Loading field map
A patient's electrocardiogram (ECG) is recorded on a monitor, but the trace is contaminated by muscle activity, power-line interference, and the patient's own movements. A neurologist examining an electroencephalogram (EEG) for seizure activity must distinguish a genuine epileptic spike from an artifact caused by eye blinking. These are the everyday challenges of biomedical signal processing: extracting reliable physiological information from recordings that are noisy, non-stationary, and often nonlinear. Over the past seven decades, five major frameworks have emerged to meet this challenge, each building on—and sometimes breaking with—the assumptions of its predecessors.
The first systematic framework for analyzing biosignals was rooted in the mathematics of stationary processes. Classical Spectral Analysis, built on the Fourier transform, assumed that the statistical properties of a signal—its mean, variance, and frequency content—do not change over time. This assumption made it possible to decompose an ECG or EEG into its constituent frequency components, revealing characteristic rhythms such as the alpha rhythm (8–12 Hz) in the EEG or the QRS complex in the ECG. Linear filtering, implemented with analog or early digital circuits, allowed engineers to remove noise by passing only the frequency band of interest. The framework's great strength was its mathematical tractability: the Fourier transform provided a complete, invertible representation, and linear filters could be designed with well-understood trade-offs between passband ripple and stopband attenuation. Yet the stationarity assumption was a severe limitation. Physiological signals are inherently non-stationary: heart rate varies with respiration, brain rhythms shift with cognitive state, and muscle artifacts appear and disappear unpredictably. Classical methods could not track these changes in real time.
The practical failure of fixed filters for applications such as fetal ECG monitoring—where the maternal heartbeat is a much larger interference that changes with respiration and movement—drove the development of Adaptive Signal Processing. Instead of designing a single filter in advance, adaptive algorithms such as the Least Mean Squares (LMS) and Recursive Least Squares (RLS) adjusted their coefficients continuously, learning the statistics of the noise and the signal from the incoming data. This allowed the filter to track slow changes in the interference, canceling it without distorting the underlying physiological signal. However, adaptive methods remained firmly within the linear framework: the filter output was still a weighted sum of its inputs, and the underlying model assumed that the signal and noise combined additively. The innovation was in the adaptation rule, not in the fundamental mathematics. Adaptive filters became workhorses for noise cancellation in ECG, EEG, and electromyography (EMG), but they could not handle signals whose frequency content changed rapidly—a transient epileptic spike, for example, lasts only tens of milliseconds and is poorly represented by a slowly adapting linear filter.
The need to capture brief, non-stationary events led to Time-Frequency Analysis, a framework that preserved the linear, transform-based worldview but abandoned the assumption of global stationarity. Instead of a single Fourier transform, time-frequency methods compute a joint representation of how the signal's energy is distributed over both time and frequency. The Short-Time Fourier Transform (STFT) achieved this by sliding a fixed window across the signal and computing the Fourier spectrum within each window. The wavelet transform went further, using windows of different lengths at different frequencies—short windows for high-frequency detail, long windows for low-frequency trends—to achieve a more flexible trade-off between time and frequency resolution. For biomedical signals, this was transformative: a neurologist could now see exactly when an epileptic spike occurred and what frequencies it contained, rather than just knowing that the overall spectrum had changed. Yet time-frequency methods still assumed that the signal could be decomposed into a sum of elementary components (sines, wavelets) that were themselves linear. They provided a richer description of non-stationarity, but they did not question the underlying linearity of the physiological system generating the signal.
A fundamentally different approach emerged from dynamical systems theory and chaos theory. Nonlinear Dynamics and Complexity Analysis argued that many physiological signals—heart rate variability, brain rhythms, tremor—are generated by nonlinear systems with a small number of interacting degrees of freedom. Instead of decomposing the signal into frequency components, this framework reconstructs the system's state space from the time series and computes measures of complexity and predictability: the largest Lyapunov exponent quantifies sensitivity to initial conditions (chaos), correlation dimension estimates the number of active degrees of freedom, and approximate or sample entropy measures the irregularity of the signal. These tools revealed that healthy physiological systems often exhibit a rich, complex variability—a healthy heart rate, for instance, fluctuates in a fractal-like manner—while disease or aging can lead to a loss of complexity (e.g., more regular, less variable heart rate in heart failure). This was a radical departure from earlier frameworks. Time-frequency analysis had localized linearity; nonlinear dynamics abandoned linearity altogether. The two frameworks coexist but answer different questions: time-frequency methods ask "what frequencies are present and when?", while nonlinear dynamics asks "what kind of dynamical system generated this signal?" The latter requires much longer, cleaner recordings and is more sensitive to noise, but it offers a mechanistic interpretation that linear methods cannot provide.
The most recent framework, Machine Learning and Deep Learning for Biosignals, represents a paradigm shift from handcrafted features to end-to-end learning. Instead of designing filters, choosing time-frequency representations, or computing complexity measures, deep neural networks learn directly from the raw signal (or minimally processed spectrograms) to perform tasks such as arrhythmia detection, sleep staging, or seizure prediction. Convolutional neural networks (CNNs) automatically learn hierarchical features—from simple edges in time or frequency to complex patterns spanning seconds—while recurrent architectures (LSTMs, GRUs) capture long-range temporal dependencies. The performance of these models on benchmark datasets has often surpassed that of traditional methods, sometimes dramatically. However, this success comes with a loss of interpretability: a deep network that classifies an ECG as "atrial fibrillation" cannot easily explain which morphological features drove its decision. This has created a living tension with the Nonlinear Dynamics framework, which prizes mechanistic understanding and interpretable complexity measures. The two frameworks disagree on what constitutes an explanation: nonlinear dynamics offers a mathematical model of the underlying system, while deep learning offers a high-accuracy mapping from input to output. Yet they are increasingly combined—complexity measures are used as features in shallow machine learning models, and deep networks are probed with techniques like saliency maps to recover interpretability.
Today, all five frameworks remain active, each with a distinct role. Classical spectral analysis and linear filtering are still the first-line tools in clinical monitoring devices because they are fast, well-understood, and approved by regulators. Adaptive filters continue to be used for real-time noise cancellation in wearable devices. Time-frequency analysis is the standard method for analyzing transient events in EEG and evoked potentials. Nonlinear dynamics provides a theoretically grounded approach to quantifying physiological complexity and is widely used in heart rate variability research and anesthesia monitoring. Machine learning and deep learning dominate large-scale diagnostic tasks and are rapidly being integrated into commercial devices. The leading frameworks—Nonlinear Dynamics and Deep Learning—agree that biosignals are nonlinear and non-stationary, but they disagree on the primacy of interpretability versus predictive power. This is not a conflict that will be resolved by one side winning; rather, the field is moving toward hybrid approaches that combine the mechanistic insight of dynamical systems with the pattern-recognition power of deep networks. A student entering biomedical signal processing today will need to be fluent in all five frameworks, because the choice of method depends on the question: are you building a real-time filter for a pacemaker, characterizing the complexity of a patient's heart rate, or training a model to screen millions of ECGs for rare arrhythmias? Each framework has its place, and the history of the subfield is the story of how engineers learned to match the tool to the signal.