A Deep Learning Approach for Estimating SpO2 Using a Smartphone Camera

The COVID-19 pandemic has highlighted the significance of both telemedicine and measuring physiological signals as peripheral oxygen saturation (SpO(2)) at home. Aiming to increase the confiability of SpO(2) readings performed by smartphones, this study presents a novel deep neural network (DNN) model that predicts SpO(2) levels using fingertip videos captured by smartphones. The model was trained and evaluated using the leave-one-out method and compared against the SpO(2) values measured by a set of oximeters. The dataset used in this work is a differential of the study, unlike most similar studies which use a breath-holding protocol that presents only a few samples of low SpO(2) values; our tests were evaluated using a publicly available dataset in which volunteers ' SpO(2) levels were measured in a range from almost 100% to approximately 70% by manipulating the percentage of oxygen inhaled by the volunteers, and the SpO(2) levels were kept on lower values for a considerable period of time. The proposed DNN model was compared against multiple variations in the ratio-of-rations (RoR) method and the model proposed by the dataset creators. The results show that the DNN outputs have a strong relationship with the true SpO(2) measurements, as indicated by a Pearson coefficient (PC) of 0.95 and a concordance correlation coefficient (CCC) of 0.95, and also a small error, measured by a root mean squared error (RMSE) of 2.92% and a mean absolute error (MAE) of 1.97%. Compared with the RoR method, which achieved a PC of 0.29, CCC of 0.21, RMSE of 8.83% and MAE of 7.03%, the proposed DNN model exhibits superior performance. Moreover, the proposed DNN model outperforms the model proposed by the dataset authors, which achieved an MAE of 5.00%. These findings demonstrate that the proposed DNN model can accurately predict SpO(2) levels using smartphone-recorded fingertip videos with lower error and higher precision than a widely used classical method.