Loading field map
Biomedical signal processing is the branch of engineering concerned with obtaining, interpreting, and modifying electrical or other physical signals that originate from living organisms, particularly humans. Its central task is to turn raw physiological measurements—voltages on the skin, pressure waves in arteries, magnetic fields from the brain—into information that is clinically useful or scientifically meaningful. The field sits at the intersection of electrical engineering, applied mathematics, and physiology, and it is defined less by a single method than by a recurring set of problems: signals are weak, noisy, non-stationary, and generated by systems whose internal workings are only partially known.
A signal, in this context, is any measurable quantity that varies over time or space and carries information about a biological process. The most familiar examples are the electrocardiogram (ECG), which records the heart's electrical activity from electrodes on the skin; the electroencephalogram (EEG), which records cortical electrical activity from the scalp; and the electromyogram (EMG), which records muscle electrical activity. Related signals include the magnetoencephalogram (MEG), blood pressure waveforms, respiratory airflow, and pulse oximetry traces. What unites them is that they are all indirect observations: the signal at the sensor is a distorted, attenuated, and contaminated version of the underlying physiological event.
Three properties make biomedical signals especially challenging. First, they are non-stationary: their statistical properties change over time. A heartbeat is not a perfectly repeating waveform; the interval between beats varies with breathing, posture, and mental state. Second, they are noisy, with noise arising from many sources—electrical interference from power lines, movement artifacts from electrodes shifting on the skin, and physiological activity from organs other than the one of interest. Third, they are coupled: signals from different systems interact. The heart rate changes with respiration, and brain activity modulates muscle tone. These properties mean that simple filtering or averaging, which works well for stationary laboratory signals, is often inadequate.
The field is organized around a small number of enduring questions. Detection asks whether a particular event or condition is present: Is there a seizure in this EEG? Is there a brief blockage in this heart rhythm recording? Estimation asks for a continuous quantity: What is the instantaneous heart rate? How deep is anesthesia? Classification asks which of several categories a signal belongs to: Is this ECG segment normal or pathological, and if pathological, which of several arrhythmias is it? Compression asks how to store or transmit signals efficiently without losing clinically important information. Modeling asks what underlying physiological process could have generated the observed signal, often with the goal of predicting future behavior.
The stakes are unusually high for an engineering discipline. Decisions made from processed signals can trigger alarms in intensive care units, guide the delivery of defibrillation shocks, determine whether a patient is declared brain-dead, or adjust the dose of a drug delivered by an infusion pump. A false positive in a screening test causes anxiety and unnecessary procedures; a false negative can be fatal. This places a premium on methods whose behavior is well understood and whose failure modes are characterized, and it explains why the field has historically been cautious about adopting black-box techniques whose decisions cannot be inspected.
The origins of biomedical signal processing lie in the late nineteenth and early twentieth centuries, when physiologists first developed instruments to record electrical activity from the body. Augustus Waller recorded the first human ECG in 1887 using a capillary electrometer, and Willem Einthoven improved it with the string galvanometer, which produced the familiar P-QRS-T waveform and earned him the Nobel Prize in 1924. These early recordings were made on photographic paper and analyzed by eye. The field was, at this stage, a branch of physiology: the goal was to understand the heart's electrical behavior, not to process signals in the modern sense.
The crucial transformation came with the availability of digital computers in the 1950s and 1960s. Analog signals could now be sampled, stored, and manipulated mathematically. Early work focused on automating what clinicians did visually: detecting the QRS complex in the ECG, counting heartbeats, and measuring intervals. The 1960s also saw the development of the fast Fourier transform (FFT), which made frequency analysis practical and opened the way to spectral analysis of EEG and other signals. By the 1970s, the field had coalesced as a distinct engineering discipline, with dedicated journals, conferences, and academic programs.
