Machine learning nominal max oxygen consumption from wearable reflective pulse oximetry with density functional theory

Saleem Abdul Fattah Ahmed Al Dajani, Frédéric Laquai

Keywords: Precision medicine, artificial intelligence, wearable sensors, hemoglobin, max oxygen consumption rate, pulsatile blood flow transcutaneous reflectance oximetry, density functional theory, random forest regression, machine learning

TL;DR: We present a novel method for AI-driven precision medicine to infer max oxygen consumption rate based on random forest regression of wearable sensor physiological measurements using inputs from human hemoglobin density functional theory models.

Abstract: Wearable sensors have revolutionized modern health towards precision medicine. Many measurements taken by wearables are not currently used for medical purposes, and their physical roots are not understood. Here we show the origin of inferring oxygen saturation from measuring estimated oxygen variation while preparing for a marathon via spectrophotometric pulsatile blood flow transcutaneous reflectance oximetry, building six years of data tracking the stages of running a marathon. Based on inputs from quantum mechanical simulations into classical wave and electromagnetic theory models, the imaginary part of the complex dielectric function of the active site of human hemoglobin is altered upon oxygen binding to the heme group, changing the extinction coefficient and selectively absorbing wavelengths of light in a way that enables its detection from the ratio of red to infrared absorbance. A fundamental nominal max oxygen consumption is inferred from quantum mechanical absorbance spectra with and without oxygen binding by training a machine learning model on 1.5 years of daily wearable reflective pulse oximetry data with an $R^2$ of 0.84. Reflectance oximetry is strongly correlated ($R^2$ of 0.96) to standard pulse oximetry, such as through earlobe or fingertip, enabling accurate, rapid, regular, and automated monitoring of vital signs with wearable sensors on smart watches and fitness trackers, and supplying artificial intelligence inferences of function-symptom relationships for precision medicine.

Submission Number: 16