4 Biosignal Adaptive Systems

• Adaptive: the system initiates/controls adaptations (user has low/no control)

• Adaptable: the user initiates/controls adaptations (user customizes, system does not initiate)

Personalization the process of changing the functionality, interface, information access and content of a system to increase its personal relevance to a user.

• There are 2 dimensions of personalization:

  • What is personalized? (functionality, interface, content…)

  • How is the personalization triggered? (user driven, system driven or hybrid)

• Hypertexts → text with clickable links to other texts (wikipedia)

• Hypermedia → extends hypertexts with multimedia (images, video, audio).

• Adaptive Hypermedia Systems → a hypermedia system that watches the user and adjusts itself to fit the user’s personality.

• Context: is any information about the situation of a user that is relevant for the interaction between the user and an application (individuals, activities, time, technologies, physical world…)

• Context-aware systems: are systems that use context to provide relevant information or services to the user (relevancy depends on user’s task).

Devices (e.g. mobile devices) can acquire, manage and process information from raw sensor data.

1. Basic features are extracted from sensors (e.g. quick movement).

2. Preprocessing is applied to derive higher level sensor information (e.g. running)

3. Classification is performed to inform the user context (e.g. jogging)

4. Applications are updated (e.g. with daily or weekly information)

• It’s a system that uses signals and patterns to recognize emotions.

• There are

  • Visible input techniques (facial activity, eye, posture, gesture, vocal and textual activity)

  • Invisible input techniques (physiological data)

• Biocybernetic loop:

  • It’s a closed-loop concept for physiological computing systems.

  • It’s composed of 3 stages of real-time data processing: collection → analysis → translation.

• Biocybernetic adaptation:

  • New take on the traditional loop, where physiological data is used to adapt the system to the user.

  • Circle: Sensor data → state prediction → adaptation.

• A brain computer interface is a system that reads the brain activity and turns it into commands or messages. User does something → brain signal gets picked up → signal gets translated into action.

• A passive brain computer is the same, but the user is not actively trying to control anything. The system quietly reads the brain activity and provides continuous information in real time about the user’s cognitive or emotional state.

• Biosignal adaptive systems are systems that collect and process biosignals from users and their environment in order to perform adaptations.

• Biosignal recording is individual and influenced by personal factors:

  • Environmental factors (work-related)

  • Non modifiable factors (age, gender, ethnicity)

  • Lifestyle factors (posture, tabaco, alcohol)

  • Neurophysiological factors (emotions, personality, mood)

  • Physiological and pathological factors (asthma, diabetes, brain damage)

What is adapted and personalized in the system.

• Modification of function allocation → Who? → Task sharing, task offloading.

• Modification of task scheduling → When? → Timing, prioritization, duration.

• Modification of interaction → How? → Style, amount, interface features

• Modification of content → What? → Quality, quantity, abstraction

When and how are behaviors or properties or the system adapted? How is the personalization triggered?

→ When to start the adaptation? For how long? And when to stop it?

• Operator triggers → operator initiated, operator measurement

• System triggers → state, mode

• Environment triggers → state, event

• Task/Mission triggers → task status, mission event

• Spatio-Temporal triggers → location, time

• Benefits

  • Increased awareness and individual development

  • Protect humans with adaptive technology

  • Create positive impact on well-being and productivity

• Pitfalls

  • Reduce privacy and security

  • Enable surveillance

  • Loose self-assessment abilities

  • Manipulate humans

  • Create unexpected backfire effects

• Adaptive systems are not directly user-controlled

• The adaptation process often takes time

• The effects of adaptation depend on the user and the context

• Formative evaluation → identify shortcomings or errors in a system in order to improve it and to guide the overall design and development

• Summative evaluation → determine the value or impact of an entire system

• It’s a step wise evaluation of the adaptation that provides insight into the different phases:

  • collection of input data

  • interpretation of the collected data

  • modeling of the current state of the world

  • deciding upon adaptions

  • applying the adaptations

Biosignal modeling can target psychological and physical user characteristics, organized into three temporal categories:

• Short-term (changes quickly, in the moment): Emotion, Attention, Perception, Flow

• Evolving over time (change over weeks/months): Mood, Memory, Attitude

• Long-term (stable, deeply embedded): Personality, Culture, Value

CRISP-DM is a freely available analytical modeling process. Its six phases are:

1. Business Understanding — define the goal

2. Data Understanding — explore available data

3. Data Preparation — clean and transform the data

4. Modeling — select and train models

5. Evaluation — assess whether the model meets the business goal

6. Deployment — roll out the model into production

• Traditional ML: Features are manually selected and identified by domain experts before training.

