Sources and References

These notes synthesize material from standard textbooks in biomedical signal processing and medical instrumentation. No instructor-specific or offering-specific materials are cited. BME 544 is cross-listed with SYDE 544 at the University of Waterloo, and the two designations refer to the same course.

• Rangayyan, R. M. Biomedical Signal Analysis, 2nd ed. Wiley-IEEE Press.
• Semmlow, J. L. and Griffel, B. Biosignal and Medical Image Processing, 3rd ed. CRC Press.
• Webster, J. G. (ed.) Medical Instrumentation: Application and Design, 4th ed. Wiley.
• Enderle, J. D. and Bronzino, J. D. Introduction to Biomedical Engineering, 3rd ed. Academic Press.
• Proakis, J. G. and Manolakis, D. G. Digital Signal Processing: Principles, Algorithms, and Applications, 4th ed. Pearson.
• Malmivuo, J. and Plonsey, R. Bioelectromagnetism. Oxford University Press.
• MIT OpenCourseWare, 6.022J / HST.542J Quantitative Physiology: Organ Transport Systems.
• PhysioNet tutorials on ECG, HRV, EEG, and polysomnography (physionet.org).

Chapter 1: Biosignal Origins and Biopotential Electrodes

1.1 What Makes a Signal a Biosignal

A biosignal is any measurable variable that varies in time or space and that reflects the state of a living system. In practice the term is dominated by electrical recordings from excitable tissue, but it also covers mechanical, acoustic, optical, thermal, and chemical variables. The signals treated in this course fall into three broad families. Bioelectric signals such as the electrocardiogram (ECG), electroencephalogram (EEG), and electromyogram (EMG) arise from transmembrane ionic currents in cardiac, neural, and muscular tissue. Biomechanical signals describe position, force, acceleration, and pressure and are acquired with strain gauges, load cells, piezoelectric transducers, and inertial units. Biochemical and optical signals include pulse oximetry, near-infrared spectroscopy, and glucose sensing.

The common thread is that every biosignal is a transduced version of a physiological process. The measurement chain always contains a source tissue, a volume conductor (the body), a sensor or electrode, an analog front end, an anti-alias filter, an analog-to-digital converter, and a digital processor. Each stage imposes its own bandwidth, noise floor, and distortion, so the processing pipeline cannot be designed without knowledge of the physics upstream of it.

1.2 The Cellular Origin of Bioelectric Signals

Excitable cells maintain a resting transmembrane potential near \( -70 \) mV through the combined action of the sodium-potassium pump and selective ion channels. The Nernst equation gives the equilibrium potential for a single ion species,

\[ E_{\text{ion}} = \frac{RT}{zF} \ln\!\left(\frac{[\text{ion}]_{o}}{[\text{ion}]_{i}}\right), \]

while the Goldman-Hodgkin-Katz equation predicts the resting potential from the relative permeabilities to \( \mathrm{Na}^{+} \), \( \mathrm{K}^{+} \), and \( \mathrm{Cl}^{-} \). When the membrane depolarizes past threshold, voltage-gated sodium channels open, an action potential propagates along the cell, and transmembrane currents flow into the surrounding extracellular fluid. These currents are the true sources of the ECG, EEG, and EMG.

Because the extracellular medium is a volume conductor, the potential at a remote surface electrode is a spatially filtered superposition of many cellular sources. The forward problem, which maps cellular currents to surface potentials, can be formulated using the quasi-static approximation to Maxwell’s equations because biological frequencies are low enough that capacitive and inductive effects inside tissue are negligible. The inverse problem, recovering the sources from surface measurements, is ill-posed and is a central theme in EEG source localization.

1.3 Electrode-Electrolyte Interface

A biopotential electrode is not a passive wire touching skin; it is a two-phase electrochemical cell. At the metal-electrolyte interface a double layer forms, giving rise to a half-cell potential, a charge-transfer resistance, and a double-layer capacitance. A useful lumped model is a voltage source \( E_{hc} \) in series with the parallel combination of a resistance \( R_{ct} \) and capacitance \( C_{dl} \), all in series with an electrolyte spreading resistance \( R_{s} \).

