Biomedical Signal Processing

Biomedical signal processing is the branch of biomedical engineering that extracts useful information from the electrical and physical signals the body produces—the heart's electrical rhythm, the brain's waves, the pressure of a pulse, the sound of breathing.

11 frameworks, from 1901 to today · Biomedical Engineering · Engineering

still a working view ended · each leads to its entry below

Overview

Biomedical signal processing is the branch of biomedical engineering that extracts useful information from the electrical and physical signals the body produces—the heart's electrical rhythm, the brain's waves, the pressure of a pulse, the sound of breathing1. Its practitioners take noisy, messy recordings made at the surface of the body or from inside it and turn them into measurements a clinician or a researcher can act on: is the heart beating regularly, is the brain responding to a flash of light, is a seizure beginning, is a patient's breathing stopping in their sleep. What is at stake is the conversion of raw physiology into decisions. A cardiologist looking at a paper strip and a machine-learning model screening a wearable recording are doing the same job at different scales: deciding what the signal means.

The galvanometer era: signals become visible

The field begins with the problem of making the body's tiny electrical currents visible at all. In 1842 the Italian physicist Carlo Matteucci showed that a frog's heart produces electrical activity, and in 1856 Rudolph von Kölliker and Heinrich Müller demonstrated that each heartbeat generates a current. The instrument that made this practical was the string galvanometer, developed by Willem Einthoven around 1901. A thin silvered quartz thread suspended in a magnetic field deflected when current passed through it, and a moving photographic plate recorded the deflection. Einthoven's electrocardiogram (ECG)—the recording of the heart's electrical activity from electrodes on the limbs—showed the familiar P, QRS, and T deflections, and he named them. He received the Nobel Prize in 1924. The ECG was the first biosignal to become a routine clinical tool, and the problem it posed—how to read a waveform reliably—defined the field for decades.

The same period saw the first brain recordings. Richard Caton, in 1875, recorded electrical activity from the exposed brains of rabbits and monkeys. Hans Berger, a German psychiatrist, recorded the first human electroencephalogram (EEG) in 1924 and published it in 1929, describing the alpha rhythm—a roughly 10 Hz oscillation seen when a subject is relaxed with eyes closed. Berger's work was met with skepticism, but it was confirmed in the 1930s by Edgar Adrian and others, and EEG became a clinical tool for diagnosing epilepsy and brain tumors.

These early instruments were analog and the "processing" was human: a trained reader scanned a paper trace for patterns. The field's first analytical ideas came from physics and mathematics. The Fourier transform—Joseph Fourier's 1822 result that any signal can be decomposed into a sum of sine waves of different frequencies—gave a way to describe a signal by its frequency content rather than its shape in time23. Applied to the EEG, it explained the alpha rhythm as a concentration of power near 10 Hz. Frequency-domain analysis became the standard vocabulary for describing brain waves and, later, heart rate variability.

The digital turn: sampling and computation

The decisive change came with digital computers. To process a signal by computer, it must first be sampled—measured at discrete instants—and converted to numbers. The sampling theorem, formulated by Harry Nyquist in 1928 and given its practical engineering form by Claude Shannon in 1949, states that a signal can be reconstructed exactly from samples taken at more than twice its highest frequency. This "Nyquist rate" told engineers how fast to sample and what went wrong (aliasing—the folding of high frequencies into low ones) when they sampled too slowly4. The analog-to-digital converter made the theorem operational, and by the 1960s and 1970s biomedical laboratories were digitizing ECG and EEG recordings.

The fast Fourier transform, an efficient algorithm for computing the Fourier transform published by James Cooley and John Tukey in 1965, made frequency analysis cheap enough to run routinely. Suddenly a long EEG record could be summarized as a spectrum in seconds. Digital filtering—the design of algorithms that pass some frequencies and block others—replaced analog circuits. The 1970s saw the first commercial computer-based ECG analysis systems, which detected QRS complexes and measured heart rate automatically, and the first digital EEG spectral analyzers.

Detection and estimation: the statistical framework

Once signals were numbers, the question became how to decide, from noisy data, what was happening. This is the domain of statistical detection and estimation theory56. Detection theory asks whether a signal is present (is there a heartbeat in this window?); estimation theory asks what its parameters are (how fast is the heart rate?)56. The Neyman–Pearson lemma, from Jerzy Neyman and Egon Pearson in the 1930s, gave the optimal test for deciding between two hypotheses when the noise statistics are known. The matched filter—the filter that maximizes signal-to-noise ratio for a known waveform in white noise—was developed in radar engineering during and after the Second World War and was imported into biosignal processing for tasks like detecting evoked responses in EEG.

A concrete case shows the power of the approach. In 1950s and 1960s, researchers wanted to detect the brain's response to a stimulus—a flash of light or a click—buried in the much larger ongoing EEG. The evoked potential is small, often a few microvolts against tens of microvolts of background7. Averaging many repetitions of the stimulus cancels the random background and leaves the consistent response7. This "signal averaging" technique, formalized in the 1950s and 1960s, made evoked-potential analysis a standard clinical test for hearing and nerve function. It is a direct application of the idea that a known signal repeated in noise can be recovered by averaging—the same logic as the matched filter.

Adaptive filtering and the non-stationary body

Classical filters assume the signal's statistics are fixed. The body violates this: heart rate changes, noise sources come and go, electrode contact shifts. Adaptive filtering addresses this by adjusting the filter's parameters continuously to minimize an error criterion8. The least-mean-squares (LMS) algorithm, published by Bernard Widrow and Marcian Hoff in 1960, and the recursive least squares family gave practical ways to do this910. In biomedical signal processing, adaptive filters became the tool for cancelling interference—for example, removing the mother's ECG from a fetal ECG recording, or cancelling power-line hum that drifts in frequency. The idea that a filter could learn the noise and subtract it was a conceptual shift: the processor became a small adaptive system rather than a fixed recipe.

Pattern recognition and the rise of machine learning

By the 1980s, the field had accumulated large databases of annotated signals—ECGs labeled by cardiologists, EEGs labeled by neurologists. The next problem was to automate the labeling. Pattern recognition and classical machine learning supplied the methods: feature extraction (computing descriptive numbers such as heart-rate variability, spectral power in bands, or waveform amplitudes), followed by a classifier (linear discriminant analysis, decision trees, support vector machines, artificial neural networks). The MIT–BIH Arrhythmia Database, released in 1980, became a standard benchmark for ECG beat classification. In the 1990s and 2000s, support vector machines and hidden Markov models were widely used for ECG and EEG classification. These systems were interpretable—each feature had a physiological meaning—and they worked well when the signal was clean and the task well defined.

Modeling the body: model-based signal processing

A parallel tradition treated the signal as the output of a physiological model and used the model to interpret the data. The circulatory system can be described as a network of elastic tubes and a pump; the ECG can be modeled as the sum of electrical sources in the heart moving through the torso. Physiological modeling and model-based signal processing use these descriptions to estimate hidden quantities—for example, estimating cardiac output from a blood-pressure waveform, or localizing the source of an epileptic discharge in the brain from scalp EEG. The approach is powerful because it constrains the answer to be physiologically plausible, but it depends on the model being right. It has been especially important in cardiac mechanics, respiratory monitoring, and source localization in EEG and magnetoencephalography.

Wearables and real-time embedded processing

The miniaturization of electronics moved biosignal processing out of the laboratory and onto the body. Holter monitors, introduced in the 1960s, recorded ECG continuously for a day on magnetic tape; by the 2000s, solid-state recorders and then wearable patches did the same at a fraction of the size. The key engineering constraint shifted: a wearable device has limited power, limited memory, and must process in real time. Embedded algorithms for QRS detection, such as the Pan–Tompkins algorithm published in 1985, were designed to run on small processors with minimal computation. The 2010s brought consumer wearables—optical heart-rate sensors, wrist-worn ECG—and with them a new problem: processing signals from noisy, uncontrolled environments where motion artifact dominates. Real-time embedded biosignal processing is now the layer between the sensor and the cloud, deciding locally what is worth transmitting.