• Deep Learning: High-level features are automatically learned from huge amounts of raw data in an incremental, layered manner — no manual feature engineering needed.

• requires less data

• requires less computing power

• offers better explainability

Both are data preparation activities, but they operate at different levels:

• Feature Engineering = vertical (column-level) operations — transforming or creating entire features/columns (e.g., encoding a "gender" column)

• Data Manipulation = horizontal (instance/row-level) operations — modifying individual data points (e.g., removing an outlier row, filling in a missing value)

Feature Engineering techniques:

1. Encoding of categorical data

2. Bucketing/Binning of numerical (continuous) data

3. Scaling numerical data (normalization or standardization)

4. Feature Creation (deriving new features from existing ones)

Data Manipulation techniques:

1. Outlier Identification — finding and handling extreme values

2. Imputation — filling in missing values

3. Balancing — fixing class imbalances in the dataset

The goal is to transform continuous numerical data into categorical (discrete) variables by grouping values into ranges called bins.

Example: Instead of using exact ages (13, 20, 23, 29...), you group them into: ≤20, 20–30, 30–40, >40.

Why use it?

• Some models handle categorical data better than continuous data

• Reduces the effect of small fluctuations or noise

• Can make patterns more visible and models more robust

Normalization (Min-Max Scaling):

• Rescales values to the range [0, 1], by comparing the current value against the highest and lowest values and finding the average.

• Formula: X_norm = (X - X_min) / (X_max - X_min)

• Use when: you know the bounds of your data, no extreme outliers

Standardization (Z-Score):

• Takes the value and substracts the mean to find the difference, then divides it by the standard deviation.

• Formula: z = (x - mean) / std.dev

• Use when: data roughly follows a normal distribution, or outliers are present

→ Important because ML algorithms are sensitive to different scales — without scaling, a feature with values in the thousands could dominate one with values between 0 and 1.

1. Sorting Method — Sort the data for each feature and scan for extremely low or extremely high values.

2. Statistical Methods:

   • Z-Score: Set a threshold for the Z-score (e.g., Z > 3); data points exceeding it are considered outliers

   • IQR (Interquartile Range): Calculate the range of the middle 50% of data, create upper and lower "fences," and flag anything outside as an outlier

3. Visualization — Create a box plot for a feature; points outside the whiskers are outliers

Imputation is the process of filling in missing values in a dataset rather than deleting entire rows or columns, which would lose information.

Two main strategies:

• Categorical Imputation: Replace missing categorical values with the most frequently occurring value in that column

• Numerical Imputation: Replace missing numerical values with the mean of that column

• A class imbalance occurs when one class (the majority class) makes up a disproportionately large portion of the training data, while other classes (the minority class) are underrepresented.

• Example: a fraud detection dataset with 99% normal transactions and 1% fraudulent ones.

• The problem: a model trained on such data learns it can achieve high accuracy by always predicting the majority class — without ever actually learning to detect the minority class. This makes the model useless for the cases you actually care about most.

1. Resampling the training set:

   1. Undersampling — Remove instances from the majority class until both classes are equal. Risk: you lose potentially useful data.

   2. Oversampling — Add more instances to the minority class until both are equal. Risk: overfitting, since you're often just duplicating existing samples.

2. k-Fold Cross-Validation

3. Ensemble Method with different resampled datasets — Train multiple models on differently resampled versions and combine them.

4. Resample with different ratios — Experiment with different class distribution targets (not necessarily 50/50).

Labeling is the process giving ground-truth class labels to some data instances, so that a supervised ML model can learn from them.

• In biosignal modeling, this means tagging a segment of physiological data with the user state that was occurring at that time.

Challenging: because costly, error-prone, and labor-intensive — often requires domain expertise and can frustrate users (boring?).

Interactive Labeling Systems are Interactive ML systems where users iteratively build and refine a model through cycles of labeling data and reviewing the model's suggestions.

• The model assists the human, and the human corrects the model — making the labeling process more efficient over time.

The Experience Sampling Method (ESM) — is a research method where participants are randomly interrupted during their daily life and asked to answer brief questions about their current psychological state (thoughts, feelings, behaviors, environment).

Key features:

• Short time reference (e.g., last 5 minutes) to avoid recall bias

• Psychologically well-established procedure

ESM is used alongside biosignal recording to capture self-reported psychological states (e.g., "how focused are you right now?") at the same timestamp as the physiological data. This creates labeled biosignal datasets — the labels from the survey + the biosignal data at that moment = training data for an ML classifier.

