L-SODA for Preictal Detection

• The paper introduces L-SODA, which employs shrinkage-regularized multinomial logistic regression to robustly estimate directed neural interactions in high-dimensional EEG data for preictal detection.
• It implements a sliding-window framework with precise statistical calibration to derive confidence intervals and quantify directed information flow preceding seizures.
• Empirical results on the CHB-MIT dataset show high sensitivity (97%), low false alarm rates (<0.5/hour), and a mean detection latency of 2–3 seconds before seizure onset.

The logit shrinkage optimized directed information assessment (L-SODA) is an information-theoretic framework for inferring directed neural interactions and classifying preictal periods in scalp electroencephalographic (EEG) data. L-SODA extends classical directed information (DI) estimation by introducing shrinkage-regularized multinomial logistic regression, enabling robust estimation of information flow in high-dimensional, multi-channel EEG settings typical of preictal seizure detection. The method addresses small-sample issues, allows for nonlinear and non-Gaussian interactions, and provides statistically robust confidence intervals, making it suitable for hypothesis testing of network dynamics in preictal and interictal brain states (Chen et al., 2014).

1. Directed Information Estimation

Directed information quantifies the information flow from a source channel $X$ to a target channel $Y$ over an observed sequence, formalized by Massey as:

$$I(X^n\to Y^n) = \sum_{i=1}^n I(X^i; Y_i \mid Y^{i-1}) = \sum_{i=1}^n \mathbb{E}\left[\ln \frac{f(X^i, Y_i \mid Y^{i-1})}{f(X^i \mid Y^{i-1}) f(Y_i \mid Y^{i-1})}\right].$$

Equivalently, it can be represented via conditional entropies as:

$$I(X^n\to Y^n) = \sum_{i=1}^n \left[ H(Y_i \mid Y^{i-1}) - H(Y_i \mid Y^{i-1}, X^i) \right].$$

These formulations capture causal, directional dependencies, which are crucial for dissecting information propagation preceding seizure onset in EEG data.

2. Shrinkage-Regularized Multinomial Logistic Regression

Traditional density estimators perform poorly in high dimensions and with limited samples. L-SODA addresses this by quantizing $X$ into $p$ discrete classes and modeling the conditional probabilities:

$$P_\beta(X^i = k \mid Y^{i-1}) = \frac{\exp\left((Y^{i-1})^T \beta_k(i-1)\right)}{\sum_{j=1}^p \exp\left((Y^{i-1})^T \beta_j(i-1)\right)},$$

with analogous modeling for $P_\beta(X^i \mid Y^i)$. The model parameters $\{\beta_k(m)\}$ are fit by maximizing

$$\hat\beta = \arg\max_\beta \left\{ \sum_{t=1}^n \sum_{k=1}^p \mathbb{I}[X_t = k] \ln P_\beta(X_t = k \mid Y_t) -\lambda \sum_{k=1}^p \|\beta_k\|_2^2 \right\},$$

where $Y$0 is the shrinkage (ridge, $Y$1) parameter. This regularization combats overfitting with limited data typical of short sliding windows in preictal detection.

3. Parameter Selection and Statistical Calibration

Optimal shrinkage is determined by minimizing mean-squared error (MSE) of the DI estimator. Strategies include $Y$2-fold cross-validation—selecting $Y$3 that yields maximal predictive DI stability—or analytic James-Stein-type shrinkage towards a weak prior, balancing estimator bias and variance analytically:

$Y$4

L-SODA provides statistical confidence intervals for estimated DI via the central limit theorem:

$Y$5

with variance estimated via delta method or bootstrap. A $Y$6 confidence interval is

$Y$7

Comparisons of DI between preictal and interictal states use a two-sample $Y$8-test, with Benjamini–Hochberg FDR correction for multiple pairwise comparisons.

4. Protocol for Preictal Detection on CHB-MIT Dataset

Application to the CHB-MIT database involves specific segmentation and feature computation steps:

• Preictal window: 10 s preceding annotated seizure onset.
• Ictal window: From seizure onset to seizure end.
• Interictal windows: ≥1 h from any seizure to avoid transitional artifacts.
• Sliding window: Segment data into 1 s windows with 50% overlap (labels determined by window center).