A second transformation began in the 1980s and 1990s with the rise of wavelet analysis, which provided a way to analyze signals at multiple resolutions simultaneously. Unlike the Fourier transform, which gives only frequency content and loses time information, wavelets can localize events in both time and frequency. This proved valuable for detecting transient features—a spike in an EEG, a brief change in heart rhythm—that are invisible in a global spectrum. The same period saw the introduction of adaptive filtering, in which filter coefficients adjust automatically to the signal, and the growing use of hidden Markov models and other probabilistic tools for signals that evolve through discrete states.
The most recent major development, beginning in the 2010s, has been the application of deep learning. Neural networks with many layers can be trained directly on raw signals to detect patterns that are difficult to specify analytically. This has produced striking results in areas such as arrhythmia detection from single-lead ECG and sleep staging from EEG. However, deep learning has also created tension within the field, because these models are often opaque: they make accurate predictions but do not explain why, and they can fail in unexpected ways on data that differs from their training set.
The field is not divided into rival schools in the way that, say, theoretical physics has competing interpretations. Rather, it contains several coexisting methodological traditions, each with its own assumptions and typical applications. These traditions overlap and are often combined in practice.
The oldest and most fundamental approach treats the signal as a waveform to be cleaned and characterized. Filtering removes unwanted components: a low-pass filter smooths high-frequency noise, a notch filter removes 50 or 60 Hz power-line interference, and a band-pass filter isolates a frequency range of interest. The design of such filters is well understood from general signal processing, but biomedical applications impose special constraints. For example, filtering an ECG must not distort the sharp QRS complex, which contains high frequencies, while still removing baseline wander caused by respiration or electrode movement. This has led to specialized techniques such as median filtering and adaptive baseline removal.
Spectral analysis computes the distribution of power across frequencies. The Fourier transform of an EEG segment reveals the classic frequency bands—delta, theta, alpha, beta, gamma—which change with sleep stage, alertness, and neurological disorders. Spectral analysis of heart rate variability, the small beat-to-beat fluctuations in heart rate, provides an indirect window into autonomic nervous system activity. A limitation of classical spectral analysis is that it assumes the signal is stationary, at least within the analyzed window. For signals that change rapidly, such as an EEG during a seizure, the assumption fails, and the result is a blurred average that may miss the event of interest.
To handle non-stationarity, methods were developed that represent a signal in both time and frequency simultaneously. The short-time Fourier transform (STFT) divides the signal into short overlapping windows and computes the spectrum of each window, producing a spectrogram. The choice of window length involves a trade-off: a short window gives good time resolution but poor frequency resolution, and vice versa. The wavelet transform improves on this by using windows of different lengths at different frequencies—short windows for high frequencies, long windows for low frequencies—which better matches the way physiological events actually occur. A QRS complex, for instance, is a brief, high-frequency event embedded in a slower waveform; wavelets can isolate it cleanly.
These methods are not competitors but complements. The STFT is simpler and remains useful when the signal is only mildly non-stationary. Wavelets are more powerful for transient detection and are now standard for analyzing EEG spikes, epileptic seizures, and the fine structure of heart sounds. A related family of methods, empirical mode decomposition, decomposes a signal into intrinsic mode functions without assuming a fixed basis, which can be useful for highly irregular signals, though its mathematical properties are less well established than those of wavelets.
A different tradition begins not with the signal itself but with a model of how the signal was generated. The most influential model is the autoregressive (AR) model, which assumes that each sample of the signal is a weighted sum of previous samples plus a small random input. AR models provide a compact parametric description of a signal's spectrum and are particularly useful for short data segments, where classical spectral estimates are unreliable. They have been applied extensively to EEG analysis, where the model coefficients themselves can serve as features for classification.
More elaborate models incorporate physiological knowledge. For example, the ECG can be modeled as the output of a dynamical system that generates a sequence of P-QRS-T complexes with variable timing; fitting such a model to a recording yields estimates of parameters that correspond to physiological quantities. State-space models, in which an unobserved physiological state evolves over time and produces noisy observations, provide a framework for tracking quantities like the depth of anesthesia or the probability of a seizure. These models are estimated using algorithms such as the Kalman filter, which recursively updates estimates as new data arrive. The advantage of model-based approaches is that they can separate signal from noise using assumptions about the underlying process, and they can provide confidence intervals for their estimates. The disadvantage is that the model may be wrong, and a misspecified model can produce confident but incorrect results.
