Classification of Healthy People and Schizophrenics Using Time- Frequency Domain Features Extracted from Electroencephalogram Signals

This study proposes a novel automated diagnosis scheme for schizophrenia that achieves 100% accuracy in distinguishing patients from healthy individuals by extracting and selecting time-frequency domain features from EEG signals and classifying them using various machine learning algorithms.

The Gist

Imagine your brain is like a bustling city with millions of cars (neurons) driving around, creating a constant flow of traffic. When the city is healthy, the traffic flows in a predictable, organized rhythm. But in people with Schizophrenia, the traffic patterns get chaotic, unpredictable, and sometimes completely gridlocked.

Currently, doctors try to diagnose this "traffic chaos" by interviewing the driver (the patient) and asking, "How does the traffic feel?" This is subjective, slow, and sometimes the driver might not even realize the traffic is messed up, or they might confuse it with other types of traffic jams (like Bipolar Disorder).

This paper proposes a smart traffic camera system (an automated computer program) that looks directly at the traffic flow (brain waves) to spot the chaos instantly and accurately.

Here is how they built this system, explained simply:

1. The Raw Material: Listening to the Brain's Radio

The researchers used EEG, which is like putting a set of microphones on the scalp to record the brain's electrical "radio waves."

• The Subjects: They recorded 14 healthy people and 14 people with Schizophrenia.
• The Setup: Everyone sat quietly with their eyes closed for 5 minutes. It's like asking everyone to sit in a quiet room and hum a tune so the microphones can hear the natural rhythm of their brains.

2. Cleaning the Signal: Tuning the Radio

Before analyzing, they had to clean up the recording.

• The Filter: Just like tuning a radio to remove static, they filtered out "noise" like heartbeats, eye blinks, and electrical hums from the power grid.
• The Result: A clean, pure recording of just the brain's activity.

3. The Detective Work: Finding the Clues (Feature Extraction)

This is the most creative part. The researchers didn't just look at the raw sound; they broke it down into three different "languages" to find hidden clues that the human eye can't see.

• Language 1: Time Domain (The Rhythm)

  • Analogy: Imagine listening to a drumbeat. Is it steady? Is it erratic?
  • They measured things like how often the signal crosses zero (how jittery it is) and how complex the pattern is. It's like checking if the drummer is keeping a simple beat or going crazy with a complex solo.

• Language 2: Frequency Domain (The Pitch)

  • Analogy: Imagine a piano. Some notes are low (bass), some are high (treble).
  • They checked how much "energy" was in the low notes (Delta/Theta) versus the high notes (Alpha/Beta). In Schizophrenia, the "bass" might be too loud, and the "treble" too quiet. They calculated the ratio between these notes to find the imbalance.

• Language 3: Time-Frequency Domain (The Movie)

  • Analogy: The first two languages are like looking at a still photo or listening to a single note. This one is like watching a movie of the signal.
  • They used a mathematical tool called a Wavelet Transform (think of it as a magical zoom lens) to see how the rhythm and pitch change over time. This revealed hidden patterns that were invisible in the other two views.

4. The Filter: Picking the Best Clues (Feature Selection)

They found 18 different types of clues (like "jitteriness," "bass volume," "complexity"). But having too many clues confuses the computer.

• The Strategy: They used a smart filter (a mix of "Mutual Information" and "Forward Selection") to ask: "Which 10 clues are the absolute best at telling a healthy brain from a sick one?"
• The Result: They narrowed it down to the top 10 "super clues" for each computer program they tested.

5. The Judges: The Classifiers

They fed these top clues into seven different "judges" (Machine Learning algorithms) to see who could make the best diagnosis:

• KNN: Looks at neighbors. "If your brain waves look like the healthy people next to you, you're healthy."
• SVM: Draws a line. "If you are on this side of the line, you are healthy; on the other, you are sick."
• Decision Tree: Asks a series of Yes/No questions. "Is the bass too loud? Yes. Is the rhythm too fast? Yes. -> Diagnosis: Schizophrenia."
• Naive Bayes: Uses probability. "Based on past data, there is a 99% chance this pattern belongs to a sick person."

The Grand Finale: The Scoreboard

The results were shocking.

• Before they picked the best clues, the computers were decent (around 90-95% accurate).
• After they used their smart filter to pick the perfect 10 clues, three of the judges (Linear SVM, Nonlinear SVM, and Decision Tree) got a perfect 100% score.

They didn't just guess; they correctly identified every single person in the study as either healthy or having Schizophrenia.

Why This Matters

Think of this as a super-powered stethoscope.

• Current Diagnosis: A doctor listening to a patient's story (which can be vague or misinterpreted).
• This New Tool: A computer looking at the brain's electrical "traffic" and saying, "I see the pattern. This is Schizophrenia," with 100% certainty on this specific dataset.

The Bottom Line:
The authors aren't saying this replaces the doctor yet. Instead, they are offering a second opinion that is fast, objective, and incredibly accurate. It's like having a co-pilot who never gets tired and never misses a detail, helping the doctor make the right call faster and cheaper.

In the future, this "traffic camera" could be used to catch the disease earlier, test different types of schizophrenia, or even help diagnose it in teenagers, potentially saving lives by catching the chaos before it takes over the city.

Technical Summary

1. Problem Statement

Schizophrenia (SZ) is a chronic, complex mental disorder characterized by neurobiological deficits, hallucinations, delusions, and disorganized thinking. Currently, diagnosis relies heavily on subjective behavioral observations and clinical interviews conducted by psychiatrists. This traditional approach faces several challenges:

• Subjectivity and Reliability: Manual screening is prone to human error and inter-rater variability.
• Misdiagnosis: SZ is frequently confused with other disorders like bipolar disorder or major depressive disorder.
• Resource Intensity: The process is time-consuming, costly, and exhausting.
• Lack of Objective Biomarkers: There is a critical need for an automated, objective, and cost-effective diagnostic tool to assist clinicians, particularly for early detection.

