Post

🧠 New preprint! How well can current algorithms *actually* detect neural "replay" in the human brain under absolutely optimal conditions? We built FASTIMAGES: A combined MEG + fMRI benchmark with KNOWN neural sequences, so replay-detection methods can finally be validated against a ground truth.

The problem: Methods like TDLM & SODA are widely used to find replay, but we rarely know *if* or *when* replay actually happened. Without invasive recordings -> no ground truth. Therefore, we created a dataset that contains clear neural sequences to benchmark current algorithms.

FASTIMAGES = 70 participants (30 MEG + 40 fMRI) who saw 5 images flashing in a fast sequences (132–612 ms onset-to-onset), plus localizer trials to train decoders. This produces clearly-expressed neural sequences with a KNOWN order and timing. See our project page for some animations.

As a first step, we recorded a localizer, in which participants saw the five images slowly, in random order. Here, apply standard techniques to train a brain decoder that will be used later for sequence detection. Decoders work as expected ~55% in MEG (peak ~150 ms) and ~70% in fMRI (peak ~4 s).

You can literally see the sequence order in the averaged decoder output of the MEG and partially in the fMRI 👀 however, individual trials are much noisier. Perfect, so let's test TDLM and SODA!

TDLM (built for MEG, Liu et al. 2021) successfully finds significant "sequenceness" at the expected time lag in ALL four speed conditions. But variance is high! Only ~half of participants are individually significant, and sequenceness is highly correlated to decoder quality, not behaviour.

SODA (built for fMRI, Wittkuhn et al. 2021) also recovers the sequences, mainly driven by the "offset" period, which is significant in every speed condition. Similarly, individual participants are not always significant, and variance between participants is high. Dependency on the decoder is less!

Effect sizes are comparable in the native modality (TDLM Cohen's d ≈ 0.77). Bootstrap power analysis shows: ~20 participants × 20 trials → 80% power. As this was an absolutely best case scenario and real effects are likely much smaller, you'll likely need to go beyond that for finding real replay.

Use FASTIMAGES as a sanity check, a testbed for new methods, and a shared benchmark for human replay detection. 🧪 🌐 Project page: cimh-clinical-psychology.github.io/FASTIMAGES-b... 📄 Preprint: www.biorxiv.org/content/10.6... 💻 Code + 💾 data 👇

💻 Code: github.com/CIMH-Clinica... 📦 FASTIMAGES BIDS datasets on GIN: • MEG: gin.g-node.org/skjerns/FAST... • fMRI: gin.g-node.org/skjerns/FAST...