Key result
Reflective PPG linear regression linked to ~25% lower RMSE for SpO2 prediction.
Why the study?
Reflective PPG sensors exhibit patient-dependent variations in the Ratio of Modulation, causing conventional SpO2 calibration curves to report inaccurate values.
Do Machine Learning algorithms improve the accuracy of SpO2 determination using reflective PPG compared to Blood Gas Analysis in healthy subjects?
Observational (n=10)
Do Machine Learning algorithms improve the accuracy of SpO2 determination using reflective PPG compared to Blood Gas Analysis in healthy subjects?
Machine learning, specifically linear regression using combined features, can accurately determine SpO2 from reflective PPG sensors independently of the subject.
Answers are grounded in this paper and cite their sources.
Supports ML refinement of reflective PPG SpO2 estimation; leaves open clinical utility pending larger patient studies.
Reflective Photoplethysmography (PPG) sensors are less obtrusive than transmissive sensors, but they present patient-dependent variations in the so-called “Ratio of Modulation” (R). Thus, the conventionally employed calibration curves for determining peripheral oxygen saturation ( S p O 2 ) may report inaccurate values. In this paper, we study the possibility of overcoming these limitations through Machine Learning (ML). For that, we show the results of applying several algorithms and feature combinations to PPG data from a human hypoxia study. The study was performed on ten healthy subjects. Their target oxygen saturation was reduced in five steps from 98- 100% to 70-77% through an oral mask. Blood Gas Analysis (BGA) was performed five times for each saturation level to measure the arterial oxygen saturation. PPG data were acquired from a reflective pulse oximeter placed in the subjects’ ear canals. PPG signals were pre-processed, and several features in the frequency and temporal domain were calculated. For the ML algorithms’ input, we explored different combinations of the features. We trained and validated the algorithms with the data from seven patients, and we tested them on three. Finally, we performed leaveone- out cross-validation to ensure the universality of the methods. The results show a good agreement of the predictions with the BGA values for Linear Regression, k- Nearest Neighbors, Stochastic Gradient Descent, and Neural Network for all input feature combinations with an average RMSE in the range of 3%. However, the performance of the Linear Regression was not beaten by the Neural Network, even for overfitting with 2000 hidden layers. The combination of R calculated with a Fast-Fourier Transform and ACRMS .red /ACRMS .ir significantly improved the results, reducing the RMSE by 25%. This work demonstrates that a straight-forward Linear Regression is capable of determining S p O 2 with reflective PPG independently of the subject if the ratio ACRMS .red /ACRMS .ir is considered simultaneously with the Ratio of Modulation.
No takes yet. Share an insight, caveat, or question.
Badiola et al. (2021) conducted an observational in Hypoxia (n=10). Machine Learning algorithms on reflective PPG vs. Blood Gas Analysis (BGA) was evaluated on Agreement of SpO2 predictions with Blood Gas Analysis (RMSE) (25% reduction in RMSE). Linear Regression using reflective PPG and the ratio ACRMS.red/ACRMS.ir alongside the Ratio of Modulation determined SpO2 with an average RMSE of ~3%, reducing the RMSE by 25%.
Synapse has enriched 10 related papers. Consider them for comparative context: