⚡ electrical engineering

Physiologically Grounded Driver Behavior Classification: SHAP-Driven Elite Feature Selection and Hybrid Gradient Boosting for Multimodal Physiological Signals

This study proposes an interpretable, multimodal framework that combines SHAP-driven feature selection, Bayesian-optimized gradient boosting models, and a weighted ensemble to achieve 80.91% accuracy in classifying driver behaviors from synchronized EEG, EMG, and GSR signals, demonstrating significant performance gains over single-modality approaches.

Original authors: Sahar Askari, Mohammad Mahdi Mirza Ali Mohammadi, Fatemeh Ensafdoust, Amin Golnari, Saeid Sanei

Published 2026-05-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sahar Askari, Mohammad Mahdi Mirza Ali Mohammadi, Fatemeh Ensafdoust, Amin Golnari, Saeid Sanei

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to figure out what a driver is about to do next—like hitting the brakes, turning the wheel, or speeding up. Usually, we look at the car: is it slowing down? Is the steering wheel moving? But this paper suggests a smarter way: listen to the driver's body before the car even moves.

The researchers built a "digital detective" that reads three different types of biological signals from a driver's body to guess their next move. Here is how they did it, explained simply:

1. The Three "Sensors" (The Team)

Instead of just one camera or one sensor, the team used a trio of biological "spies" working together:

The Brain (EEG): Like a high-speed microphone picking up the driver's thoughts and focus. It tells us what the driver is planning to do.
The Muscles (EMG): Like a tension meter on the driver's arms and legs. It detects the physical twitch or preparation to move a pedal or turn a wheel.
The Sweat (GSR): Like a stress alarm. It measures how "jumpy" or aroused the driver is (e.g., getting ready for a sudden stop).

The Analogy: Imagine trying to guess if a friend is about to jump off a diving board.

Looking at the pool (the car) tells you they might jump.
Listening to their brain (EEG) tells you they are thinking about jumping.
Feeling their muscles (EMG) tells you their legs are tensing up.
Checking their sweat (GSR) tells you they are excited or nervous.
The paper argues that you need all four pieces of information to be 100% sure.

2. Cleaning the Messy Data (Preprocessing)

Real-life signals are messy. A driver's head moves, they blink, and electrical noise interferes.

The Filter: The researchers used a sophisticated "noise-canceling headphone" technique (called ICA) to strip away the static and keep only the pure brain and muscle signals.
The Window: They didn't look at the whole day; they looked at tiny 2-second "snippets" right before and after a driving action, like looking at a photo right before a sneeze.

3. Picking the Best Clues (Feature Selection)

The system initially found over 500 different clues (features) in the data. That's too many for a computer to process quickly, and it might get confused by useless noise.

The Detective's Notebook (SHAP): They used a smart tool called SHAP to ask, "Which clues actually matter?"
The Cut: They threw away the weak clues and kept only the top 250 most important ones. This is like a detective ignoring 500 irrelevant witnesses and focusing only on the 250 who saw the crime.

4. The Brain Trust (The Ensemble)

Instead of using one computer program to make the guess, they built a "committee" of two powerful AI models: XGBoost and LightGBM.

The Strategy: Think of these as two expert judges. One is great at spotting patterns in complex data, and the other is great at speed and handling large lists.
The Vote: They don't just let one judge decide. They take a weighted vote. In this case, the "LightGBM" judge had a slightly louder voice (65% of the vote), while the "XGBoost" judge had 35%. Together, they reached a consensus.

5. The Results: Did It Work?

The team tested this system on a massive dataset of 30 drivers performing four actions: Braking, Changing lanes, Accelerating, and Turning.

The Score: The team got it right 80.91% of the time.
The Comparison:
- If they used only the brain signals (EEG), they got about 73% right.
- If they used only the muscle signals, they got about 60% right.
- If they used only the sweat signals, they got about 40% right.
- The Lesson: By combining all three, the system became significantly smarter. It's like a puzzle: you can't solve it with just one piece.

6. Why It Makes Sense (Interpretability)

One of the biggest problems with AI is that it's a "black box"—you don't know why it made a decision. This paper fixed that.

The Proof: The system explained its own logic. It showed that when predicting a Brake, it relied heavily on the "stress alarm" (GSR) and brain waves. When predicting a Turn, it looked closely at the "muscle tension" (EMG) to see if the driver was leaning left or right.
The Takeaway: The AI isn't just guessing; it's using real, biological logic that matches how human bodies actually work.

Summary

This paper presents a new, transparent way to predict what a driver is about to do by listening to their brain, muscles, and stress levels all at once. By cleaning the data, picking the best clues, and letting two smart AI models vote together, they created a system that is much more accurate than looking at just one signal or using older, simpler computer methods. It proves that to understand a driver's mind, you need to listen to their whole body.

Technical Summary: Physiologically Grounded Driver Behavior Classification

Problem Statement
Road traffic injuries remain a critical global public health concern, with approximately 94% of crashes attributed to human factors such as distraction, fatigue, and inattention. While conventional driver-monitoring systems rely on vehicle kinematics (e.g., steering variability, lane deviation) or vision-based techniques (e.g., eye tracking), these methods often capture the outcomes of cognitive failures rather than the underlying physiological processes. Although electroencephalography (EEG) offers millisecond-level temporal resolution of neural activity, it suffers from limited spatial resolution and susceptibility to noise in real-world environments. Furthermore, deep learning approaches, while promising, often lack interpretability and require extensive labeled data, hindering their deployment in safety-critical, resource-constrained scenarios. There is a need for an interpretable, scalable framework that leverages multimodal physiological signals (EEG, electromyography [EMG], and galvanic skin response [GSR]) to decode driving behaviors with high accuracy and transparency.

