Presented at the NeurIPS 2018 Workshop on Machine Learning for Health (ML4H), this research explores the application of Bayesian Deep Learning to transform Electroencephalography (EEG)—a low-cost, non-invasive tool—into a precise instrument for monitoring Alzheimer’s disease (AD) severity.
Introduction
Alzheimer’s Disease remains a significant global health challenge. While "gold standard" biomarkers exist, they often rely on expensive imaging like MRI/PET or invasive procedures such as lumbar punctures. EEG offers a widely available alternative, but using it to objectively measure disease progression is difficult due to the typically small and sparse nature of clinical datasets.
Standard deep learning models often struggle in this domain, frequently overfitting on small samples and providing only "point estimates" without indicating predictive confidence. In clinical settings, the lack of uncertainty bounds can lead to misleading or potentially dangerous interpretations. This study addresses these gaps by employing Bayesian Neural Networks (BNNs) to provide both accurate predictions and reliable uncertainty estimates.
Methodology
The researchers utilized the PRODEM (Prospective Dementia Database Austria) cohort, one of the largest prospective AD EEG trials, involving 188 patients. The target metric was the Mini-Mental State Examination (MMSE) score, a 30-point scale measuring cognitive impairment. Input data consisted of 820 quantitative EEG (QEEG) features derived from continuous recordings, which were reduced to five principal components via Principal Component Analysis (PCA).
To address the challenges of small data and uncertainty, the study compared three modeling approaches:
- Hamiltonian Monte Carlo (HMC) BNN: A sophisticated Markov Chain Monte Carlo technique that explores parameter space using physics-inspired dynamics. It is considered the gold standard for Bayesian inference.
- Monte Carlo (MC) Dropout BNN: A more computationally efficient approximation that treats dropout during test time as a method for sampling from the model's posterior distribution.
- Standard Neural Network (NN): A traditional non-Bayesian baseline.
Results and Discussion
The HMC BNN consistently outperformed the other models in both accuracy and uncertainty calibration. As shown in the performance comparison below, the HMC BNN achieved a significantly lower Mean Squared Error (MSE) and Standardized Mean Squared Error (SMSE) compared to the standard NN and MC Dropout variants.
Key findings include:
- Superior Accuracy: The HMC BNN achieved an aggregate MSE of 12.17, compared to 16.67 for MC Dropout and 33.71 for the standard NN.
- Reliable Uncertainty: The SMSE, which accounts for predictive variance, was substantially lower for HMC (124.02) than for MC Dropout (948.16). This indicates that HMC provides far more reliable "error bars," which are critical for clinical decision-making.
- Bayesian Regularization: The Bayesian framework acted as a natural controller for model complexity, preventing the overfitting that typically plagues standard neural networks on small datasets like the 188-patient cohort used here.
Conclusion
This study demonstrates that Bayesian Deep Learning, specifically using Hamiltonian Monte Carlo, can effectively leverage low-cost EEG data to track Alzheimer's severity. While HMC is computationally intensive, its ability to provide calibrated uncertainty estimates makes it a superior choice for clinical applications where understanding the limits of a model's knowledge is as important as the prediction itself. Future work may explore end-to-end Bayesian models that utilize raw EEG signals and more efficient inference methods for larger datasets.
Summary updated