The machine learning tradition treats signal processing as a classification or regression problem. The classical version, dominant from the 1980s through the 2000s, involved a two-stage pipeline: first, feature extraction, in which the signal is reduced to a set of numbers (e.g., spectral powers, wavelet coefficients, morphological measurements); second, classification, in which a statistical or machine learning algorithm (e.g., a support vector machine, a decision tree, a linear discriminant) maps the features to a label. This approach requires substantial domain expertise to design good features, but the resulting classifiers are often interpretable: one can see which features matter and why.
Deep learning represents a break with this pipeline. A deep neural network learns both the features and the classifier from raw data, given enough labeled examples. Convolutional neural networks (CNNs) can be applied directly to time-domain signals, and recurrent networks or transformers can capture temporal dependencies. Deep learning has achieved state-of-the-art results on many benchmark tasks, particularly when large datasets are available, such as the PhysioNet challenge datasets for arrhythmia detection. However, it has important limitations. It requires large amounts of labeled data, which are expensive to obtain in medicine. It is sensitive to differences between the training distribution and the deployment setting—a model trained on hospital A's equipment may fail on hospital B's. And its decisions are difficult to explain, which matters in a clinical context where physicians must justify their actions. The field is currently grappling with these issues, exploring methods for interpretability, uncertainty quantification, and domain adaptation.
These traditions are not historical stages that replaced one another; they coexist and are frequently combined. A typical modern system for detecting cardiac arrhythmias might use wavelet-based denoising as a preprocessing step, an AR model to estimate spectral features, and a machine learning classifier to make the final decision. The choice of approach depends on the signal, the question, and the constraints. Classical filtering is fast, well understood, and appropriate when the noise has known characteristics. Model-based methods are preferable when physiological knowledge is available and when uncertainty estimates are needed. Machine learning is attractive when the patterns are too complex to specify analytically and when large datasets exist.
The deepest divide in the field today is between explainable, model-based methods and opaque, data-driven methods. Proponents of the former argue that clinical decisions require understanding, and that a model whose reasoning cannot be inspected is dangerous. Proponents of the latter argue that accuracy is paramount, that the brain itself is an opaque pattern recognizer, and that the relevant test is whether the system improves patient outcomes, not whether a human can follow its reasoning. This debate is not resolved, and it is likely that the future lies in hybrid systems that combine the strengths of both—for example, using deep learning to propose features and model-based methods to verify them.
Contemporary biomedical signal processing is characterized by several converging trends. Wearable and mobile devices have moved signal acquisition out of the hospital and into daily life. Smartwatches record continuous ECGs and photoplethysmograms, and consumer EEG headsets are marketed for sleep tracking and meditation. This creates new opportunities for long-term monitoring and early detection, but also new challenges: signals recorded in uncontrolled environments are noisier, and the volume of data is enormous, requiring automated analysis that can run on low-power devices.
Data sharing and open benchmarks have transformed the field. Large public datasets, such as those hosted by PhysioNet, allow researchers to compare methods on common data, and annual challenges have driven rapid progress on specific tasks. This has accelerated the adoption of machine learning, since the necessary training data are now available for many problems.
Integration with clinical workflows remains the ultimate test. A signal processing method that works in a research setting must be validated on real patient populations, integrated into electronic health records, and approved by regulatory agencies before it affects patient care. This pipeline is slow and expensive, and many promising methods never make the transition. The field is therefore not only about algorithms but also about the practical engineering of reliable, safe, and usable systems.
The enduring identity of biomedical signal processing lies in its commitment to extracting reliable information from imperfect measurements of living systems. Its methods have changed dramatically over a century—from photographic paper to deep neural networks—but its core questions have remained remarkably stable. As sensors become cheaper, algorithms more powerful, and data more abundant, the field's importance is likely to grow, but so too will the need for careful validation and for methods whose limitations are as well understood as their capabilities.