Electrodes are classified as polarizable or non-polarizable depending on whether charge crosses the interface. A stainless-steel dry electrode is nearly perfectly polarizable and behaves like a capacitor; its impedance rises sharply at low frequencies, which is catastrophic for DC-coupled signals such as slow cortical potentials. A silver-silver chloride (Ag/AgCl) electrode with a chloride-containing gel is the canonical non-polarizable electrode: chloride ions exchange freely between the AgCl coating and the electrolyte, the half-cell potential is stable, and the impedance is low and flat down to DC. For this reason Ag/AgCl is the de facto standard for ECG, EEG, and surface EMG.

1.4 Motion, Offset, and Skin Preparation

The dominant artifact in wearable biopotential recording is not line noise but electrode motion. When the electrode moves relative to the skin, the double-layer charge is disturbed and the half-cell potential transiently changes, producing low-frequency offsets that can exceed the signal by two orders of magnitude. Mitigation strategies include light skin abrasion to reduce stratum corneum impedance, use of hydrogels that stabilize the interface, mechanical decoupling so that cable motion does not transmit to the electrode, and active driven-right-leg circuits that reduce common-mode voltage.

Chapter 2: Instrumentation Amplifiers and Front-End Design

2.1 Requirements of a Biopotential Front End

The front end stands between microvolt-level signals and millivolt-level line interference riding on the patient at common mode. Four requirements dominate its design: very high input impedance so that the electrode impedance does not attenuate the signal, very high common-mode rejection ratio (CMRR) so that mains interference is suppressed, low input-referred noise so that the signal is not buried, and adequate dynamic range so that DC offsets do not saturate subsequent stages.

The workhorse circuit is the three-op-amp instrumentation amplifier, which combines two non-inverting buffers with a difference amplifier. Its differential gain is set by a single external resistor, its input impedance is essentially that of the op-amp inputs (gigaohms for JFET or CMOS parts), and its CMRR can exceed 100 dB when laser-trimmed on a monolithic die such as the AD620 or INA128.

2.2 Common-Mode Rejection and the Driven Right Leg

Even with an excellent instrumentation amplifier, the patient capacitively couples to mains and floats at tens of millivolts common mode. The driven-right-leg (DRL) circuit senses the common-mode voltage at the amplifier inputs, inverts and amplifies it, and feeds it back to the body through a current-limited resistor. This negative feedback loop lowers the effective common-mode voltage by the loop gain, often by 40 to 60 dB, and is a hallmark of modern ECG and EEG amplifiers. The feedback resistor must be large enough to guarantee patient safety under fault conditions; standards such as IEC 60601 set maximum leakage currents in the tens of microamperes.

2.3 Filtering and AC Coupling

Biopotential signals typically occupy a narrow band on top of a large low-frequency offset. ECG diagnostic mode nominally spans 0.05 to 150 Hz, EEG spans 0.5 to 70 Hz, and surface EMG spans 20 to 500 Hz. Because the DC offset from mismatched electrode potentials can reach hundreds of millivolts, an AC-coupling high-pass filter is inserted early in the chain. Its time constant must be long enough not to distort the slowest signal component: a 0.05 Hz cutoff requires a 3.2 s time constant, which in a single RC stage demands a large resistor-capacitor product and careful leakage budgeting.

A low-pass filter provides the anti-alias function before sampling. In a conventional two-stage design a passive or low-order active low-pass filter limits the bandwidth, followed by a higher-order active filter (Butterworth, Bessel, or elliptic) to give a sharp cutoff below the Nyquist frequency. Bessel filters are preferred when preserving waveform morphology matters, because they have linear phase in the passband.

2.4 Isolation and Safety

Any amplifier connected to a patient must provide galvanic isolation between the patient-connected part and the mains-powered part. This is achieved with optocouplers, transformers, or capacitive isolation barriers. Isolation limits fault currents in the event of mains breakdown and breaks ground loops that would otherwise inject interference. Modern designs often use digital isolators after the ADC, shifting the isolation barrier to the digital domain where it is cheaper and more linear.

Chapter 3: Noise and Interference

3.1 Physical Noise Sources

Thermal (Johnson-Nyquist) noise in a resistor has a flat voltage spectral density

\[ S_{v}(f) = 4 k_{B} T R, \]

where \( k_{B} \) is Boltzmann’s constant, \( T \) is absolute temperature, and \( R \) is the resistance. The rms thermal noise of a \( 100 \,\mathrm{k\Omega} \) electrode over a 500 Hz bandwidth is already near the microvolt level and rivals the intrinsic noise of many op-amps, so electrode impedance must be driven down aggressively. Shot noise in diode and transistor junctions, and flicker (1/f) noise in CMOS devices, add to the input-referred noise of the amplifier. Op-amps for biopotential applications are specified by input voltage noise density in \( \mathrm{nV}/\sqrt{\mathrm{Hz}} \) and input current noise density in \( \mathrm{fA}/\sqrt{\mathrm{Hz}} \).

3.2 Powerline Interference

Fifty- or sixty-hertz interference enters the recording through three main mechanisms. Capacitive coupling from mains wiring to the patient injects common-mode current that the CMRR must reject. Capacitive coupling to the lead wires injects differential current that CMRR cannot reject and that requires shielding, twisted-pair leads, and balanced electrode impedances. Magnetic coupling through the loop area enclosed by the leads is mitigated by minimizing that loop area. When residual line interference remains in the recording, a narrow notch filter centered at the mains frequency is often applied. Digital notch filters realized as IIR biquads or as FIR designs with a zero pair on the unit circle at the offending frequency are both common.

3.3 Motion Artifact and Physiological Interference

Motion artifact has been introduced above as an electrode phenomenon. Physiological interference refers to the contamination of one biosignal by another. In surface EMG from the trunk, the ECG appears as a strong rhythmic artifact; in EEG, eye blinks produce large frontal deflections and heart beats introduce a ballistocardiographic rhythm. Because these interferers often share spectral content with the signal of interest, linear filtering is insufficient and adaptive or blind-source-separation techniques (Chapter 9) are called for.

Chapter 4: Sampling, Quantization, and Analog-to-Digital Conversion

4.1 The Sampling Theorem

A bandlimited signal \( x(t) \) with maximum frequency \( f_{\max} \) can be recovered exactly from its samples \( x[n] = x(nT_{s}) \) provided that the sampling rate \( f_{s} = 1/T_{s} \) exceeds \( 2 f_{\max} \). The reconstruction is

\[ x(t) = \sum_{n = -\infty}^{\infty} x[n]\, \mathrm{sinc}\!\left(\frac{t - nT_{s}}{T_{s}}\right). \]

In practice the signal is never strictly bandlimited, so the anti-alias filter must suppress out-of-band energy to below the quantization noise floor. A practical rule is to choose \( f_{s} \) between three and five times the highest frequency of interest, giving a transition band in which the anti-alias filter can roll off gracefully.

4.2 Quantization and the ADC

Quantization with \( B \) bits over a full-scale range \( V_{\mathrm{FS}} \) introduces an additive noise with variance \( q^{2}/12 \), where \( q = V_{\mathrm{FS}} / 2^{B} \). The ideal signal-to-quantization-noise ratio for a full-scale sinusoid is

\[ \mathrm{SQNR} = 6.02 B + 1.76 \;\;\text{dB}. \]

For a typical ECG amplifier with a gain of 500 and a \( \pm 5 \,\mathrm{mV} \) input range, a 12-bit ADC already provides better resolution than the electrode noise floor, and 16- to 24-bit sigma-delta ADCs are standard in modern front ends. Sigma-delta converters exploit oversampling and noise shaping to push quantization noise out of the signal band, and they pair naturally with decimation filters that also serve as anti-alias filters.

4.3 Aliasing in Practice

When anti-alias filtering is inadequate, high-frequency interference folds into the signal band and cannot subsequently be removed. EMG activity above the anti-alias cutoff, spike noise from switching electronics, and radiofrequency pickup are frequent culprits. A practical diagnostic is to record with the subject at rest and examine the spectrum for suspiciously narrow peaks or a rising noise floor at high frequencies.

Chapter 5: Spectral Analysis

5.1 The Discrete Fourier Transform

The discrete Fourier transform (DFT) of a length-\( N \) signal is

\[ X[k] = \sum_{n=0}^{N-1} x[n] \exp\!\left(-j \frac{2\pi k n}{N}\right),\quad k = 0, 1, \ldots, N-1. \]

The fast Fourier transform (FFT) computes this sum in \( O(N \log N) \) time. Because biosignals are rarely periodic over the record length, directly applying the DFT to a finite segment produces spectral leakage: energy at frequency \( f_{0} \) smears into neighboring bins. Windowing the segment with a taper such as Hann, Hamming, or Blackman reduces leakage at the cost of slightly wider main lobes.

5.2 The Periodogram and Its Variance

The periodogram estimate of the power spectral density (PSD) is \( |X[k]|^{2}/N \). It is asymptotically unbiased but its variance does not decrease with record length, so a single long periodogram is still a noisy estimate. The Welch method addresses this by dividing the record into overlapping segments, applying a window to each, computing the periodogram of each, and averaging. With \( K \) segments the variance falls roughly as \( 1/K \) while the frequency resolution is determined by the segment length. Typical choices are 50 percent overlap and Hann windows of 1 to 4 s duration for EEG, giving sub-hertz resolution.

5.3 Multitaper and Parametric Methods

The multitaper method uses a set of orthogonal discrete prolate spheroidal sequences (Slepian tapers) and averages their periodograms. It provides the optimal trade-off between bias and variance for a given frequency resolution and is particularly attractive for short EEG segments during event-related paradigms. Parametric methods fit an autoregressive (AR) model of order \( p \),

\[ x[n] = -\sum_{k=1}^{p} a_{k}\, x[n-k] + e[n], \]

and compute the PSD from the estimated coefficients. AR spectra are smooth and can resolve closely spaced peaks, but model order selection is delicate; criteria such as AIC or BIC guide the choice.

Chapter 6: ECG Processing

6.1 Genesis of the ECG

Each heartbeat begins with pacemaker activity in the sinoatrial node and propagates through atrial myocardium, the atrioventricular node, the bundle of His, and the Purkinje fibers into ventricular muscle. The surface ECG records the dipolar field generated by this coordinated depolarization and repolarization. The P wave reflects atrial depolarization, the QRS complex reflects ventricular depolarization, and the T wave reflects ventricular repolarization. Twelve standard leads (six limb leads and six precordial leads) project the cardiac dipole onto different spatial directions, enabling the clinician to localize ischemia, infarction, hypertrophy, and conduction disturbances.

6.2 Preprocessing

ECG preprocessing typically removes baseline wander below about 0.5 Hz using a high-pass filter or a cubic spline fit, suppresses mains interference with a narrow notch, and limits the bandwidth to roughly 40 Hz for monitoring or 150 Hz for diagnostic use. Care must be taken not to introduce ringing that mimics pathology; linear-phase filters are generally preferred and zero-phase filtering by forward-backward application is common offline.

6.3 QRS Detection

Reliable detection of QRS complexes underpins heart-rate estimation, arrhythmia analysis, and synchronous averaging of T-wave or P-wave features. The Pan-Tompkins algorithm is the canonical approach and proceeds in five steps: a bandpass filter with passband roughly 5 to 15 Hz emphasizes the QRS, a derivative approximates slope, a squaring operation emphasizes large slopes and enforces positivity, a moving-window integration produces an envelope, and an adaptive threshold selects candidate peaks. A refractory period of about 200 ms prevents double detection within a single beat, and a secondary lower threshold performs a searchback when no beat is found within 1.66 times the running average RR interval.

