Skip to content

Latest commit

 

History

History
54 lines (43 loc) · 3.83 KB

DESIGN_PROPOSAL_BIDS_DYNAMICS.md

File metadata and controls

54 lines (43 loc) · 3.83 KB

EEG/MEG Tooling Landscape Analysis

Executive Summary

For a new tool competing with osl-dynamics while offering BIDS support and a lightweight footprint, the strongest strategic direction is to fork or extend MNELAB.

Why?

  • GAP: osl-dynamics lacks native BIDS support, requiring data conversion.
  • SOLUTION: MNELAB has robust, native BIDS integration (via mne-bids).
  • OPPORTUNITY: Integrating "Dynamic Connectivity" (HMMs) into the MNELAB GUI creates a tool that solves the "BIDS-to-Analysis" friction point that currently exists in the OSL ecosystem.

Tool Analysis Matrix

Feature osl-dynamics MNELAB generic MNE-Python EEGLAB (Octave) SigViewer
Primary Focus HMMs / Dyn. Conn. GUI Frontend General Processing Processing Viewing
BIDS Support ❌ (Custom formats) ✅ (Native Read/Write) ✅ (via mne-bids) ⚠️ (Plugin required) ❌
Lightweight? Medium (Dependencies) ✅ (GUI wrapper) ❌ (Large dep tree) ❌ (Heavy legacy) ✅ (C++)
Octave Support ❌ ❌ ❌ ✅ (Slow/Glitchy) ❌
Language Python Python Python MATLAB/Octave C++

Detailed Findings

1. The Competitor: osl-dynamics

  • Strengths: Specialized in Hidden Markov Models (HMMs) and Dynamic Network Modes (DyNeMo). Best-in-class for fast scale functional connectivity.
  • Weaknesses:
    • Data Ingestion: Does not digest BIDS datasets directly. Users must convert .bids -> .npy or .mat manually.
    • User Interface: Primarily code/script-based.
  • Your Angle: A tool that applies these advanced dynamic models directly to the BIDS directory structure without an intermediate conversion step.

2. The Frontend: MNELAB

  • Status: Active, Python-based GUI.
  • Architecture: It is effectively a "thin" GUI wrapper around mne-python. This makes it the ideal "lightweight" starting point. It inherits mne-bids capabilities for free.
  • Recommendation: Instead of building a GUI from scratch, fork MNELAB and add a "Dynamic Connectivity" tab. This tab could behave like osl-dynamics but read directly from the loaded BIDS structure.

3. The Constraint: Octave Support

  • EEGLAB: Runs on Octave but is ~2x slower and has significant graphical glitches (scrolling issues). Plugins often fail.
  • FieldTrip: Similar issues; intended for MATLAB.
  • Conclusion: Supporting Octave will likely cripple a modern "lightweight" tool. The Python ecosystem (MNE/MNELAB) offers superior performance and BIDS compliance. If "lightweight" is the goal, avoiding the MATLAB Runtime (required for SPM-Python) and Octave overhead is critical.

4. Lightweight Alternatives

  • MNE-CPP: Very lightweight (C++) and performant for real-time. relying on Python for BIDS. Good for acquisition but likely too low-level for dynamic connectivity modeling.
  • Dynamax: A generic JAX library for probabilistic state space models (HMMs).
    • Strategy: Use Dynamax as the backend engine. It is JAX-based (unlike glhmm or osl-dynamics), enabling high-performance, differentiable modeling. You can replicate osl-dynamics features by feeding Time-Delay Embedded (TDE) data into Dynamax's specialized HMMs.

Proposed "Killer App" Architecture

To satisfy your requirements (BIDS + Compete with OSL + Lightweight + Modern/JAX):

  1. Base: Fork MNELAB (provides GUI + MNE-BIDS).
  2. Engine: Integrate Dynamax (JAX) for the modeling core.
  3. Workflow:
    • Load: Open BIDS folder (MNELAB native).
    • Preprocess: Standard MNE-Python cleaning + Time-Delay Embedding (TDE) step.
    • Analyze: New "Dynamics" Plugin -> Runs Dynamax HMM on TDE data.
    • Visualize: Plot State Probabilities and Transition Matrices (New Widgets).