2. Methodology

The study proposes an automated machine learning framework to classify EEG signals into "Healthy" and "Schizophrenic" categories using a multi-domain feature extraction and hybrid selection approach.

A. Dataset and Preprocessing

• Data Source: EEG recordings from the Atieh Clinical Neuroscience Center.
• Subjects: 28 total participants (14 healthy controls, 14 diagnosed with SZ).
• Acquisition: 19 electrodes (10-20 system), 500 Hz sampling rate, resting state with eyes closed (5 minutes of data used per subject).
• Preprocessing:
  • Filtering: Finite Impulse Response (FIR) band-pass filter (0.5–45 Hz) to remove power line interference and high-frequency noise.
  • Artifact Removal: Independent Component Analysis (ICA) to eliminate non-brain artifacts (e.g., eye blinks, muscle movement, heartbeat).

B. Feature Extraction

The authors extracted 18 feature models across three distinct domains to capture the complexity of EEG signals:

1. Time Domain:
   • Zero Crossing Rate (ZCR)
   • Hjorth Mobility and Complexity
   • Higuchi's Fractal Dimension
   • Lempel-Ziv Complexity (LZC)
2. Frequency Domain:
   • Mean Power in Delta, Theta, Alpha, Beta, and Gamma bands.
   • Delta-to-Alpha Power Ratio (a known biomarker for SZ).
   • Spectral Entropy.
3. Time-Frequency Domain (via Discrete Wavelet Transform - DWT):
   • Wavelet: Daubechies order 6 (db6) selected for optimal similarity.
   • Features: Kurtosis, Hurst Exponent, Shannon Entropy, Log Energy Entropy, Rényi Entropy, and Tsallis Entropy.

C. Feature Selection

To address the "curse of dimensionality" and prevent overfitting (given the small dataset of 28 subjects), a hybrid feature selection strategy was employed:

1. Mutual Information (MI): Used as a scalar filter to remove irrelevant features based on their dependence on the target class.
2. Sequential Forward Feature Selection (SFFS): A vector-based wrapper method that iteratively adds features to maximize classification accuracy. The top 10 feature combinations were selected for each classifier.

D. Classification Algorithms

Seven machine learning classifiers were evaluated:

• K-Nearest Neighbors (KNN) and Weighted KNN (WKNN)
• Linear and Nonlinear Support Vector Machines (SVM with RBF kernel)
• Decision Tree (DT)
• Linear Discriminant Analysis (LDA)
• Naive Bayes (NB)

E. Validation

To ensure robustness with a limited dataset, two cross-validation techniques were used:

• 10-Fold Cross-Validation.
• Leave-One-Out Cross-Validation (LOOCV): Used as the primary metric, where $N=28$ iterations were performed (training on 27, testing on 1).

3. Key Contributions

• Multi-Domain Integration: Unlike previous studies focusing on a single domain, this work integrates Time, Frequency, and Time-Frequency features to provide a comprehensive view of brain dynamics in SZ.
• Hybrid Feature Selection: The combination of Mutual Information (for speed/irrelevance filtering) and SFFS (for optimal subset selection) effectively reduced dimensionality while retaining high discriminative power.
• Identification of Key Biomarkers: The study identified that Lempel-Ziv Complexity (Time), Spectral Entropy (Frequency), and Logarithmic Energy Entropy (Time-Frequency) were the most consistent and effective features across all classifiers.
• High Accuracy with Small Data: The framework achieved near-perfect classification accuracy despite a small sample size (14 per class), demonstrating the efficacy of the proposed feature engineering.

4. Results

The performance was evaluated using Accuracy, Sensitivity, and Specificity.

• Single Domain Performance:
  • Time Domain: Best accuracy was 96.42% (Linear SVM).
  • Frequency Domain: Best accuracy was 100% (LDA).
  • Time-Frequency Domain: Best accuracy was 100% (Naive Bayes).
• Integrated Domain Performance (After Feature Selection):
  • When combining all three domains and applying the hybrid feature selection, three classifiers achieved 100% accuracy using LOOCV:
    1. Linear SVM (100% Sensitivity, 100% Specificity)
    2. Nonlinear SVM (100% Sensitivity, 100% Specificity)
    3. Decision Tree (100% Sensitivity, 100% Specificity)
  • Naive Bayes also performed exceptionally well, achieving 98% accuracy with 10-fold cross-validation and 95% with LOOCV.
  • KNN generally underperformed compared to SVM and Tree-based models, highlighting the non-linear separability of the data.

5. Significance and Conclusion

• Clinical Utility: The proposed system offers a highly accurate, automated, and non-invasive complementary tool for psychiatrists. It can reduce misdiagnosis rates and provide objective evidence to support clinical decisions.
• Efficiency: By identifying a small subset of highly relevant features (e.g., entropy and complexity measures), the method reduces computational load, making it suitable for real-time or portable diagnostic applications.
• Future Directions: The authors suggest applying this framework to:
  • Differentiate between positive and negative symptoms of schizophrenia.
  • Classify various subtypes of schizophrenia.
  • Analyze adolescent populations to compare developmental stages of the disorder.

In summary, this paper demonstrates that a hybrid approach combining multi-domain EEG feature extraction with advanced machine learning classifiers can effectively distinguish schizophrenia patients from healthy controls with 100% accuracy, offering a promising avenue for objective psychiatric diagnosis.