Wavelet-based detectors exploit the scale-selective response of QRS complexes to dyadic wavelet transforms; they are robust in low-signal-to-noise conditions and are widely used in ambulatory monitors. Deep learning detectors trained on PhysioNet databases such as MIT-BIH have surpassed classical algorithms in sensitivity and specificity on noisy records, at the cost of data hunger and interpretability.

6.4 Heart Rate Variability

Once beats are annotated, the series of RR intervals is resampled to a uniform grid (commonly 4 Hz) and analyzed in time and frequency domains. Time-domain metrics include SDNN, the standard deviation of normal-to-normal intervals, and RMSSD, the root mean square of successive differences, which reflects parasympathetic vagal tone. Frequency-domain analysis, usually via Welch or Lomb-Scargle to handle irregular sampling, partitions the spectrum into very-low-frequency (below 0.04 Hz), low-frequency (0.04 to 0.15 Hz), and high-frequency (0.15 to 0.4 Hz) bands. The high-frequency band is dominated by respiratory sinus arrhythmia and is interpreted as a vagal index; the low-frequency band reflects mixed sympathetic and parasympathetic activity and is influenced by baroreflex dynamics.

ECG band	Passband	Purpose
Monitoring	0.5 to 40 Hz	Rhythm analysis, bedside display
Diagnostic	0.05 to 150 Hz	Preservation of ST segment and high-frequency notching
Pan-Tompkins internal	5 to 15 Hz	QRS emphasis for detection

Chapter 7: EEG Processing

7.1 Rhythms and Generators

Scalp EEG reflects the summed postsynaptic potentials of large populations of cortical pyramidal neurons oriented radially to the skull. It is conventionally partitioned into delta (below 4 Hz), theta (4 to 8 Hz), alpha (8 to 13 Hz), beta (13 to 30 Hz), and gamma (above 30 Hz) bands. Alpha rhythm, most prominent over posterior cortex with eyes closed, reflects thalamo-cortical oscillations; beta is associated with alert cognition and motor planning; gamma is implicated in perceptual binding. Sleep stages are defined by stereotyped combinations of these rhythms with sleep spindles, K-complexes, and slow oscillations.

7.2 Artifact Rejection

EEG is contaminated by ocular, muscular, cardiac, and movement artifacts. Eye blinks produce large frontal deflections; saccades produce step-like contamination. Simple approaches reject epochs that exceed an amplitude threshold or correlate too strongly with a reference electrooculogram (EOG). More sophisticated approaches use blind source separation. Independent component analysis (ICA) separates the multichannel EEG into maximally independent components, and components whose scalp maps and time courses match known artifact signatures are removed before reconstruction. This has become routine in research-grade EEG pipelines such as EEGLAB and MNE-Python.

7.3 Event-Related Potentials

When EEG is time-locked to a repeated stimulus and averaged across many trials, ongoing activity that is uncorrelated with the stimulus averages toward zero while the stimulus-locked response emerges. The resulting event-related potential (ERP) contains components such as the P300, a positive deflection around 300 ms post-stimulus that indexes attention and novelty, and the N400, which indexes semantic processing. Signal-to-noise ratio improves as \( \sqrt{N} \) with the number of averaged trials, and careful experimental design must control for baseline drifts and trial-to-trial latency jitter that smears the average.

7.4 Source Localization Basics

Scalp potentials are a spatially smoothed view of cortical currents. The forward problem uses a volume-conductor head model (spherical shells or realistic boundary-element or finite-element meshes derived from MRI) to predict scalp potentials from assumed dipole sources. The inverse problem seeks source distributions consistent with measurements. It is underdetermined (more sources than sensors) and requires regularization. Minimum-norm estimates, weighted minimum-norm variants such as sLORETA, and beamformers such as LCMV are the main families. Combined EEG-MEG and EEG-fMRI recordings and individualized head models based on MRI substantially improve localization.

Chapter 8: EMG Processing

8.1 Motor Unit Action Potentials

A motor unit consists of a motor neuron and the muscle fibers it innervates. When the motor neuron fires, all its fibers contract nearly simultaneously and their action potentials superimpose into a motor unit action potential (MUAP). The surface EMG is the spatiotemporal summation of many MUAP trains, weighted by the volume-conductor transfer function from fiber to skin. Because fibers at different depths contribute with different low-pass characteristics, surface EMG is richer in low frequencies than intramuscular recordings.

