Sampling is not a formality
Sampling rate determines what can be measured and what is lost. ECG morphology, EEG oscillations, and respiration trends all impose different temporal demands. If the acquisition rate is too low, aliasing can distort physiologically important structure before any downstream model sees it.
The Nyquist condition is basic, but in physiology the practical question is tougher: what is the highest relevant frequency after accounting for noise, sensor bandwidth, and clinically meaningful morphology?
Filtering requires scientific judgment
Filtering is used to suppress motion artifacts, baseline drift, power-line noise, muscle contamination, and other nuisances. But every filter encodes a claim about which components matter. An aggressive high-pass filter may clean a trace while also distorting slow dynamics; a narrow notch may remove interference but introduce ringing or phase issues.
For biosignals, filtering is never just cosmetic preprocessing. It changes the evidence available for interpretation.
Time and frequency views are complementary
Some physiological structure is easiest to see in waveform morphology, while other structure emerges in spectral or time-frequency representations. Heart-rate variability, neural oscillations, tremor content, and respiratory modulation all depend on frequency-domain reasoning.
That is why Fourier methods, spectral density estimation, and wavelet or short-time analyses remain central. They provide different views of the same underlying physiology.
Stationarity is often only approximate
Classical signal processing often assumes stable statistics over time. Physiological data frequently violate that assumption. Sleep stages change, stress responses evolve, movement artifacts appear suddenly, and interventions alter baseline behavior.
This is why windowing, adaptive estimation, and state-space thinking are so important. A biosignal model that assumes one fixed regime may fail on the most interesting real-world segments.
Artifacts are part of the problem
In physiological sensing, artifact is not a marginal issue. It is a central property of the data. Motion, poor contact, sensor displacement, sweating, ambient light, and muscle activity can all reshape the signal. A robust pipeline must distinguish between physiological variation and acquisition failure.
That distinction matters scientifically and operationally. Otherwise the model may become good at detecting sensor conditions rather than biological state.
Why the classical toolkit still matters
Classical concepts are what make modern modeling credible. Good sampling preserves the relevant bandwidth. Good filtering protects morphology. Good spectral analysis exposes structure. Good artifact handling prevents false confidence. These are not alternatives to learning-based models; they are part of the measurement discipline that makes those models trustworthy.