Physiological signals are inherently hierarchical. When a clinician looks at an electrocardiogram (ECG), they aren't just looking at a sequence of numbers; they are looking for millisecond-level "spikes" (the QRS complex), second-level heartbeats, and minute-level rhythmic trends. Traditional machine learning models often treat these signals as "flat" sequences, forcing a single temporal resolution onto data that lives across many.
Recent research has converged on a singular insight: to truly understand the human body, models must "zoom in" and "zoom out" simultaneously. This direction, often termed multiscale or multi-granularity representation learning, moves away from fixed-window analysis toward architectures that mirror the nested rhythms of biology.
The Problem of Scale Disparity
In standard signal processing, there is a fundamental trade-off. If you look at a signal with a fine-grained lens, you capture sharp, transient events like epileptic spikes or heart valve clicks, but you lose the "big picture" of the overall physiological state. Conversely, if you look at a coarse-grained lens to see long-term trends, the vital high-frequency fluctuations are smoothed away as noise.
Wenpo Yao and colleagues argue that analyzing data at a single resolution overlooks the complex dynamics that only emerge across multiple scales. This is particularly true in disease states; for instance, Jianbo Gao and Jing Hu demonstrated that biological signals like heart rates operate across scales ranging from milliseconds to minutes, and failing to capture this multiscale nature leads to an incomplete understanding of system complexity.
Hierarchical Architectures: Tokens, Beats, and Rhythms
To solve this, researchers are building models that explicitly encode the hierarchy of the body. Instead of treating an ECG as a long string of tokens, Fuying Wang and Lequan Yu introduced MELP, which aligns signals with clinical text across three levels: individual tokens, heartbeats, and overall rhythms.
Similarly, Nan Huang and Xiang Zhang developed Medformer, a Transformer that uses "cross-channel patching" to capture diagnostic markers that appear at distinct temporal scales. By using different patch sizes, the model can attend to high-frequency fluctuations and low-frequency trends at the same time. This hierarchical approach is also central to SleepMaMi, where Keondo Park and colleagues separate the analysis of fine-grained signal textures (micro-structures) from long-term temporal rhythms (macro-structures) to improve sleep staging.
Progressive Fusion and Gated Mechanisms
Capturing multiple scales is only half the battle; the model must also know how to combine them. Yiling Zhang and Hao Wang proposed the Multi-scale Progressive Gated Transformer (MPGT), which uses a specialized gating mechanism to fuse local fluctuations with global trends. This prevents the "global" information from overwhelming the "local" details, ensuring that a sudden, tiny spike in an EEG isn't washed out by a minute of steady background activity.
In the clinical setting, this fusion is vital for interpretability. Liantao Ma and Yasha Wang developed AdaCare, which uses scale-adaptive feature extraction to mirror clinical intuition. By using varying receptive fields, the model can filter out noise while remaining sensitive to urgent, acute medical events.
Self-Supervised Multiscale Discovery
A newer frontier in this area is using self-supervised learning to "discover" these scales without human labels. Megha Thukral and Viswam Nathan introduced Masked Multiscale Reconstruction (MMR), which uses wavelet decomposition to force a model to reconstruct masked signals across different frequency scales. This teaches the model to understand everything from the shape of a single pulse to the long-term rhythm of heart rate variability.
Similarly, the HiMAE framework treats physiological signals as multi-resolution structures rather than flat sequences. By masking and reconstructing data at different temporal resolutions, the model learns which health markers reside at which scales—for instance, identifying that millisecond-level heart rhythms are better for cardiac health, while minute-level cycles are better for sleep analysis.
Why It Works: The Fractal Nature of Biology
The reason multiscale approaches outperform "flat" models is that biological systems are often multifractal. As Yufan Hu and Yixuan Yuan note in their work on fMRI synthesis, neural signals exhibit scale-invariant patterns where activity fluctuations look similar across different time resolutions.
By building models that respect this hierarchy—from the MNL-Network for epilepsy to XTSFormer for irregular clinical events—researchers are moving closer to an artificial intelligence that "sees" the body the way a doctor does: as a complex, nested system of rhythms within rhythms.