Biosignals: Dynamic biometrics
“Physiological phenomenon, a body variable that can be measured and monitored. Since the number of physiological mechanisms is nearly unlimited, the diversity of biosignals is huge.”
Core applications of biosignal processing:
❏ Monitoring
❏ Diagnosis
❏ Prognosis
❏ Identification (info in a device-system)
❏ BCI / Neuroprosthetic Control
MONITORING
Continuous or periodic measurement of physiological signals to observe the real-time status of a person or biological system.
➢ Goal: Detect trends, variations, or abnormal events as they occur (in real time). Examples:
💓 ECG: Tracking heart rate and rhythm in real time (e.g., Holter).
🧠 EEG: Observing brain activity for seizure detection.
💪 EMG: Measuring muscle activation during rehabilitation.
DIAGNOSIS
Use of biosignals to detect and classify diseases or dysfunctions by recognizing specific patterns.
➢ Goal: Determine what is wrong based on signal abnormalities Examples:
💓 ECG: Detecting arrhythmias, ischemia, or myocardial infarction.
🧠 EEG: Identifying epileptic activity or sleep disorders.
💪 EMG: Diagnosing neuromuscular disorders.
PROGNOSIS
Prediction of the future evolution or outcome of a disease or physiological state using biosignal analysis, patient data and models.
➢ Goal: Estimate what will happen next or the likely outcome. Examples:
💓 ECG: Predicting risk of cardiac arrest or arrhythmia recurrence.
🧠 EEG: Forecasting epileptic seizure.
💪 Gait signals: Estimating fall risk in elderly patients.
Identification (information in a device-system)
Process of recognizing or verifying the identity of a subject or system using information extracted from biosignals or device data.
➢ Goal: Ensure that the signal is correctly assigned to its origin (person or sensor). Examples:
👤 Biometric ID using ECG or EEG (unique signal patterns).
📌 Sensor identification: Detecting which electrode or device collected the signal.
🔐 User authentication in wearable systems.
BCI / Neuroprosthetic Control
Use of physiological signals to control external devices or systems directly from neural or
muscular activity — bypassing damaged neural pathways or enabling hands-free control.
➢ Goal: To translate biological signals into commands for real-time control of assistive or robotic devices — restoring or augmenting motor function and communication.
Biosignals processing → Translates physiological activity into actionable commands
It bridges neural intent → device control through algorithms like:
Feature extraction / Classification / Real-time decoding and feedback loops Biosignals Used:
🧠 EEG: brain activity for motor imagery or intent detection.
💪 EMG: muscle activation signals to control prosthetic movements.
🧠 ECoG / LFPs: higher-resolution brain signals for invasive BCIs.
BCI / Neuroprosthetic Control: Processing Steps:
Processing Steps:
1. Signal acquisition and preprocessing (filtering, artifact removal).
2. Feature extraction (e.g., frequency bands).
3. Classification / decoding (detecting user intent).
4. Command translation to control devices.
5. Feedback loop for adaptation and learning.
BCI / Neuroprosthetic Control - Processing Steps:
Signal acquisition and preprocessing
Signal acquisition and preprocessing
Acquisition: Capture clean and reliable biosignals from the body or brain. Data are digitized and sent to the processing unit.
○ Preprocessing: Removes unwanted components (artifacts) such as eye blinks, or electrical noise.
→ Common techniques: filtering (band-pass, notch), baseline correction, Independent Component Analysis (ICA).
Result: a cleaner signal that represents true neural or muscular activity.
BCI / Neuroprosthetic Control - Processing Steps: 2. Feature Extraction
- Feature Extraction
Goal → This step transforms raw data into measurable variables that can be classified. Examples:
- EEG: Extract power in specific frequency bands (e.g., mu and beta rhythms for motor imagery).
- EMG: Compute the envelope or root-mean-square (RMS) of muscle activity to quantify contraction strength.
BCI / Neuroprosthetic Control - Processing Steps: 3. Classification / Decoding
Goal → Interpret the extracted features to determine the user’s intention. Methods:
- Machine learning algorithms (e.g., SVM, LDA, neural networks) are trained to recognize patterns associated with specific actions or thoughts.
- Example: distinguishing between “move left,” “move right,” or “no movement.” The classifier outputs a decision or probability representing the detected command.
Machine learning algorithms:
Biosignals are complex, variable, and noisy — they differ across
people, sessions, and even within a single recording.
→ Traditional rule-based systems struggle to capture these patterns.
Machine learning allows systems to:
● Automatically detect patterns in complex, multidimensional signals.
● Adapt to individual users (e.g., specific EEG patterns during motor imagery).
● Generalize across time and conditions (e.g., detect “move left” even if the signal slightly
changes).
● Continuously improve with feedback or training data.
Machine learning
The process usually follows these main stages:
Training phase:
○ The system learns from labeled data (e.g., EEG segments tagged as “left hand,”
“right hand,” or “rest”).
○ It finds mathematical patterns (features) that best separate these categories
Testing phase:
○ The trained model receives new (unlabeled) data in real time.
○ It predicts the most likely class — i.e., what the user intends to do.
○ This approach is essential for decoding user intention in BCIs and
neuroprosthetics.
There are different classification algorithms → One algorithm cannot solve all problems! Choosing an algorithm depends on its speed, accuracy, training time, amount of data
required, interpretation of the trained model, etc.
Linear Discriminant Analysis (LDA)
Linear Discriminant Analysis (LDA): Supervised learning approach mainly used for multi-class classification problems.
→ LDA performs dimensionality reduction while preserving class-discriminative information.
In summary, when doing LDA-based multiclass supervised classification:
1. We build a training set of labelled observations for each class
2. We use those training sets to estimate the class parameters
3. We classify a new observation with the maximum discriminant function
Support Vector Machines (SVM)
Support Vector Machines (SVM): Supervised learning methods used for classification, regression, and outlier detection.
→ Main goal of SVM is to build a hyperplane (or a set of hyperplanes) in a high-dimensional space that best separates the data classes.
A good separation is achieved by the hyperplane that has the largest margin (distance) from the nearest data points of any class.
Random Forest (RF)
Random Forest (RF): machine learning algorithm that combines the outputs of multiple decision trees to produce a single result.
Due to its flexibility and robustness, it is used for both classification and regression tasks.
😊 Reduces the overfitting in decision trees 😓 Longer training period and complexity
😊 Can automatically handle missing values. 😓 Random forests do not generalize well with
completely new data
BCI / Neuroprosthetic Control - Processing Steps: 4. Command Translation
Goal → Convert the decoded intention into a real, actionable command for an external device.
This is where the biosignal becomes functional control of a robotic or assistive system.
BCI / Neuroprosthetic Control - Processing Steps: 5. Feedback Loop (Adaptation & Learning)
Goal → Improve performance through continuous interaction and adaptation.
This feedback loop enables closed-loop control, enhancing precision, accuracy, and usability over time.
