Few-shot meta-learning for pre-symptomatic detection of Covid-19 from limited health tracker data

Abstract

Detecting (or screening for) Covid-19 even before symptoms fully manifest, could enable patients to receive timely and life-saving treatment. Prior work has demonstrated that heart rate and step data from low-end wearables analyzed using deep learning models can detect Covid-19 reliably. However, significant individual differences in vital sign manifestation (high inter-subject variability) present a challenge to the generalization of deep learning models across diverse users. The limited amount of data in many medical scenarios further exacerbates this issue. Consequently, neural network models that can learn from limited vital sign data and varied inter-subject patterns are compelling. Meta-learning has emerged as a powerful technique for tackling various machine learning challenges, including insufficient data, domain shifts across datasets, and issues with generalization. This study proposes MetaCovid, a deep adaptation framework that employs meta-learning to address the variability of vital sign manifestation between subjects using only two days of data in order to detect Covid-19 before symptoms manifest. MetaCovid leverages heart rate and step measurements collected from consumer-grade health trackers over the preceding 2 days, extracts 45 digital bio-markers (features), which along with raw data, are fed into a deep GRU-based network with an attention mechanism, followed by uncertainty filtering. MetaCovid is trained using OC-MAML, a one-class few-shot MAML variant that adapts to the target distribution/user using only samples from the majority class. MetaCovid generalized well across two relatively small, publicly available Covid-19 datasets, achieving a recall of 0.81 and 0.92, and detecting 61% (14 out of 23) and 50% (17 out of 34) of users infected with Covid-19 before symptom onset. When OC-MAML was excluded from MetaCovid in an ablation study, the F₂ score dropped by 36%, highlighting that meta-learning indeed facilitates adaptation of deep sensing models to varying vital sign patterns. Notably, MetaCovid outperforms the current state-of-art method by predicting Covid-19 early on day $N$ using heart rate and step measurements from only the preceding 2 days compared to 28 days, reducing data requirements by 93%. To the best of our knowledge, our study is the first to propose utilizing meta-learning to mitigate vital sign variability with limited data for Covid-19 screening. We believe that MetaCovid will pave the way for innovative Covid-19 interventions that are accurate even with limited data and help contain the spread of infectious diseases in the future.