Hyperparameter tuning is the process of finding the best combination of hyperparameters — it happens before/during training and uses the validation dataset to compare configurations.

1. Training Dataset — the data the model learns from; used to fit the model parameters. Typically ~80% of the training portion.

2. Validation Dataset — used to evaluate the model while tuning hyperparameters; gives an unbiased evaluation of different model configurations. Typically ~20% of the training portion.

3. Test Dataset — held out completely until the very end; used for a final, unbiased evaluation of the finished model. Never used during training or tuning.

The key rule: the test set must never influence model development.

1. The Hyperparameter Tuner proposes a set of hyperparameters

2. The model is trained on the Training Dataset using those hyperparameters

3. The trained model is evaluated on both the Training Dataset (to check for underfitting) and the Validation Dataset(to check for overfitting/generalization)

4. The validation evaluation results are fed back to the Hyperparameter Tuner

5. The tuner proposes new hyperparameters and the loop repeats

6. The best configuration found is then evaluated once on the Test Dataset

Cross-Validation is a technique to get a better estimate of model performance on unseen data by rotating which portion of the training data is used as the validation set.

Three main strategies:

1. k-Fold Cross-Validation: Split training data into k equal parts (folds). Train on k-1 folds, validate on the remaining 1 fold. Repeat k times (each fold serves as validation once). Average the results. → Good general approach.

2. Stratified k-Fold: Same as k-fold, but ensures each fold contains the same class distribution as the full dataset. → Better for imbalanced datasets.

3. Leave-One-Out: Each single data point is used once as the validation set. Computationally very expensive but useful for very small datasets.

• Problem: N-dimensional space where each dimension represents a hyperparameter and each point represents one model configuration (highly complex with lots of different combinations.

• Hyperparameter tuning is about searching through the multi-dimensional space of hyperparameter combinations to find the best configuration:

  • Grid Search: Define a grid of all possible hyperparameter value combinations and evaluate every single point in the grid. Exhaustive but very slow — gets worse exponentially as you add more hyperparameters.

  • Random Search: Define a bounded search space and randomly sample combinations from it. Faster and often surprisingly effective — frequently finds good configurations with far fewer evaluations than Grid Search.

Ensemble Methods train multiple models to solve the same problem and combine their outputs to get better results than any single model could achieve.

Two main meta-algorithms:

1. Bagging (e.g., Random Forest): Train multiple homogeneous models independently on different random subsets of data, then combine results through averaging or majority vote. Reduces variance/overfitting.

2. Boosting (e.g., AdaBoost, XGBoost): Build models sequentially, where each new model focuses on the mistakes of the previous one. They build on each other in an adaptive way. Reduces bias.

Analytical models must meet functional requirements → perform the task stakeholders expect.

• There is also a range of non-functional requirements that probably should be fulfilled:

  • Accuracy — correct predictions

  • Fairness — no bias against demographic groups

  • Explainability — decisions should be understandable to humans

  • Robustness — stable performance under noisy or shifted data

  • Security/Privacy — handles sensitive data safely

  • Energy Consumption — computationally efficient

Model Evaluation and XAI are applied continuously — not just at the end — throughout the modeling process:

• A Created Model and a Deployed Model are both subjects of Continuous Model Evaluation

• Both the Data Scientist/Analyst and other users interact with the model through XAI and Model Evaluation

• XAI and Model Evaluation thus serve as a feedback loop between the model and the people using or developing it

These four outcomes capture every possible result of a classification prediction:

• TP (True Positive) — correctly detected targets

• TN (True Negative) — correctly found non-targets

• FP (False Positive) — wrongly found targets (falsely flagged as target)

• FN (False Negative) — wrongly found non-targets (missed a real target)

XAI methods can be:

• model specific — only works for one type of model

• model agnostic — works for any model regardless of model type

And they can be:

• global — explains the behavior of the whole model

• local — explains a single prediction

• Partial Dependence Plots — shows the marginal effect one or two features have on the predicted outcome of a machine learning model

• Permutation Feature Importance — measures the increase in prediction error of the model after the feature's values have been permuted (shuffled). A large error increase = the feature is important.

• LIME — approximates the predictions of the underlying black box model to explain individual predictions locally

• SHAP — determines the positive or negative impact of each feature on an individual prediction

A dashboard is a visual display of the most important information needed; consolidated and organized on a single screen so the information can be monitored at a glance.

Key characteristics:

• Like a car's dashboard, BI&A provides decision makers with a "managerial cockpit" to "drive" the business

• Primary data source is a data warehouse, from which data is extracted through preprocessing and transformation steps