever day (white hatched areas) and grouped by self-reported temperature on fever day. Individuals reporting temperatures in the range of (38–39 °C) are in blue (n = 621), and (39+ °C) are in red (n = 103). Lines represent the mean z-score normalized wearable metric across all participants in the respective group for each night, and shaded regions are the 95% confidence interval of the mean.

Figure 4

Performance of the fever detection classifier following a five-fold cross-validation scheme. Shaded areas indicate a 95% confidence interval. (a) The mean Receiver Operator Characteristic curve (ROC) across iterations. The mean area under the curve is 0.85. (b) The mean Precision–Recall curve (PRC) across iterations. The average precision was 0.25. (c) The reliability plot (or calibration curve) across iterations. The mean Brier score was 0.0018. (d) Box plots indicating the classifier predicted probability, binned by self-reported body temperature.

Figure 5

Explanation of the fever detection classifier. Features are ranked from most (top) to least (bottom) important based on the mean permuted importance across 30 permutations. NB: Night before [non]-fever day; NA: night after [non]-fever day; days of the week (i.e., Sunday) indicate the ground truth day; error bars: 95% confidence interval of the mean.