EEG Classification for Multiple Disorders

• Multi-disorder EEG classification is an approach that leverages machine learning and signal processing to automatically differentiate between various neurological conditions using EEG data.
• By applying covariance-based descriptors on the SPD manifold and multi-domain feature extraction, the method robustly addresses challenges like noise, non-stationarity, and class imbalance.
• Adaptive segmentation and sensitivity-calibrated thresholds enhance clinical utility by ensuring high recall and reducing inter-rater variability in disorder screening.

Multi-disorder EEG classification refers to the automated identification and differentiation of multiple clinically relevant neurological or psychiatric disorder categories using electroencephalography (EEG) signals. This approach leverages advanced ML, signal processing, and mathematical representations to address the heterogeneity, non-stationarity, and class imbalance endemic to real-world clinical data. Two contemporary methodological paradigms stand out: (1) segmenting EEG into homogeneous intervals represented by covariance structure and classified via multiple-instance learning (MI) on the Riemannian manifold of symmetric positive definite (SPD) matrices (Sadatnejad et al., 2017), and (2) extracting multi-domain feature sets in time, frequency, complexity, and connectivity, followed by sensitivity-calibrated ML classification across a spectrum of disorders (Samanta et al., 27 Dec 2025). This article synthesizes the principles, algorithms, and clinical implications of these approaches.

1. Data Acquisition and Preprocessing

Multi-disorder EEG classification requires rigorously curated datasets with harmonized protocols:

• Dataset Characteristics: For large-scale multi-disorder screening, the Harvard Electroencephalography Database (HEEDB) subset “I0002” provides 553 valid sessions from 191 patients, spanning eleven target disorders with severe positive-to-negative class imbalance (e.g., “Peripheral NS Disorders” 34:519; “Seizure Disorders” 340:213) (Samanta et al., 27 Dec 2025).
• EEG Recording Standards: Typical configurations include 16-channel longitudinal bipolar (“double-banana”) montages derived from the 10–20 system. For MI learning approaches, 22-channel (ADHD/BMD) or 19-channel (depression/normal) referential montages are used, band-pass filtered (0.3–50 Hz), optionally partitioned into canonical frequency sub-bands (Sadatnejad et al., 2017).
• Preprocessing Procedures:
  • Bipolar referencing suppresses common-mode noise (Samanta et al., 27 Dec 2025).
  • Quality control enforces full channel completeness.
  • Artifact correction (eye blinks, EMG) may not be explicitly executed; the MI segmentation pipeline is robust to moderate artifact contamination via statistical aggregation (Samanta et al., 27 Dec 2025).

2. Feature Extraction and Representation

Two main strategies define the feature space for multi-disorder EEG classification.

2.1 Covariance-Based Descriptors on the SPD Manifold

• Elementary Segmentation: Continuous EEG is partitioned into overlapping windows of length $T_0$ (1–2 s, 50% overlap). Each yields a data matrix $E(t,T_0)\in\mathbb{R}^{d\times N}$, and a sample covariance $C(t,T_0) = \frac{1}{N-1}E(t,T_0)E(t,T_0)^\top$, which is SPD (Sadatnejad et al., 2017).
• Adaptive Segmentation: Homogeneity boundaries are detected by affine-invariant geodesic distances $$d_{AI}(C_i, C_{i+1}) = \|\log(C_i^{-1/2}C_{i+1}C_i^{-1/2})\|_F$$, merging windows until statistical change points are crossed. This yields segments whose covariances are robust, homogeneous SPD descriptors (Sadatnejad et al., 2017).
• Bag Representation: Each subject is represented as a “bag” $X_i = \{C_{i,1}, ..., C_{i,n_i}\}$ of segment covariances (unlabeled at the segment level).

2.2 Multi-Domain Feature Engineering

• Window-Level Aggregation: Nonoverlapping 60 s windows; per window, extract $d > 3,000$ features including (Samanta et al., 27 Dec 2025):
  • Time-domain statistics: Mean, variance, skewness, excess kurtosis, line length, Hjorth parameters.
  • Spectral features: Welch PSD estimates, relative band power (δ: 0.5–4 Hz, θ: 4–8 Hz, α: 8–13 Hz, β: 13–30 Hz), spectral edge frequency (SEF$_{95}$).
  • Complexity measures: Shannon entropy, Hjorth complexity.
  • Functional connectivity: Pearson correlation for all channel pairs (captures zero-lag synchrony).
• Aggregation and Normalization: For each feature: mean, SD, and percentiles (5th, 50th, 95th) over time windows. Features are median-imputed for missing data and z-score normalized per channel.

3. Classification Algorithms and Learning Frameworks

3.1 Multiple-Instance Learning on the SPD Manifold

• Bag-Level Kernel SVM: Similarity between subject bags computed by summing Riemannian Gaussian kernels over all cross-segment covariance matrix pairs:

$$K_{MI}(X, Y) = \sum_{C\in X}\sum_{C'\in Y} k(C, C')$$

with $$k(C, C') = \exp\big[-d_{AI}(C, C')^2/(2\sigma^2)\big]$$ (Sadatnejad et al., 2017).

• SVM Training: Standard dual-form SVM is trained on the resultant Gram matrix. Labels apply at the bag (subject) level, not segment (Sadatnejad et al., 2017).