8.2 Envelope and Amplitude Estimation

For gross force estimation and prosthetic control, what matters is not the individual MUAPs but the slowly varying amplitude of the EMG. Standard pipelines high-pass the signal at 10 to 20 Hz to remove motion artifact, rectify (full-wave) to produce a strictly positive signal, and then low-pass smooth at a few hertz, yielding the linear envelope. Alternative estimators include the root-mean-square over a sliding window and a Bayesian amplitude estimator that models the EMG as a stochastic process with a slowly varying standard deviation.

8.3 Decomposition and Force Estimation

EMG decomposition recovers the firing times of individual motor units from the composite signal. Template-matching algorithms detect candidate MUAPs, cluster them by shape, and assign each to a unit; blind-source-separation approaches such as convolutive ICA applied to high-density surface grids (e.g., 8 by 8 electrodes) can decompose activity during moderate contractions. Decomposition reveals motor-unit recruitment thresholds, firing-rate modulation, and synchronization, and provides a physiologically grounded input to force estimation models.

Force estimation from EMG uses either a static nonlinear map from envelope to force (often a second-order polynomial) or a Hill-type muscle model that accounts for activation dynamics, force-length, and force-velocity relations. Accurate force estimation is central to myoelectric prosthetic control.

Chapter 9: Adaptive Filtering

9.1 The Wiener Filter

The Wiener filter is the linear filter that minimizes the mean-squared error between its output and a desired signal. For a finite-impulse-response filter of length \( M \), the optimal coefficient vector is

\[ \mathbf{w}_{\text{opt}} = \mathbf{R}^{-1} \mathbf{p}, \]

where \( \mathbf{R} \) is the autocorrelation matrix of the input and \( \mathbf{p} \) is the cross-correlation vector between input and desired signal. In biosignal practice the statistics are unknown and time-varying, so one uses adaptive algorithms that estimate the coefficients online.

9.2 LMS, NLMS, and RLS

The least-mean-squares (LMS) algorithm updates the coefficients by stochastic gradient descent,

\[ \mathbf{w}[n+1] = \mathbf{w}[n] + \mu\, e[n]\, \mathbf{x}[n], \]

where \( e[n] \) is the instantaneous error and \( \mu \) is a step size chosen small enough for stability. The normalized LMS divides the step by the instantaneous input energy, decoupling stability from input level. Recursive least squares (RLS) tracks the exact least-squares solution with a forgetting factor and converges much faster than LMS at the cost of higher complexity.

9.3 Adaptive Noise Cancellation

A classic biomedical application is fetal ECG extraction. A chest electrode records the mixed maternal and fetal ECG; an abdominal electrode records a reference dominated by the maternal component. An adaptive filter shapes the maternal reference to cancel the maternal contribution in the mixed signal, leaving the fetal ECG as the residual. Similar architectures cancel the ECG from diaphragmatic EMG and the EOG from frontal EEG channels.

Chapter 10: Time-Frequency Analysis

10.1 Motivation

Biosignals are rarely stationary. Sleep EEG transitions between stages on a scale of minutes; EMG bursts last hundreds of milliseconds; ECG morphology changes beat to beat during arrhythmia. A stationary spectrum hides these dynamics. Time-frequency analysis localizes spectral content in time, trading frequency resolution for time resolution according to the Heisenberg uncertainty principle,

\[ \Delta t \cdot \Delta f \ge \frac{1}{4\pi}. \]

10.2 Short-Time Fourier Transform

The short-time Fourier transform (STFT) slides a window \( w[n] \) along the signal and computes the DFT of each windowed segment,

\[ X(n, \omega) = \sum_{m} x[m]\, w[m-n]\, e^{-j\omega m}. \]

The squared magnitude is the spectrogram. The window length sets a fixed trade-off between time and frequency resolution that applies uniformly across the spectrum. This is adequate for signals whose frequency content changes on a single timescale but is limiting for signals with both slow low-frequency rhythms and brief high-frequency transients.