For each window:

• Preprocessing: Bandpass 0.5–70 Hz, notch 59–61 Hz to remove line noise; optional ICA or regression for artifact removal.
• Channel handling: All available scalp channels (16–23 channels); optional mapping to Brodmann or lobar regions.
• Local DI estimation: For each channel pair, quantize $Y$9 to $$I(X^n\to Y^n) = \sum_{i=1}^n I(X^i; Y_i \mid Y^{i-1}) = \sum_{i=1}^n \mathbb{E}\left[\ln \frac{f(X^i, Y_i \mid Y^{i-1})}{f(X^i \mid Y^{i-1}) f(Y_i \mid Y^{i-1})}\right].$$0 via Lloyd–Max; $$I(X^n\to Y^n) = \sum_{i=1}^n I(X^i; Y_i \mid Y^{i-1}) = \sum_{i=1}^n \mathbb{E}\left[\ln \frac{f(X^i, Y_i \mid Y^{i-1})}{f(X^i \mid Y^{i-1}) f(Y_i \mid Y^{i-1})}\right].$$1 remains continuous. Fit the shrinkage-regularized logistic regression model and compute:

$$I(X^n\to Y^n) = \sum_{i=1}^n I(X^i; Y_i \mid Y^{i-1}) = \sum_{i=1}^n \mathbb{E}\left[\ln \frac{f(X^i, Y_i \mid Y^{i-1})}{f(X^i \mid Y^{i-1}) f(Y_i \mid Y^{i-1})}\right].$$2

• Feature aggregation: Form the $$I(X^n\to Y^n) = \sum_{i=1}^n I(X^i; Y_i \mid Y^{i-1}) = \sum_{i=1}^n \mathbb{E}\left[\ln \frac{f(X^i, Y_i \mid Y^{i-1})}{f(X^i \mid Y^{i-1}) f(Y_i \mid Y^{i-1})}\right].$$3 DI matrix; vectorize (upper-triangle) for individual windows; aggregate over the 10 s preictal period for mean, standard deviation, and maximal DI statistics.

5. Classification and Performance Metrics

The feature vectors are input to a support vector machine (SVM) with RBF kernel (LibSVM implementation; kernel bandwidth $$I(X^n\to Y^n) = \sum_{i=1}^n I(X^i; Y_i \mid Y^{i-1}) = \sum_{i=1}^n \mathbb{E}\left[\ln \frac{f(X^i, Y_i \mid Y^{i-1})}{f(X^i \mid Y^{i-1}) f(Y_i \mid Y^{i-1})}\right].$$4, cost $$I(X^n\to Y^n) = \sum_{i=1}^n I(X^i; Y_i \mid Y^{i-1}) = \sum_{i=1}^n \mathbb{E}\left[\ln \frac{f(X^i, Y_i \mid Y^{i-1})}{f(X^i \mid Y^{i-1}) f(Y_i \mid Y^{i-1})}\right].$$5). Two-fold cross-validation is performed within each patient to avoid data contamination. Preictal and interictal windows are balanced either by subsampling or class-weighting the SVM.

Performance is quantified using:

• Sensitivity (true positive rate): Fraction of preictal windows detected.
• Specificity (true negative rate): Fraction of interictal windows correctly rejected.
• False alarms per hour: Interictal windows misclassified as preictal, normalized to total interictal time.
• Detection latency: Time between first preictal detection and clinical seizure onset.

6. Empirical Results and Comparative Evaluation

On CHB-MIT data (173 annotated seizures, 22 pediatric subjects):

• Sensitivity: 97.1% of seizures detected within the 10 s preictal window.
• Detection latency: Mean of 2.8 s lead time before seizure onset.
• False alarm rate: 0.2–0.5 per hour, lower than both unregularized DI and energy-based methods.

Comparative results indicate:

• Energy-based SVM detectors (Shoeb et al.): 96% sensitivity, ~4 s latency, higher false-alarm rate.
• Unregularized DI (Quinn et al.): ~95% sensitivity, ~3.5 s latency (Chen et al., 2014).

L-SODA's enhanced performance in early warning derives from sensitivity to directional, non-Gaussian dependencies in EEG and the use of shrinkage regularization to prevent overfitting in short analysis windows. Average expected lead time is 2–3 s prior to onset, with sensitivity ≈97% and false alarms <0.5/hour.

7. Significance for Preictal EEG Analysis

L-SODA enables high-resolution separability of preictal and interictal periods in clinical EEG by leveraging robust estimates of information flow under minimal assumptions. Its framework supports per-patient adaptation, regionally resolved interaction mapping (e.g., Brodmann areas), and rigorous statistical testing of changes in connectivity preceding seizures. The approach is especially advantageous in datasets with limited annotated preictal periods and high-dimensional multi-channel signals, characteristic of preictal seizure detection tasks in clinical neurology (Chen et al., 2014).