Machine learning COVID-19 detection from wearables

Nestor, Bret; Hunter, Jaryd; Kainkaryam, Raghu; Drysdale, Erik; Inglis, Jeffrey B.; Shapiro, Allison; Nagaraj, Sujay; Ghassemi, Marzyeh; Foschini, Luca; Goldenberg, Anna

doi:10.1016/s2589-7500(23)00045-6

Cited by 10 publications

(7 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Furthermore, certain binary classification performance metrics (i.e., AUROC, accuracy) can lead to misleading notions of performance when used on datasets that exhibit extreme class imbalance, as in these analyses where the number of non-fever days far outnumber fever days. Such a class imbalance is common in illness detection studies [ 27 ]. Accordingly, we attempted to report all metrics in a way that did not overestimate the performance.…”

Section: Discussionmentioning

confidence: 99%

“…This work also differs from other illness detection studies in both study design and the wearable device used to gather data. We performed these analyses retrospectively, and the performance should be verified in a prospective manner [ 27 ]. Furthermore, differences in commercially available wearable device sensors (i.e., the ability to collect HRV, HR, temperature, and other physiological metrics) have led to substantial differences in the features used to train illness detection classifiers.…”

Section: Discussionmentioning

confidence: 99%

“…In order to ensure applicability, we implemented a relatively simple, commonly used ensemble classifier based on the standard implementation of a Histogram-Based Gradient-Boosting Classification Tree from the sklearn Python (Open source) package v1.2.0 (sklearn.ensemble.HistGradientBoostingClassifier) with all hyper-parameters left at default. Models of this variety are commonly used for physiological anomaly detection [ 8 , 10 , 27 , 28 ]. For training and testing, we followed a five-fold stratified cross-validation scheme with a user split as previously outlined in Merill et al [ 15 ], where each model was trained on data from a subset of participants and tested on another subset.…”

Section: Methodsmentioning

confidence: 99%

Section: Onset Latency Secondsmentioning

confidence: 99%

See 3 more Smart Citations

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

Kasl,

Keeler Bruce,

Hartogensis

et al. 2024

Sensors

View full text Add to dashboard Cite

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.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Discussionmentioning

confidence: 99%

Section: Methodsmentioning

confidence: 99%

Section: Onset Latency Secondsmentioning

confidence: 99%

See 2 more Smart Citations

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

Kasl,

Keeler Bruce,

Hartogensis

et al. 2024

Sensors

View full text Add to dashboard Cite

show abstract

“…In addition, although not discussed in the present work, tracking of cardiovascular metrics and movement patterns during sleep could also pave the way for integrated arrhythmia detection and other cardiovascular events [ 21–23 ], influenza/coronavirus disease monitoring [ 24 ], as well as, fall detection [ 25 ] in these vulnerable groups that could be shared in real-time with a caregiver or clinical provider. As aging in place becomes more common, one needs to carefully evaluate cost–benefit trade-offs associated with a particular use case of CTs.…”

mentioning

confidence: 99%