Deep learning and the present

The most recent turning point is the application of deep learning11. Instead of hand-designing features, deep neural networks learn representations directly from raw or lightly processed signals11. Convolutional neural networks, which exploit local structure in time series, became the standard architecture for ECG and EEG classification in the 2010s. Recurrent networks and, later, transformer architectures were applied to longer sequences12. The change is methodological: the feature extractor is learned rather than specified, which helps when the relevant pattern is subtle or unknown, but it makes the model harder to interpret and raises questions about generalization across patients and devices. Large public datasets and challenges—such as the PhysioNet/Computing in Cardiology challenges, running since 2000—accelerated the shift by providing shared benchmarks.

Today the field spans a continuum from embedded real-time algorithms on a wearable to cloud-based deep-learning analysis of long recordings. The classical tools—Fourier analysis, filtering, detection theory, adaptive filtering—remain the foundation and are still used where interpretability, power, or data are limited. The open questions are the ones the history has produced: how to make learned models trustworthy and generalizable across populations and devices, how to handle the non-stationarity of physiology, how to process at the edge without draining a battery, and how to combine the interpretability of model-based approaches with the flexibility of learned ones.

References

  1. Biosignal, Wikipedia
  2. Fourier transform (History), Wikipedia
  3. Fourier analysis, Wikipedia
  4. Nyquist–Shannon sampling theorem, Wikipedia
  5. Detection theory, Wikipedia
  6. Estimation theory, Wikipedia
  7. Evoked potential, Wikipedia
  8. Adaptive filter, Wikipedia
  9. Least mean squares filter, Wikipedia
  10. Recursive least squares filter, Wikipedia
  11. Deep learning, Wikipedia
  12. Recurrent neural network, Wikipedia

The frameworks of Biomedical Signal Processing

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1901-Present

    Electrocardiography and Cardiac Rhythm Analysis

    How is the heart's rhythm read from its electrical trace?

    Fly to its tower in Nousfarer

  2. 1901-1975

    Electrodiagnostic waveform interpretation

    What do the shapes of electrical waveforms from the body mean?

    Fly to its tower in Nousfarer

  3. 1924-Present

    Electroencephalography and Evoked-Potential Analysis

    What do the brain's electrical waves reveal about its activity?

    Fly to its tower in Nousfarer

  4. 1930-Present

    Frequency-domain analysis

    How can a signal be described by its frequency content?

    Fly to its tower in Nousfarer

  5. 1950-Present

    Detection theory

    How can a signal be detected in noise?

    Fly to its tower in Nousfarer

  6. 1960-Present

    Adaptive filter

    How can a filter adjust itself to changing noise?

    Fly to its tower in Nousfarer

  7. 1965-Present

    Digital signal processing

    How can computers process sampled signals efficiently?

    Fly to its tower in Nousfarer

  8. 1970-Present

    Model-based physiological signal processing

    How can physiological models interpret measured signals?

    Fly to its tower in Nousfarer

  9. 1980-2015

    Pattern recognition

    How can machines automatically classify signal patterns?

    Fly to its tower in Nousfarer

  10. 1985-Present

    Real-time embedded biosignal processing

    How can biosignals be processed in real time on small devices?

    Fly to its tower in Nousfarer

  11. 2012-Present

    Deep learning

    How can deep networks learn directly from raw signals?

    Fly to its tower in Nousfarer

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Electrocardiography and Cardiac Rhyth…Electrodiagnostic waveform interpreta…Electroencephalography and Evoked-Pot…Frequency-domain analysisDetection theoryAdaptive filterDigital signal processingModel-based physiological signal proc…Pattern recognitionReal-time embedded biosignal processi…Deep learning

Beyond the field

Where Biomedical Signal Processing's ideas came from in other fields, and what they led to there: the strongest 12 of 59 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Biomedical Signal Processing is a planet in Nousfarer

Its city has a tower for each of these 11 frameworks. Land, walk in, and climb one concept at a time.

Land on Biomedical Signal Processing

Related fields

Fields its ideas are linked to

More in Biomedical Engineering

See where Biomedical Signal Processing sits on the Map of Human Knowledge