Deep Learning for Sleep Heart Rate Estimation from Accelerometers: Toward Population-Scale Cardiac Insight Without Optical Sensors

Problem

This work addresses the challenge of recovering cardiac information from wrist accelerometry during sleep without relying on optical heart-rate sensing. The authors highlight the need for a method that can provide accurate heart rate estimates using non-invasive accelerometer data, which is particularly relevant for large-scale population studies where optical sensors may not be feasible.

Method

The authors propose SeqSmoother, a transformer-based architecture designed as a temporal corrector for heart rate estimation. The model integrates several components: spectral descriptors, a frequency anchor derived from Nightbeat, and physics-motivated sub-harmonic features. The training process utilizes wrist accelerometry data, with ECG data serving as a reference for heart rate labels and to assign quality weights to the training labels. Specific details regarding the training compute resources used are not disclosed.

Results

The proposed method achieves a participant-macro Mean Absolute Error (MAE) of 1.60 beats per minute (bpm), which is significantly higher than the baseline performance of Nightbeat, which reports an MAE of 0.615 bpm. Additionally, the final estimates from Nightbeat cover 72.85% of the SeqSmoother-eligible out-of-fold grid. The model also demonstrates a sub-harmonic ratio Area Under the Receiver Operating Characteristic (AUROC) of 0.972 for identifying candidates for harmonic lock-on, indicating strong performance in this aspect.

Limitations

The authors note a trade-off between accuracy and availability, specifically regarding the balance between learned temporal modeling and quality-gated signal processing. This suggests that while the model may improve accuracy, it could also complicate the availability of high-quality signals necessary for optimal performance. Other limitations are not explicitly discussed in the text.

Why it matters

The implications of this research are significant for advancing non-invasive cardiac monitoring technologies. By enabling accurate heart rate estimation from wrist accelerometry, this work opens avenues for large-scale health monitoring applications, particularly in populations where traditional optical sensors are impractical. The findings could lead to improved health insights and monitoring capabilities in various settings, including clinical and remote health environments.