10.3 Wavelets

The continuous wavelet transform decomposes a signal onto scaled and translated copies of a mother wavelet \( \psi \),

\[ W(a, b) = \frac{1}{\sqrt{a}} \int x(t)\, \psi^{*}\!\left(\frac{t - b}{a}\right) dt. \]

Small scales \( a \) yield short windows that localize high frequencies in time; large scales yield long windows that resolve low frequencies in frequency. This multiresolution behavior matches the structure of many biosignals and has made wavelets the method of choice for detecting transient events such as epileptiform spikes, ECG QRS complexes, and sleep spindles. The discrete wavelet transform (DWT), implemented by a cascade of quadrature mirror filters, provides an orthogonal multiresolution decomposition useful for denoising and compression. Wavelet denoising thresholds the detail coefficients, shrinking small coefficients presumed to be noise while preserving large coefficients presumed to be signal.

10.4 Hilbert-Huang and Empirical Mode Decomposition

The Hilbert-Huang transform decomposes a signal into intrinsic mode functions by empirical mode decomposition (EMD) and then computes instantaneous amplitude and frequency of each via the Hilbert transform. It is fully data-driven, does not assume stationarity, and has been applied to HRV, EEG, and gait analysis, though its theoretical foundations are less rigorous than those of the STFT or wavelet transform.

Chapter 11: Force, Pressure, and Position Sensors

11.1 Strain Gauges and Load Cells

A metal-foil strain gauge exploits the change in resistance of a thin conductor under strain,

\[ \frac{\Delta R}{R} = G_{F}\, \epsilon, \]

where \( G_{F} \) is the gauge factor (about 2 for metal foils, up to 100 for semiconductor gauges) and \( \epsilon \) is the strain. Four gauges arranged in a full Wheatstone bridge cancel temperature drift and double the sensitivity. Load cells for biomechanics (force plates in gait analysis, handgrip dynamometers, prosthetic socket sensors) are precision aluminum or steel elements with multiple bridges that resolve three-axis force and three-axis moment. The bridge output is typically a few millivolts per volt of excitation and requires an instrumentation amplifier followed by a high-resolution ADC.

11.2 Pressure Sensors

Physiological pressure measurement covers a wide range. Intravascular blood pressure is measured with fluid-filled catheters coupled to diaphragm-based strain-gauge transducers or, for higher-fidelity measurements, with tip-mounted piezoresistive or fiber-optic sensors. The frequency response of the fluid column must extend well beyond the heart rate; a resonant frequency above 20 Hz and critical damping are desired to reproduce the dicrotic notch. Airway pressure in respiratory monitoring uses similar diaphragm sensors with lower frequency response requirements.

Non-invasive blood-pressure measurement by oscillometry inflates a cuff, then deflates it while detecting pressure oscillations; the mean arterial pressure corresponds to the maximum oscillation amplitude, with systolic and diastolic estimated from fixed ratios. Continuous non-invasive methods such as volume-clamp (Penaz) finger cuffs and pulse-transit-time estimators extend this to beat-to-beat tracking.

11.3 Position, Velocity, and Acceleration

Joint and limb kinematics can be measured by optical marker tracking (Vicon, OptiTrack), inertial measurement units (IMUs), electromagnetic trackers, and goniometers. IMUs combine three-axis accelerometers, gyroscopes, and often magnetometers, fused by a Kalman or complementary filter to yield orientation. Accelerometer noise density and gyroscope drift determine attitude error growth, and the practical upper bound on integration time without position drift is seconds to minutes. Optical systems achieve sub-millimeter accuracy but are restricted to controlled laboratory volumes, while IMUs are wearable but prone to drift.

11.4 Piezoelectric and Capacitive Transducers

Piezoelectric materials (quartz, PVDF, PZT) produce charge proportional to strain and are natural choices for high-frequency force and acoustic measurements such as phonocardiography and swallowing assessment. They respond only to dynamic signals and require charge amplifiers with very high input impedance. Capacitive sensors, in which displacement changes the separation between plates, provide excellent low-noise performance and are used in MEMS accelerometers and in some pressure sensors.