3.2 Disorder-Aware Machine Learning with Threshold Calibration

• Model Selection:
  • XGBoost (gradient-boosted decision trees) deployed for disorders with focal/transient signatures (e.g. seizures), using explicit hyperparameter tuning (Samanta et al., 27 Dec 2025).
  • Multi-Layer Perceptron (MLP) classifiers for diffuse or developmental disorders.
• Class Imbalance and Weighted Loss: Training uses inverse-frequency weighting.
• Sensitivity-Oriented Thresholding: Decision thresholds per disorder, $\tau_k$, are tuned such that recall (sensitivity) $\geq 80\%$, optimizing on validation ROC/PR curves (Samanta et al., 27 Dec 2025).
• Evaluation Strategy: Patient-level splits (80% train, 20% test), metrics include accuracy, precision, recall, F1, ROC-AUC, and PR-AUC.

4. Comparative Performance and Key Results

The following summarizes results for both MI-learning-based and large-scale calibration-based multi-disorder EEG classification, focusing exclusively on data-supported findings.

4.1 MI Riemannian Manifold Approach (Sadatnejad et al., 2017)

• Disorders: ADHD vs. bipolar mood disorder (BMD); depression vs. healthy control.
• Accuracy Metrics (eyes-closed, 3–50 Hz average):

Method	ADHD vs. BMD	Depression vs. Normal
Batch-Cov+SVM	74.4%	58%
Mean-Cov+SVM	62.8%	56%
MI-fixed-Cov+SVM	79.1%	86%
MI-adaptive-Cov+SVM	88.4%	90%

• Significant (paired t-test, $p<0.05$) improvements for MI-adaptive-Cov+SVM over all baselines.

4.2 Sensitivity-Calibrated Multi-Disorder ML (Samanta et al., 27 Dec 2025)

• Disorders: 11 categories, including seizure, peripheral NS, developmental, cognitive, movement, cerebral/vascular, sleep, etc.
• Recall-Optimized Test Performance (selected):

Disorder	Model	Threshold	Accuracy	Recall	Precision	F1	ROC-AUC
Seizure Disorders	XGBoost	0.43	80.2%	91.2%	79.5%	84.9%	85.9%
Peripheral NS	MLP	0.13	99.1%	85.7%	100%	99.0%	98.5%
Developmental Delay	MLP	0.03	94.6%	83.3%	50.9%	95.0%	93.9%

• Absolute recall gains for low-prevalence disorders after threshold optimization reach $+42.8$ to $+50.0$ percentage points.

5. Feature Importance, Interpretability, and Physiological Correlates

• Seizure Disorders (XGBoost Gain): Top predictors include gamma relative power variability, median spectral edge frequency, spectral entropy, and extreme amplitude measures. These correlate with established markers of epileptiform discharges and transient synchronization (Samanta et al., 27 Dec 2025).
• Peripheral NS Disorders (MLP Weights): Most important features are zero-crossing rates, spectral edge, and band power ratios, reflecting increased myogenic or conduction artefacts and alterations in oscillatory complexity.
• Covariance Descriptors: In the MI framework, covariance features capture multichannel spatial structure and demonstrate robustness to amplitude scaling and moderate noise. Riemannian distances between covariance matrices respect the nonlinear geometry inherent to the SPD manifold, providing a discriminative representation (Sadatnejad et al., 2017).

6. Architecture Robustness, Calibration, and Generalization

• Robustness to Noise/Artifacts: The MI-adaptive approach maintains accuracy gains up to $+16\%$ over single-covariance SVM even at SNR $= 0$ dB, indicating stability against noise (Sadatnejad et al., 2017). Feature aggregation in the multi-domain pipeline mitigates effects of isolated artifacts (Samanta et al., 27 Dec 2025).
• Calibration and Trade-offs: Post-hoc threshold selection prioritizes recall, often at some precision cost—a deliberate alignment with clinical triage priorities, especially in rare or subtle disorder categories (Samanta et al., 27 Dec 2025).
• Generalization: Leave-one-out and patient-level validation ensure out-of-sample robustness. The MI-kernel framework automatically emphasizes “positive–positive” (i.e., disorder-related) cross-subject segment matches while deemphasizing irrelevant or noisy segments.

7. Clinical and Methodological Implications

Multi-disorder classification frameworks—when rigorously evaluated—offer the following advantages:

• Screening and Triage: Automated, sensitivity-oriented pre-screening accelerates identification of high-risk patterns in large clinical EEG cohorts, supporting scalable, standardized triage (Samanta et al., 27 Dec 2025).
• Reduction of Inter-Rater Variability: Objective classifier outputs serve as a baseline first-pass assessment, reducing interpretation subjectivity.
• Guidelines for Practice: Reported frameworks provide concrete recommendations for EEG dataset curation, feature extraction, model selection, post-training thresholding, and deployment monitoring (Samanta et al., 27 Dec 2025).
• Extension and Future Directions: Prospective work includes validation of these frameworks in multi-center contexts, the inclusion of deeper complexity/connectivity metrics (e.g., sample entropy, phase-locking value, coherence), and integration into clinical decision-support systems.

In sum, contemporary multi-disorder EEG classification integrates adaptive signal segmentation, geometrically informed representations, and sensitivity-calibrated ML, yielding robust, interpretable, and clinically actionable models for multi-pathology EEG analysis (Sadatnejad et al., 2017, Samanta et al., 27 Dec 2025).