A hybrid feature selection approach for the early diagnosis of Alzheimer’s disease

Esteve Gallego-Jutglà, Jordi Solé‐Casals, François Vialatte, Mohamed Elgendi, Andrzej Cichocki, Justin DauwelsJournal of Neural Engineering 2015 · DOI

Abstract

The new features selection method described in this work may be a reliable tool that could help to design a realistic system that does not require prior knowledge of a patient's status. With that aim, we explore the standardization of features for MCI and Mild AD data sets with promising results.

Introduction

Early diagnosis of Alzheimer’s Disease (AD) remains one of the most significant challenges in modern neurology. While definitive cures are currently unavailable, early intervention is critical for delaying symptom progression and improving patient quality of life. Electroencephalography (EEG) offers a promising diagnostic path due to its non-invasive nature, low cost, and portability. However, the high level of noise in EEG data and the complexity of identifying specific biomarkers for cognitive decline have historically limited its clinical application.

This research addresses these challenges by introducing a hybrid feature selection method designed to identify the most discriminative frequency ranges and synchrony measures. By moving beyond traditional frequency bands, this approach captures the subtle "slowing" of brain activity that characterizes the early stages of Alzheimer’s.

Methodology

The study proposes a two-step hybrid framework to refine EEG data into a robust set of diagnostic features.

Standard EEG analysis typically relies on five fixed frequency bands (Delta, Theta, Alpha, Beta, and Gamma). However, pathological changes in AD do not always align with these arbitrary boundaries. The researchers conducted an exhaustive search across 435 possible frequency ranges between 1 and 30 Hz to identify the specific spectral windows that provide the highest diagnostic sensitivity.

2. Hybrid Feature Selection via OFR

To select the most effective features from a vast pool of candidates, the team utilized Gram-Schmidt Orthogonal Forward Regression (OFR). This algorithm iteratively selects features that provide the most significant amount of new (orthogonal) information relative to the features already chosen, ensuring a compact and highly efficient feature set.

3. Overfitting Mitigation: The Random Probe Method

Given the relatively small size of clinical datasets, overfitting is a primary concern. To ensure the model's generalizability, the researchers implemented a "Random Probe" method. By introducing 5,500 noise-based features into the selection process, the algorithm only retained real features that significantly outperformed the noise, acting as a rigorous statistical filter.

Experimental Setup and Datasets

The methodology was validated using two distinct datasets to assess performance across different stages of cognitive impairment:

  • Mild Cognitive Impairment (MCI) Dataset: Comprising 22 patients who eventually progressed to AD and 38 healthy controls.
  • Mild AD Dataset: Comprising 17 patients diagnosed with mild Alzheimer’s and 24 healthy controls.

Both datasets utilized resting-state EEG (eyes closed), focusing on 20-second segments of artifact-free data to ensure signal purity.

Key Insights and Results

The Significance of Brain "Slowing"

The study confirmed the "slowing" hypothesis: as AD progresses, brain energy shifts from higher-frequency activity to lower-frequency "slow" waves. Relative Power (RP) emerged as the most critical single feature. Interestingly, the optimal frequency ranges for diagnosis were found to be stage-specific. For MCI, the most discriminative range was 2–9 Hz, while for Mild AD, it was 4–7 Hz.

The Role of Synchrony

While power measures provided a strong baseline, the inclusion of synchrony measures—such as Coherence and Granger Causality—was essential for achieving peak accuracy. These measures reflect the functional connectivity of the brain, capturing how effectively different regions communicate.

Classification Performance

The hybrid approach achieved exceptional results using Linear Discriminant Analysis (LDA) with leave-one-out cross-validation:

  • MCI Diagnosis: 95% accuracy using 11 optimized features.
  • Mild AD Diagnosis: 100% accuracy using only 4 optimized features.

Even when restricted to a single optimized Relative Power feature, the system achieved 78.33% accuracy for MCI and 97.56% for Mild AD, highlighting the power of selecting the correct frequency range over standard bands.

Discussion and Limitations

While the results are promising, the authors identify several areas for further refinement. The small sample size remains a limitation, though the Random Probe method was employed to mitigate this. Additionally, in the Mild AD dataset, an age gap between patients and controls suggests that some of the detected signal could be influenced by natural aging processes. Finally, the current study used global scalp averages; the authors suggest that regional analysis (e.g., focusing on the temporal or parietal lobes) could yield even more specific biomarkers.

Conclusion

This hybrid feature selection method represents a significant step toward a standardized, automated diagnostic tool for Alzheimer’s. By optimizing frequency ranges and leveraging orthogonal regression, the study demonstrates that EEG can provide highly accurate, early-stage diagnostic markers without requiring prior knowledge of a patient’s clinical status.

Summary updated

Abstract

The new features selection method described in this work may be a reliable tool that could help to design a realistic system that does not require prior knowledge of a patient's status. With that aim, we explore the standardization of features for MCI and Mild AD data sets with promising results.
BibTeX
@article{gallegojutgla2015hybrid,
  author = {Esteve Gallego-Jutglà and Jordi Solé‐Casals and François Vialatte and Mohamed Elgendi and Andrzej Cichocki and Justin Dauwels},
  title = {{A hybrid feature selection approach for the early diagnosis of Alzheimer’s disease}},
  journal = {Journal of Neural Engineering},
  year = {2015},
  doi = {10.1088/1741-2560/12/1/016018},
}
Figure