Chapter 12: Applications

12.1 Sleep Staging

Polysomnography records EEG, EOG, EMG (typically chin and legs), ECG, respiratory effort, airflow, and oxygen saturation. Sleep technologists score 30-second epochs into wake, REM, and three non-REM stages (N1, N2, N3) according to AASM rules based on the presence of alpha, spindles, K-complexes, slow waves, and REM with low chin EMG. Automated sleep staging has matured from hand-crafted features into deep networks trained on large open datasets, and clinical-grade accuracy is now routinely achieved. Beyond stage scoring, sleep analytics quantify arousal indices, sleep-disordered breathing events, and periodic limb movements.

12.2 Brain-Computer Interfaces

A brain-computer interface (BCI) translates brain activity into commands for an external device. Non-invasive BCIs based on scalp EEG exploit several paradigms. Motor-imagery BCIs use the event-related desynchronization of mu and beta rhythms over sensorimotor cortex when subjects imagine limb movement, classified by common spatial patterns and linear discriminant analysis or by deep networks such as EEGNet. P300 spellers flash letters in rows and columns and detect the P300 response to attended targets. Steady-state visual evoked potentials (SSVEPs) flicker targets at distinct frequencies and detect the matching response in occipital EEG. Invasive BCIs based on electrocorticography or intracortical microelectrode arrays achieve much higher bandwidth and have enabled cursor control, robotic-arm control, and recently speech decoding in paralyzed patients.

12.3 Prosthetic Control

Myoelectric prostheses use surface EMG from residual musculature to control hand, wrist, and arm actuators. Direct control maps one or two EMG channels onto one or two degrees of freedom, requiring the user to co-contract and mode-switch to access additional movements. Pattern recognition control trains a classifier (linear discriminant analysis, support vector machines, or neural networks) on features (mean absolute value, waveform length, autoregressive coefficients, wavelet coefficients) extracted from multiple electrodes; at run time the classifier identifies the intended movement class and a proportional estimator modulates the speed. High-density EMG and EMG decomposition promise more natural control by exposing the underlying motor-unit discharges to the decoder. Targeted muscle reinnervation surgically relocates nerves to expose richer EMG signals and has extended myoelectric control to shoulder-disarticulation amputees.

12.4 Cardiac Monitoring and Arrhythmia Detection

Ambulatory ECG monitors (Holter, event, and patch recorders) record days to weeks of data, with classification performed offline or on-device. Modern arrhythmia detectors combine Pan-Tompkins-style QRS detection with rhythm classifiers trained on large annotated datasets such as the PhysioNet/Computing in Cardiology Challenge corpora. Detection of atrial fibrillation from RR-interval irregularity is the headline application for consumer wearables, and photoplethysmography on the wrist has extended this capability beyond ECG-equipped devices.

12.5 Respiratory and Metabolic Monitoring

Pulse oximetry estimates arterial oxygen saturation from the ratio of red and infrared absorption modulations with the cardiac cycle, a remarkable example of multi-wavelength spectroscopy combined with signal processing. Capnography tracks end-tidal CO2 in exhaled breath and provides real-time feedback on ventilation during anesthesia and resuscitation. Near-infrared spectroscopy extends the principle to tissue oxygenation in muscle and brain, with cortical hemodynamics forming the basis of functional near-infrared spectroscopy (fNIRS) BCIs.

12.6 A Closing Synthesis

The thread running through this course is that no stage of biomedical signal processing can be understood in isolation. Electrodes and amplifiers set the noise floor and the bandwidth; sampling and quantization set the resolution; spectral, time-frequency, and adaptive methods reveal structure that naive filtering would destroy; and application-specific knowledge of physiology determines which features matter. The practitioner who can move fluidly between these levels, choosing a filter in light of the volume conductor, an ADC in light of the electrode noise, and a classifier in light of the motor-unit physiology, is the one who builds biomedical instruments that actually work in the clinic and in the field.