Liset M de la PridaJul 4, 2023
Our new @biorxivpreprint preprint šŸ“ is out!! https://www.biorxiv.org/content/10.1101/2023.07.02.547382v1 by Andrea Navas-Olive and Adrian Rubio. šŸ‘‰šŸ¼An opensource ML toolbox for sharp-wave ripple detection https://github.com/PridaLab/rippl-AI Trained with šŸ­ data and applied to šŸ™Š in collaboration with @karihoffman and Sam Abbaspoor #DeepCode
Show threadEl DuvelleJul 4, 2023
@LMPrida @biorxivpreprint @karihoffman
This looks very useful and is definitely needed! Automated ripple classification is so difficultā€¦
Is there a figure showing something like % agreement between the models and the ā€œground truthā€ - say, some human classification?
In our lab weā€™ve found that automated methods only go as high as ~50% while between humans can be >70% (if I remember correctly)
Andrea Navas-OliveJul 12, 2023
Show threadkari hoffman

@elduvelle @LMPrida @biorxivpreprint @cogneurophys Iā€™m stoked that it worked out so well! The assessment followed closely the methods in Navas-Olive CNN paper https://elifesciences.org/articles/77772 using F1 (balanced accuracy) to reflect both precision and recall (i.e. sensitivity). So both FN and FPs count against the score, equally. The human raters were around .7 and the monkey data started at ~.5 and reached ~.6 (same as mouse levels!) after retraining. A pleasant surprise, given visible differences in the SWR phenotype between rodent and primate clades!

I think Andrea will post more details soon, but meanwhile, some relevant keywords for interested folks (can you think of others we should use?)

#neuroscience #MemoryReplay #learningandmemory #hippocampus #ripples #SWR #replay #cnn #lstm #openscience #hackathon #oscillations

Deep learning-based feature extraction for prediction and interpretation of sharp-wave ripples in the rodent hippocampus

A new method is described to identify sharp-wave ripples from the rodent hippocampus with deep learning techniques, which may help to identify and characterize previously undetected physiological events.

eLife
Jul 4, 2023 at 7:12pmWeb
Show threadLiset M de la PridaJul 4, 2023
@karihoffman @elduvelle @cogneurophys Exactly, experts are at 0.7 F1 so the models work pretty well. In our eLife paper we actually showed how F1 of the CNN improves if we use the consolidated GT (union of two expertsā€™ GTs). This means that some FP of the model are actually in the GT of some expert, calling for community tagging to improve performance. As for the models, here is a comparison for Precision and Recall (the diagonal is blanked).
Show threadEl DuvelleJul 4, 2023
@LMPrida @karihoffman @cogneurophys very cool! Thank you for the explanation! Iā€™ll definitely need to read this. Maybe we should have some informal meeting at some point about all the different techniques that ripple labs use to do their categorization :)
Show threadkari hoffmanJul 5, 2023

@elduvelle @LMPrida @cogneurophys yes and then turn it into some document or consensus statement! Kidding, a little bit. Last yearā€™s paper scratched the surface: https://www.nature.com/articles/s41467-022-33536-x

Have a look at the PridaLab/rippl-AI github for the various models. MvdM already had a nice pipeline for annotation, a few years back. I guess you all would have detection algorithms to test, too. Iā€™m happy to offer NHP data for testing ā€” something improved from what we thought of doing w/ MvdM a few years ago. Or maybe we formalize a benchmark ā€œdata menagerieā€ that all algoā€™s should try and classify. The present paper used some mouse and linear-array NHP data, but one could add to that.

A consensus statement on detection of hippocampal sharp wave ripples and differentiation from other fast oscillations - Nature Communications

While the contribution of sharp wave ripples in memory consolidation and decision-making is established in rodent models, our understanding of their role in human memory is incomplete. Here, the authors discuss common methodological challenges in detecting, analyzing, and reporting sharp wave ripples, then they suggest practical solutions to distinguish them from other high-frequency events

Nature
Show threadEl DuvelleJul 5, 2023
@karihoffman @LMPrida @cogneurophys Haha yeahā€¦ thatā€™s a great paper! I was surprised at the frequency bands recommended for ripple detection, until I actually looked at spectrograms for many ripples - itā€™s definitely lower than what I thought! But also ripples are so variable, in so many ways, even within the same individualā€¦
Iā€™m in Mattā€™s lab now and the manual annotation pipeline is amazing! And yes we have different auto detection methods but they are not super satisfyingā€¦
Building a shared annotated ā€œground truthā€ dataset for different species would be super useful! We have a former student who might even have some human data! Alsoā€¦ do bats have ripples?!
Show threadkari hoffmanJul 5, 2023
@elduvelle @LMPrida @cogneurophys they do! https://www.nature.com/articles/nn1829. Pigs too. Unlike hippocampal theta, the SWR seems pretty well conserved.
Hippocampal cellular and network activity in freely moving echolocating bats - Nature Neuroscience

The hippocampus is crucial for episodic and spatial memory. In freely moving rodents, hippocampal pyramidal neurons show spatially selective firing when the animal passes through a neuron's 'place-field', and theta-band oscillation is continuously present during locomotion. Here we report the first hippocampal recordings from echolocating bats, mammals phylogenetically distant from rodents, which showed place cells very similar to those of rodents. High-frequency 'ripple' oscillations were also rodent-like. Theta oscillation, however, differed from rodents in two important ways: (i) theta occurred when bats explored the environment without locomoting, using distal sensing through echolocation, and (ii) theta was not continuous, but occurred in short intermittent bouts. The intermittence of theta suggests that models of hippocampal function that rely on continuous theta may not apply to bats. Our data support the hypothesis that theta oscillation in the mammalian hippocampus is involved in sequence learning and hence, theta power should increase with sensory-input rateā€”which explains why theta power correlates with running speed in rodents and with echolocation call rate in bats.

Nature
Show threadEl DuvelleJul 5, 2023
@karihoffman @LMPrida Nice! Thank you for the ref - and the discussion! Whoā€™s nextā€¦ whales? šŸ˜
Show threadkari hoffmanJul 5, 2023
@elduvelle @LMPrida good question! Must do some researchā€¦