LEARNING BLOOD OXYGEN FROM RESPIRATION SIG

Abstract

Monitoring blood oxygen is critical in a variety of medical conditions. For almost a century, pulse oximetry has been the only non-invasive method for measuring blood oxygen. While highly useful, pulse oximetry has important limitations. It requires wearable sensors, which can be cumbersome for older patients. It is also known to be biased when used for darkskinned subjects. In this paper, we demonstrate, for the first time, the feasibility of predicting oxygen saturation from breathing. By eliminating the dependency on oximetry, we eliminate bias against skin color. Further, since breathing can be monitored without body contact by analyzing the radio signal in the environment, we show that oxygen too can be monitored without any wearable devices. We introduce a new approach for leveraging auxiliary variables via a switcher-based multi-headed neural network model. Empirical results show that our model achieves good accuracy on multiple medical datasets.

1. INTRODUCTION

Oxygen saturation refers to the amount of oxygen in the blood -that is the fraction of oxygensaturated hemoglobin relative to the total blood hemoglobin. Normal oxygen levels range from 94% to 100%. Lower oxygen can be dangerous, and if severe, leads to brain and lung failure (Taylor et al., 1951; Díaz-Regañón et al., 2002; Lapinsky et al., 1999) . Oxygen Monitoring is important for patients suffering from lung diseases such as COVID-19, pulmonary embolism (PE), and chronic obstructive pulmonary disease (COPD) who are susceptible to reduced blood oxygen (NIH, 2020; Nordenholz et al., 2011) . It is also recommended in old people since lung functions deteriorate with age putting the elders at a high risk of low oxygen (NLM, 2020). Today, the only non-invasive approach for measuring oxygen saturation (SpO2) is pulse oximetry. It works by shining light on one's figure (or other body parts). Since oxygen-saturated hemoglobin has a bright red color, oxygen saturation can be measured by estimating the absorbance of the redcolored waves relative to other colors. While pulse oximetry is a very useful technology, it can be unsuitable for some of the most vulnerable patients. Oximetry requires the patient to wear a finger clip or other sensors, which can be cumbersome and hard to remember for older patients with dementia or cognition problems. Further, since it relies on measuring light absorbance, it is affected by skin color and tend to overestimate blood oxygen in dark-skinned subjects (Feiner et al., 2007) . In this paper, we introduce a new approach for monitoring blood oxygen that could be more appropriate for older patients or those with dark skin color. We propose to learn oxygen saturation from respiration signals. Recent advances in wireless sensing have demonstrated the feasibility of obtaining accurate measurements of the respiration signal from analyzing the radio waves that bounce off people's bodies, without any physical contact (Adib et al., 2015; Yue et al., 2018) . Hence, if oxygen saturation is predictable from breathing, we can sense one's oxygen in a contactless way without any wearable sensors. Further, unlike pulse oximetry, such oxygen sensing method would be oblivious to skin color. We model the problem as sequence-to-sequence learning and introduce a neural network that maps respiration signals to the corresponding oxygen saturation. A key innovation in our model is a new design for leveraging auxiliary variables. The standard approach in the literature for leveraging auxiliary variables would either provide those variables to the model as inputs or try to predict them as auxiliary tasks. We demonstrate empirically that for our problem neither of these approaches is beneficial. We show that in some cases, the gradient of the model can take vastly different and even divergent values given the auxiliary variable. In such a scenario, it is better to learn different models for different values of the variable. This motivates us to propose an alternative design for incorporating auxiliary information. We introduce a multi-headed model where the value of the auxiliary variable is used to gate the learning through the appropriate head. We evaluate our model on multiple medical datasets, and compare it with various baselines. Our empirical results show that the average absolute error in predicting oxygen saturation on the tested medical data is 1.6%, which is comparable to the accuracy of consumer pulse oximeters (whose average error ranges from 0.4% to 3.5% (Lipnick et al., 2016) ). Further, by testing the model on data that include the RF modality, we show that the model generalizes to breathing signals extracted from radio waves, opening the door for continuous contactless monitoring of blood oxygen. We also use longitudinal data from one COVID-19 patient as a qualitative case study, showing that the predicted oxygen values are consistent with the patient's recovery process. In summary, this paper makes the following contributions: • The paper is the first to predict blood oxygen from respiration signals. This provides a novel non-invasive way for monitoring oxygen saturation, and improves our understanding of the relationship between these two physiological signals. 

2. RELATED WORK

Predicting Oxygen Saturation. Past work on oxygen prediction focuses on inferring future oxygen values from recent measurements. Several papers use auto-regressive models that take past SpO2 readings and predict the SpO2 in the near future (ElMoaqet et al., 2013; 2014; 2016) . Others use an auto-regressive model to model the photoplethysmogram (PPG) signal, which is the raw data obtained from a pulse oximeter (Lee et al., 2011) . Some past work predicts changes in SpO2 in the five minutes after adjusting ventilator setting (Ghazal et al., 2019) . Our work differs from these methods in that it predicts oxygen saturation from breathing, a completely different physiological signal. Further, we show the feasibility of oxygen prediction without body contact by inferring oxygen from radio-based breathing estimates. Contactless Sensing with Radio Frequency Signals. The past decade has seen a rapid growth in research on passive sensing using radio frequency (RF) signals. Early work has demonstrated the possibility of sensing one's respiration and heart rate using radio signals (Adib et al., 2015) . Building on this work, researchers have found that by carefully analyzing the RF signals that bounce off the human body, they can monitor a variety of health metrics including sleep, respiration, heart rate, gait, falls, and even human emotions (Nguyen et al., 2016; Wang et al., 2017; Yue et al., 2018; Hsu et al., 2017; Wang et al., 2016; Tian et al., 2018; Zhao et al., 2016; Jiang et al., 2018) & Weston, 2008; Dong et al., 2015) , reinforcement learning (Wilson et al., 2007), and visual odometry (Valada et al., 2018) . Our work explores a third way of using auxiliary variables. We focus on partitioning the data space by those auxiliary variables based on analyzing how they affect the gradient of the primary task. We show that a wise partition facilitates model learning and leads to better results.



• The paper introduces a new approach for leveraging auxiliary variables. We analyze the gradient of a vanilla model with respect to the variable; if the gradient is divergent or vastly different for different values of the variable, we use a multi-headed neural network gated by the variable.• The paper is the first to demonstrate the feasibility of contactless oxygen prediction from radio signals. Such a design can facilitate in-home oxygen monitoring, particularly for older patients who may have problems remembering to wear and charge oxygen sensors.

. Our work adds a new piece to the global picture of predicting vital signals from radio waves. ) provides an alternative approach for leveraging auxiliary variables by taking them as secondary supervisors. It has been shown beneficial in multiple domains including object detection (Mordan et al., 2018), sequence modelling (Collobert

