The Resolution Hypothesis: Revealing the Correspondence Between Time Scales and Health Outcomes

PPG is a multi-layered signal that captures heart rate, vascular tone, autonomic regulation, respiratory coupling, and even long-term behavioral rhythms, but most models handle it along a single time axis. Here, the resolution hypothesis is proposed, holding that the temporal resolution at which meaning emerges differs depending on the health outcome.

The core idea redefines representation learning not as "which features matter" but as "at which time scale does the information emerge." Fine-grained morphological changes over short time spans fit signals like arrhythmia better, while slow changes over long time spans better fit sleep stages or chronic cardiovascular indicators.

As a way to implement this, a U-Net style hierarchical convolutional encoder-decoder is proposed. As layers are stacked deeper, the receptive field grows larger, indirectly capturing long-term dependencies, while skip connections preserve the high-resolution waveform information from shallow layers. In contrast, transformers connect every time point directly through self-attention, so the computational cost grows as O(L^2).

When this architecture is trained with masked autoencoding, the process of masking and reconstructing certain segments lets the model learn local continuity and long-term context together. More importantly, the intermediate representations themselves become probes that reveal resolution-specific information. By attaching linear classifiers to different layers, one can check which health tasks depend on which time scales.

The observed pattern is clear.

• Cardiovascular outcomes: Hypertension and arrhythmia detection perform better at deeper layers, i.e., coarser temporal resolution.
• Sleep stage classification: Slow changes in heart rate variability and vascular tone matter most, favoring medium-to-coarse resolution.
• Blood test abnormalities: Hemoglobin and electrolyte imbalances are sensitive to fine, high-frequency distortions in the waveform, so they show up better at shallower layers, i.e., finer resolution.

As a result, the hierarchical convolutional model shows competitive or better performance at a much smaller scale than transformer-based models. The authors connect this not merely to an efficiency issue but to the claim that when the structure of the signal aligns with the model's inductive bias, learning becomes easier and the representations become more meaningful.

The conclusion is clear. Rather than compressing the signal into a single monolithic embedding, wearable health models should reveal and interpret the internal multi-resolution hierarchy. Doing so enables both lightweight, privacy-friendly continuous monitoring that runs on-device and clinically explainable interpretation at the same time.