A recent UK report highlighted the fact that measurements of blood oxygenation using pulse oximeters may be inaccurate for patients with darker skin tones. The photo-plethysmography signals collected by pulse oximeters and a multitude of wearable devices are often combined with machine learning to derive physiological parameters that may be useful for health monitoring. So it is natural to question whether machine learning accuracy when using photoplethysmography signals as input can also lead to inaccurate results for darker skin tones, particularly when a model may be trained on data from subjects with predominantly lighter skin tones.

To test this, we chose a binary classification problem of classifying systolic blood pressure as high (≥ 140 mm Hg) or not high (< 140 mm Hg). We used the Aurora BP dataset which includes photoplethysmography signals, blood pressure readings and Fitzpatrick skin tone classification. We trained a model using only skin tone class 1 data (lightest skin tone) and then tested all the other skin tone data using this model. We found that the accuracy dropped for all the other skin tone classes. We then trained a model using data from all skin tones and the accuracy for the darker skin tones improved except class 5.