6.4 Feature Extraction and Pattern Recognition

Feature Extraction from Biomedical Signals

Feature extraction and pattern recognition turn raw biomedical signals into something actionable. Instead of working with thousands of data points from an EEG, ECG, or EMG recording, you pull out a compact set of features that capture what matters, then use pattern recognition to classify or group those signals. This pipeline is at the core of automated diagnosis, patient monitoring, and brain-computer interfaces.

Time-Domain and Frequency-Domain Features

Time-domain features describe a signal's amplitude and how it behaves over time. You compute them directly from the raw signal values:

• Mean captures the signal's central tendency (its average value over a window).
• Variance quantifies how spread out the signal is around that mean. A high-variance ECG segment might indicate irregular cardiac activity.
• Skewness measures asymmetry in the amplitude distribution. Positive skewness means a longer tail to the right; negative skewness means a longer tail to the left.
• Kurtosis measures how peaked or flat the distribution is relative to a normal distribution. High kurtosis means sharp peaks (concentrated energy); low kurtosis means a flatter, more uniform spread.

These features are straightforward to compute and work well for characterizing overall signal shape in EEG and ECG analysis.

Frequency-domain features reveal what frequencies are present and how strong they are. You obtain them by transforming the signal out of the time domain:

• The Fourier transform decomposes a signal into its constituent frequency components, showing the magnitude of each frequency present.
• Power spectral density (PSD) analysis quantifies how signal power is distributed across frequency bands. For EEG, this lets you compare power in alpha (8–13 Hz), beta (13–30 Hz), and other clinically relevant bands.
• Common frequency-domain features include:
  • Spectral centroid: the "center of mass" of the frequency spectrum, indicating where most spectral energy is concentrated.
  • Spectral entropy: a measure of how complex or irregular the frequency distribution is. A flat spectrum (white noise) has high entropy; a single dominant frequency has low entropy.
  • Spectral flux: how much the frequency spectrum changes over time, useful for detecting transitions in EMG or EEG signals.

Wavelet-Based Features and Feature Selection

Frequency-domain analysis tells you what frequencies are present but not when they occur. Wavelet transforms solve this by providing a time-frequency representation, decomposing the signal into wavelet coefficients at different scales and positions.

• Wavelet coefficients capture localized patterns and transient events, making them ideal for signals with sudden changes (like QRS complexes in ECG or spike-wave discharges in EEG).
• Different wavelet families suit different signals. Haar wavelets are simple and good for detecting sharp transitions. Daubechies and Symlet wavelets offer smoother representations and are commonly used for EEG and ECG analysis.

Why extract features at all? Raw biomedical signals can have thousands of samples per second. Feeding all of that into a classifier is impractical and often counterproductive. Feature extraction reduces dimensionality while preserving the most discriminative information for the task at hand.

Choosing the right feature type depends on your signal and your goal:

• Time-domain features work well for overall amplitude and temporal properties (e.g., heart rate variability from ECG).
• Frequency-domain features are best for spectral content and dominant rhythms (e.g., EEG band power, EMG frequency analysis).
• Wavelet-based features excel at capturing localized time-frequency patterns and transient events across EEG, ECG, and EMG.

Dimensionality Reduction for Feature Sets

Even after feature extraction, you can end up with a large number of features. Dimensionality reduction trims this set down to the most informative components, which helps avoid overfitting, speeds up computation, and often improves classifier performance.

Principal Component Analysis (PCA)

PCA is a linear technique that finds new axes (principal components) along the directions of maximum variance in your feature space.

How PCA works, step by step:

1. Compute the covariance matrix of your feature set.
2. Calculate the eigenvectors and eigenvalues of that covariance matrix.
3. Rank the eigenvectors by their eigenvalues (largest first). Each eigenvector is a principal component, and its eigenvalue indicates how much variance it captures.
4. Project the original features onto the top-$k$ principal components to get a reduced representation (the principal component scores).

The key decision is how many components to keep. A common rule is to retain enough components to explain at least 95% of the total variance. You can visualize this with a cumulative explained variance plot or a scree plot, where you look for an "elbow" where additional components contribute diminishing returns.

Independent Component Analysis (ICA) and Dimensionality Reduction Benefits

ICA takes a different approach from PCA. Instead of maximizing variance, ICA finds a linear transformation that maximizes statistical independence between components, assuming the observed signal is a mixture of independent, non-Gaussian sources.

• In EEG processing, ICA is widely used to separate neural sources from artifacts (eye blinks, muscle activity, line noise). Each independent component ideally corresponds to a distinct physiological or artifactual source.
• ICA can reveal latent variables with meaningful interpretations, such as activity from different brain regions contributing to a scalp EEG recording.

Why dimensionality reduction matters for pattern recognition:

• Curse of dimensionality: As the number of features grows, the amount of training data needed for reliable classification grows exponentially. With limited patient data (common in biomedical applications), high-dimensional feature spaces lead to poor generalization.
• Computational efficiency: Fewer features mean faster training and prediction.
• Reduced overfitting: Eliminating redundant or noisy features helps models generalize to new patients and recording sessions.

