Fusion4Sleep

🧠 Multimodal Fusion Framework for Accurate Insomnia Detection

📄 Abstract

Insomnia, a prevalent sleep disorder, adversely impacts cognitive and physical health, highlighting the need for accurate and efficient diagnostic tools. This study employs multimodal biosignal analysis, integrating EEG, EMG, EOG, and ECG to classify individuals as insomniac or non-insomniac. Preprocessing techniques, such as resampling, are applied to reduce computational load, followed by feature extraction methods tailored to each signal modality. These methods capture frequency-domain, time-domain, and entropy metrics, which reveal physiological traits associated with insomnia, such as irregular brain activity, altered muscle atonia, and abnormal eye movements.

Early and late fusion techniques are employed for the classification task—early fusion achieved a better accuracy of 95%, compared to 92.275% for late fusion, with the highest overall F1 score of 95%. Explainable AI (XAI) concepts such as LIME (Local Interpretable Model-Agnostic Explanations) and SHAP (Shapley Additive Explanations) were used to determine the most influential features in classification. By leveraging multimodal data fusion, this methodology captures intricate physiological interactions to improve disease diagnosis and provides a non-invasive, interpretable detection of insomnia—paving the way for enhanced therapeutic interventions.

---

📁 Repository Structure

File/Folder	Description
BSP_1.ipynb	Baseline models for individual signal modalities
BSP_1_EarlyFusion.ipynb	Implementation of Early Fusion strategy
BSP_1_LateFusion.ipynb	Implementation of Late Fusion strategy
BSP-ATT.ipynb	Attention-based fusion modeling
Norm_BSP_1_EarlyFusion.ipynb	Normalized Early Fusion model
BSP_XAI.ipynb	Explainable AI using LIME and SHAP
combined_data.xlsx	Dataset with extracted features
csv_output/	Folder for result/output files
README.md	Project documentation

---

🧪 Dataset Description

The dataset used is sourced from the Sleep Disorders Research Center (SDRC). It contains recordings of 22 subjects (11 normal + 11 insomniac), sampled at 256 Hz, including:

• EEG (14 channels)
• EOG (6 channels)
• EMG (3 channels)
• ECG (1 channel)

Data was downsampled to 32 Hz and segmented into 60-second overlapping epochs for efficient processing and analysis.

---

🔬 Methodology

🔄 Data Preprocessing

• Downsampling to 32 Hz
• Epoch segmentation (60s windows with 10s overlap)

📊 Feature Extraction

Each modality has tailored feature extraction methods:

• EEG: Wavelet Coefficients, Spectral Entropy, Sample Entropy, RMS
• ECG: HRV metrics (mean RR, SDNN, RMSSD), Entropy, RMS
• EOG: Zero-Crossing Rate, Entropy, RMS
• EMG: Hjorth Parameters, Entropy, RMS

🤖 Fusion Techniques

• Early Fusion: Combines all modality features before model input
• Late Fusion: Combines predictions from modality-specific models
• Voting Classifier used for late fusion aggregation
• Models used: SVM, Random Forest, Gradient Boosting, KNN, AdaBoost, MLP, Extra Trees, SGD, Logistic Regression

---

🧠 Explainable AI (XAI)

To enhance interpretability:

• LIME explains individual predictions and highlights contributing features

Figure: LIME’s Prediction probabilities computed for both classes with feature values

• SHAP provides global and local interpretability of model outputs

Figure: SHAP’s Prediction probabilities computed for both clases with feature values.

• Both methods emphasize the importance of ECG and EMG features in classification

---

📊 Performance Summary

Method	Accuracy	Precision	Recall	F1 Score
Early Fusion	95%	94%	97%	95%
Late Fusion	92%	91%	93%	92%
Best Classifiers	Random Forest, MLP, Extra Trees	All with 95% Accuracy

---

🧭 Workflow

Figure: Overall workflow for data preparation, feature extraction, fusion, classification, and explainability._

---

🧰 Setup Instructions

📦 Requirements

• Python 3.8+
• numpy, pandas, matplotlib, scikit-learn, scipy
• XAI: lime, shap
• Jupyter Notebook or Google Colab