Methodology
The study proposes a framework utilizing the Multimodal Physiological Driving Behavior (MPDB) dataset, which contains synchronized 64-channel EEG, 4-channel EMG, and single-channel GSR signals from 30 participants performing simulated driving tasks. The methodology consists of four primary stages:

Preprocessing and Synchronization:
- Signal Processing: Raw signals (originally 1000 Hz) were downsampled to 500 Hz. EEG data underwent band-pass filtering (0.5–40 Hz), artifact removal via Independent Component Analysis (ICA) using the Infomax algorithm, and exclusion of non-cerebral channels, resulting in 59 usable EEG channels. EMG signals were filtered (20–240 Hz) with notch filtering to remove power-line interference, while GSR was filtered (0.1–35 Hz).
- Segmentation: Events were extracted using a 2.0-second window (from $t = -0.5$ s to $t = +1.5$ s relative to event onset).
- Fusion: Multimodal alignment was achieved via hybrid cross-modality matching with strict Event-ID verification, followed by baseline correction and channel-axis concatenation.
Feature Extraction:
A domain-specific pipeline extracted 503 initial features, categorized into:
- Time-Domain: Line length (LL), variance of derivatives, and maximum absolute change.
- Frequency-Domain: Power Spectral Density (PSD) estimated via Welch's method across specific bands (e.g., Delta, Theta, Alpha, Beta, Gamma for EEG; Low, Mid, High for EMG).
- Derived Indices: Global indicators such as the EEG Alpha/Theta ratio and EMG asymmetry index.
SHAP-Based Elite Feature Selection:
To address high dimensionality and prevent overfitting, a SHAP (SHapley Additive exPlanations) based strategy was employed. An auxiliary XGBoost model computed global feature importance across all samples and classes. The top 250 features were retained, reducing dimensionality by approximately 50% while preserving 81% of the model's predictive power. This selection was performed exclusively on the training data to prevent information leakage.
Hybrid Ensemble Learning:
The final classification utilized a weighted soft-voting ensemble of XGBoost and LightGBM models. Hyperparameters for both models were optimized using Bayesian optimization (via the Optuna framework) with the macro-F1 score as the objective. The optimal ensemble weighting was determined to be $\alpha = 0.35$ for XGBoost and $0.65$ for LightGBM. Class imbalance was addressed using balanced sample weights inversely proportional to class frequency.

Key Contributions

Robust Pipeline: Development of a preprocessing and feature extraction pipeline incorporating ICA-based artifact removal, multimodal synchronization, and streaming processing for the MPDB dataset.
SHAP-Driven Selection: Implementation of a SHAP-based elite feature selection approach that significantly reduces dimensionality while enhancing model efficiency and interpretability.
Hybrid Ensemble Framework: Creation of a state-of-the-art ensemble combining XGBoost and LightGBM, optimized via Bayesian methods, specifically tailored for multimodal physiological data.
Interpretability: Comprehensive analysis using SHAP values to validate the physiological plausibility of the model, linking feature importance to established neurophysiological markers.

Results
The proposed framework was evaluated on a held-out test set ( $n = 3,577$ ) across four driving behavior classes: Brake, Change (Lane Change), Throttle, and Turn.

Performance: The ensemble achieved a test accuracy of 80.91% and a macro-F1 score of 0.79.
Comparison: This significantly outperformed traditional machine learning baselines, including Random Forest (72.99% accuracy), KNN (64.91%), SVM (53.56%), and Logistic Regression (56.05%).
Ablation Studies:
- Multimodal Necessity: The full multimodal approach yielded an 8% accuracy improvement over the best single-modality model (EEG-only, 72.85%).
- Modality Synergy: Pairwise fusion of EEG and EMG achieved 76.15% accuracy, indicating strong cognitive-motor synergy, though the full fusion (including GSR) provided an additional 4.7% gain.
- Class Performance: The model showed highest reliability for Brake (F1: 0.83) and Turn (F1: 0.81) behaviors. The Throttle class was the most challenging (F1: 0.69), likely due to subtler physiological manifestations compared to abrupt maneuvers.
Interpretability: SHAP analysis revealed that EEG contributed the majority of predictive weight (75.2%), followed by EMG (19.4%) and GSR (5.4%). However, specific high-impact features included GSR phasic components (critical for braking) and EMG asymmetry (critical for turning), validating the complementary nature of the modalities.

Significance
The paper claims that this work offers a practical and transparent alternative to end-to-end deep learning for multimodal driver behavior recognition. By combining physiologically informed feature engineering with explainable tree-based ensemble learning, the framework achieves high accuracy without the "black-box" limitations of deep neural networks. The results validate that multimodal fusion is essential for robust driver monitoring, as no single physiological stream is sufficient to capture the complexity of dynamic driving events. The study establishes a new state-of-the-art baseline on the MPDB dataset, demonstrating that carefully engineered features combined with optimized gradient boosting can rival or exceed complex deep learning architectures while providing critical interpretability for safety-critical applications.