The prevelence of obstructive sleep apnea (OSA) has significantly increased around the world. The word “apnea” is equivalent to “no breath,” which is the main symptom of this condition. The challenge of detecting sleep disorders from consumer wearable sensors is attracting more and more researchers in the field. Some of them adopted deep learning approaches, while others combined feature engineering with conventional machine learning techniques. A study used at-home oximetry data to estimate the Apnea/Hypopnea Index (AHI) and achieved high intra-class correlation coefficients within 0.889-0.924 [1]. They then classified apnea severity using three typical AHI thresholds: 5 events/hour, 15 events/hour, and 30 events/hour, respectively. Another study using a CNN-based model to classify OSA severity and achieved a macro-accuracy of 80.51 %, a mean sensitivity of 79.95%, a mean specificity of 93.05% [2]. However, compared to clinical equipment, data obtained from wearable devices are low in frequency and resolution, making it challenging to capture all vital information for diagnosing OSA. The development and evaluation of a single model with the ability to reach high diagnostic performance using consumer trackers are still pending. In this study, we aimed to classify the apnea severity from consumer-grade low-granularity oximetry data.