MambaPD: Bidirectional Selective State Space Model for Parkinson's Disease Detection from Resting-State EEG

Ayesha ZulfiqararXiv 2026 · arXiv

TL;DR

MambaPD introduces a bidirectional selective state space model designed to detect Parkinson’s Disease (PD) from resting-state EEG data. By employing a strict subject-independent validation protocol, the paper demonstrates that while Mamba-based architectures are highly competitive—particularly in sensitivity—the "95%+ accuracy" often reported in literature is frequently an artifact of data leakage, with real-world performance currently hovering around 70%.

What Problem Does This Paper Address?

Parkinson’s Disease is typically diagnosed only after 60-80% of dopaminergic neurons are already lost. While resting-state EEG (rs-EEG) shows early biomarkers like "alpha slowing" and reduced connectivity, building reliable automated detectors is difficult.

The author identifies two major hurdles:

  1. Architectural Limitations: Traditional Deep Learning models struggle with EEG. CNNs are too local, LSTMs suffer from memory bottlenecks over long sequences, and Transformers have a quadratic computational cost that makes them prone to overfitting on small medical datasets.
  2. The "Leakage" Crisis: Many existing studies claim near-perfect accuracy (>95%) by splitting data at the "epoch" level. This means segments from the same patient appear in both training and testing sets, allowing the model to simply "memorize" individual brain signatures rather than learning generalizable PD biomarkers.

Methods

The proposed framework, MambaPD, leverages the recent "Mamba" architecture—a selective State Space Model (SSM) that processes sequences in linear time.

The pipeline consists of several key stages:

  • Preprocessing: Raw EEG is filtered (0.5–40 Hz) to capture relevant bands (delta to low gamma) and resampled to 256 Hz.
  • Channel Mixer: A lightweight module that learns spatial relationships across electrodes at each time point before temporal processing begins.
  • Patch Tokenization: The 4-second EEG epochs are divided into 8 non-overlapping patches (0.5s each), roughly the duration of fast oscillatory events like beta spindles.
  • Bidirectional Mamba Encoder: Unlike standard Mamba, which is unidirectional, MambaPD uses two stacks (forward and backward) to ensure the model has full temporal context for every segment of the signal.
  • Selective Mechanism: The core of Mamba is its ability to dynamically "choose" what to remember or forget based on the input. In an EEG context, this allows the model to potentially ignore uninformative "noise" and focus on transient diagnostic events.
Overview of the MambaPD framework

Datasets

The model was evaluated on two public datasets using a strict subject-independent 5-fold cross-validation:

  • UNM (University of New Mexico): 56 subjects (28 PD, 28 Healthy Controls) recorded with 63 EEG channels.
  • SanDiego (UC San Diego): 31 subjects (15 PD, 16 Healthy Controls) recorded with 40 EEG channels.

Crucially, the author ensured that no data from a test subject was ever seen during the training phase, providing a more realistic estimate of clinical utility.

Key Insights and Ideas

  • The Reality Check on Accuracy: When subject identity leakage is removed, accuracy drops significantly. The study found that most models (TCN, Transformers, Mamba) perform similarly, with accuracies between 60% and 72%. This suggests that the "choice of model" might be less important than the "quality of the validation protocol."
  • Sensitivity vs. Specificity: MambaPD showed a distinct profile, achieving the highest sensitivity (0.824) on the UNM dataset. This makes it a strong candidate for screening tools, where missing a potential PD case is more costly than a false alarm.
  • Fold Variance is the Real Boss: The paper notes that the difference in performance between different "folds" (groups of subjects) was often larger than the difference between the models themselves. This highlights how sensitive these models are to individual patient variability in small datasets.

Key Results

MambaPD was compared against four baselines: Temporal Convolutional Networks (TCN), BiLSTM with Attention, EEGNet, and a Patch Transformer.

  • UNM Dataset: MambaPD achieved an accuracy of 0.714, statistically indistinguishable from the Patch Transformer (0.715) and TCN (0.708). However, it led the pack in AUC (0.765) and Sensitivity (0.824).
  • SanDiego Dataset: This smaller cohort proved more challenging. TCN performed best (0.703), while MambaPD scored 0.631.
Comparison of Accuracy and AUC-ROC across models

Limitations and Future Work

The author is transparent about several limitations:

  • Small Sample Size: With only 31–56 subjects, the test folds contain as few as 6 people. A single "outlier" subject can swing the results significantly.
  • Optimistic Estimates: Checkpoints were selected based on the held-out fold's performance, which likely inflates the reported numbers slightly.
  • Hyperparameter Tuning: Baselines were not individually optimized, which might disadvantage them in the comparison.

Future Directions: The author suggests moving toward multi-site validation, repeated runs to account for variance, and implementing nested model selection to provide truly unbiased performance metrics for clinical deployment.

Summary updated

Abstract

Typical diagnosis of Parkinson's disease (PD) occurs only after significant dopaminergic loss, while resting-state EEG abnormalities like alpha slowing and reduced coherence occur earlier. While many published EEG-based PD detectors claim accuracy greater than 95%, many use epoch-level splits which leak subject identity. I introduce MambaPD, a bidirectional selective state space model for resting-state EEG-based PD detection, blending the channel mixing, patch tokenization and bidirectional Mamba encoder. It is compared with four baselines (TCN, BiLSTM with attention, EEGNet and a patch Transformer) in two public datasets (UNM, SanDiego) using subject-independent 5-fold cross-validation (checkpoints were selected on the held-out fold, so absolute estimates are optimistic). On UNM the accuracy is between 0.67 and 0.71, and MambaPD (0.714) is not statistically distinguishable from the Transformer (0.715), and TCN (0.708), with the highest sensitivity (0.82), while the lowest specificity (0.52). For the smaller SanDiego cohort, accuracy is in the range of 0.60 to 0.70, with the highest scores (0.703) from TCN and the lowest (0.601) from EEGNet, while MambaPD scores 0.631, with the lowest specificity (0.46) among the five models. The differences among the models on UNM are minor compared to the differences among folds. I report the whole pipeline, per-fold results, and the shortcomings of the protocol.
BibTeX
@misc{zulfiqar2026mambapd,
  author = {Ayesha Zulfiqar},
  title = {{MambaPD: Bidirectional Selective State Space Model for Parkinson's Disease Detection from Resting-State EEG}},
  year = {2026},
  eprint = {2611.12026},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.12026},
}
Figure