To choose the optimal number of reduced dimensions, you can evaluate:

• Explained variance (for PCA): what fraction of total variance is retained.
• Reconstruction error: the difference between the original features and their reconstruction from the reduced set.
• Downstream classifier performance: sometimes the best check is whether reducing dimensions actually improves classification accuracy on a validation set.

Pattern Recognition in Biomedical Signals

Once you have a compact, informative feature set, pattern recognition algorithms learn to classify or group signals. The choice between supervised and unsupervised methods depends on whether you have labeled data.

Supervised Learning Algorithms

Supervised learning requires a labeled dataset where each signal segment is tagged with a known class (e.g., "normal sinus rhythm" vs. "atrial fibrillation"). The algorithm learns a mapping from features to labels, then predicts labels for new, unseen data.

Support Vector Machines (SVM) find the hyperplane that maximally separates classes in the feature space. For data that isn't linearly separable, kernel functions (polynomial, radial basis function) map features into a higher-dimensional space where a separating hyperplane exists.

Decision Trees recursively split the feature space based on the most informative features at each node. The result is a tree structure where each leaf corresponds to a class label. A major advantage is interpretability: you can trace exactly which features and thresholds led to a decision.

Artificial Neural Networks (ANN) consist of layers of interconnected neurons. Input features propagate forward through hidden layers, and the network learns its weights via backpropagation. ANNs can model complex, non-linear relationships but require more data and computational resources than SVM or decision trees.

Typical supervised classification tasks in biomedical signal processing:

• Detecting arrhythmias from ECG features
• Classifying EEG into sleep stages or seizure vs. non-seizure
• Recognizing intended hand gestures from EMG signals

Unsupervised Learning Algorithms and Algorithm Selection

When you don't have labeled data, unsupervised learning discovers structure on its own.

K-means clustering partitions data into $k$ clusters by iteratively assigning each instance to the nearest cluster centroid, then updating centroids based on the new assignments. You need to specify $k$ in advance.

Hierarchical clustering builds a tree-like structure (dendrogram) representing nested groupings based on similarity. You can cut the dendrogram at different levels to get different numbers of clusters, which is useful for exploratory analysis.

Unsupervised methods are valuable for:

• Identifying patient subgroups with similar signal characteristics
• Detecting anomalies or outliers in biomedical recordings
• Exploring data when clinical labels are unavailable or uncertain

Choosing the right algorithm comes down to a few key factors:

Performance Evaluation of Feature Extraction and Pattern Recognition

Building a model is only half the job. You need rigorous evaluation to know whether your feature extraction and classification pipeline actually works on data it hasn't seen before.

Evaluation Metrics

Accuracy is the ratio of correctly classified instances to total instances. It gives a quick overall picture, but it can be misleading with imbalanced classes. If 95% of ECG segments are normal, a model that always predicts "normal" gets 95% accuracy while missing every abnormal case.

Sensitivity (also called recall or true positive rate) is the proportion of actual positives correctly identified. In medical diagnostics, high sensitivity is critical because missing a true positive (e.g., failing to detect a seizure) can have serious consequences.

Specificity (true negative rate) is the proportion of actual negatives correctly identified. High specificity reduces false alarms and unnecessary interventions.

There's often a trade-off between sensitivity and specificity. Adjusting the classification threshold increases one at the expense of the other.

Additional Evaluation Techniques

Beyond accuracy, sensitivity, and specificity, several other metrics give a fuller picture:

• Precision: of all instances predicted positive, how many actually are? Precision = $\frac{\text{True Positives}}{\text{True Positives} + \text{False Positives}}$. High precision means few false alarms.
• F1-score: the harmonic mean of precision and recall, $F1 = 2 \cdot \frac{\text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}}$. Particularly useful when classes are imbalanced.
• AUC-ROC: the area under the receiver operating characteristic curve, which plots true positive rate vs. false positive rate across all classification thresholds. An AUC of 1.0 means perfect discrimination; 0.5 means no better than random guessing.

Cross-validation provides more reliable performance estimates than a single train/test split:

1. K-fold cross-validation: Split the data into $k$ equally sized folds. Train on $k-1$ folds, validate on the remaining fold. Repeat $k$ times so every fold serves as the validation set once. Average the metrics across all folds.
2. Leave-one-out cross-validation (LOOCV): A special case where $k$ equals the number of instances. Each instance gets its own turn as the validation set. This is thorough but computationally expensive for large datasets.

Cross-validation is especially important in biomedical applications where datasets tend to be small. It gives you a more honest estimate of how the model will perform on new patients.

Performance evaluation also guides practical decisions: comparing different feature sets, tuning hyperparameters (number of PCA components, SVM kernel type, number of clusters in k-means), and confirming that the full pipeline generalizes beyond the training data.