Sam Abbott

epiforecasts.ioReflections on two years estimating effective reproduction numbers – EpiForecastsOver the last two years we have estimated reproduction numbers daily for several thousand locations, presented these estimates as a curated data set and visualised them at epiforecasts.io/covid. In th...

13 Dec 2024

In 2022 @sbfnk.bsky.social and I wrote this blog on what it was like estimating the effective reproduction number daily for 2 years and if it was helpful epiforecasts.io/posts/2022-0...

Sam Abbott reposted

James Hay

@jameshay.bsky.social

Tracking West Nile virus dynamics using viral loads from trapped mosquitoes Very proud of this paper led by @punya-alahakoon.bsky.social in collaboration with Joseph Fauver at UNMC, combining agent based modelling, trapped mosquito data and pooled testing for WNV elifesciences.org/articles/111...

elifesciences.orgTracking West Nile virus dynamics using viral loads from trapped mosquitoesMosquito-pool Ct values reflect biologically meaningful variation in viral load, while integrating viral kinetics improves prevalence estimates and helps distinguish potentially infectious from non-in...

Sam Abbott reposted

The Ferrari Lab

@theferrarilab.bsky.social

19h

New Preprint: A classic property of epidemic time series is that they get increasingly stochastic and episodic at smaller spatial scales. Metapopulation theory says that dynamics that can't persist at the local scale CAN be rescued and stabilized at larger scales ... which is unbelievably cool.

samabbott.co.ukComparing methods for imputing missing symptom onsetsComparing methods for imputing missing symptom-onset dates, with a JavaScript module and a Python package.

19h

Another casual afternoon in LLM world with a package now existing in 4 languages...

Sam Abbott@seabbs.bsky.social

Had some fun with llms knocking up a little interactive and supporting packages to show various methods for imputing missing onsets under epidemic settings samabbott.co.uk/epidemic-imp...

samabbott.co.ukComparing methods for imputing missing symptom onsetsComparing methods for imputing missing symptom-onset dates, with a JavaScript module and a Python package.

21h

Had some fun with llms knocking up a little interactive and supporting packages to show various methods for imputing missing onsets under epidemic settings samabbott.co.uk/epidemic-imp...

github.combackcalc/report.md at master · epiforecasts/backcalcEvaluating approaches to backcalculating cases counts by date of infection from cases counts by date of report - epiforecasts/backcalc

21h

Related in spirit to this demo from 2020 but much fancier due to llm support. github.com/epiforecasts... (this was due to a period of the pandemic when I got weird about putting my name on stuff)

After quite a long time sitting as a development package epidist is now on CRAN. It extends brms with a framework for epidemiological extensions.

In general, I wish as a field this was a more common approach as it lets you have the generality/robustness of widely used tools alongside the tweaks we specifically need

And avoids the we could extend this simple demo tool to do X,Y,Z but won't, as its a PhD project or even worse here is 5 years of papers each adding a small feature that is already widely available outside of domain.

After quite a long time sitting as a development package epidist is now on CRAN. It extends brms with a framework for epidemiological extensions.

I am aware of a few users of the development package but please reach out if running into issues or finding feature gaps. Also, please cite this, as this kind of work is very, very, very hard to find funding for (in my experience impossible but can but try!)

Extending brms (and more generally widely available non-domain specific tools) has been something I have wanted to do for a long time so great to see this come together.

After quite a long time sitting as a development package epidist is now on CRAN. It extends brms with a framework for epidemiological extensions.

epidist.epinowcast.orgFitting to published estimates with the meta modelFitting to summarised, potentially biased, published estimates jointly with individual level data

Within the package are extensions for estimated delays accounting for common ID epi biases epidist.epinowcast.org/articles/epi... and a method for federated/meta analysis of potentially biased literature estimates. epidist.epinowcast.org/articles/met...

epiforecasts.ioReal-Time Case Fatality Ratio via a Mixture-Cure Survival ModelEstimates a real-time case fatality ratio (CFR) from line-list data using a Bayesian mixture-cure survival model, registered as an epidist model type so the CFR and the onset-to-death delay both take ...

The framework has also been used by @sbfnk.bsky.social to estimate Case fatality ratios epiforecasts.io/cfrnow/index...

Sam Abbott reposted

JuliaCon 2026

@juliacon.org

SpeedyWeather.jl: Towards a differentiable and GPU-capable general circulation model — Maximilian Gelbrecht, Milan Klöwer https://youtu.be/STMDrYShKYA

youtu.beSpeedyWeather.jl: Towards a differentiable and GPU-capable general circulation modelThe Julia Programming Language

Sam Abbott reposted

Billy Quilty

@bquilty.bsky.social

Will be interesting to see what "sticks" from this vibe-coded software revolution. It's clear that you can get to impressive ends really easily (just don't think about the means too much), but maintaining that is much more difficult (see all those "games in a browser" that worked - for about a day).

epiforecasts.ioDashboard | BVDOutbreakSizeDocumentation for BVDOutbreakSize

New release of our external modelling of the BVD outbreak in the DRC which uses public Sitrep data only and is experimenting with very LLM augmented infectious disease outbreak modelling. This release adds health zone level estimates. epiforecasts.io/BVDOutbreakS...

In a follow up I also aim to tackle spatial mixing more deeply as this is curently a modulated gravity model for both provinces and then within health zones. The two options I am exploring are increasing the flexibility of the modulation or backing the central estimate to be flowminder based.

See the Github for further issues and in progress work and modify.

epiforecasts.ioDashboard | BVDOutbreakSizeDocumentation for BVDOutbreakSize

The model code and analysis were drafted by a language model, then reviewed and revised under human oversight. The named authors are responsible for that oversight.

Note in this release I see a spike in Rt just before detection. I don't think this is real and I think is driven by an artefact in the symptom onset data (see the Rt estimates by data set for this). There are likely other issues that I will aim to deal with in future releases.

epiforecasts.ioDashboard | BVDOutbreakSizeDocumentation for BVDOutbreakSize

epiforecasts.ioLimitations | BVDOutbreakSizeDocumentation for BVDOutbreakSize

As in our previous work this work has a lot of limitations: epiforecasts.io/BVDOutbreakS... It is also essentially an unfunded side project of mine with all the limitations that brings both in focus and resources.

This work adds an external view of the current situation, drawing on our understanding of real-time infectious disease dynamics and the infection process behind the observed surveillance counts. We are developing it and encourage feedback, so please get in touch. We support reuse and adaptation.

epiforecasts.ioDashboard | BVDOutbreakSizeDocumentation for BVDOutbreakSize

This release contains quite a bit of refactoring and attempts to make tools for the Bayesian workflow accessible. It continues to run in Github CI and so that constrains what can be done. Here we have added a Markov-Melded health zone patch model to give lower-level outputs.

cdcgov.github.ioStratified models - AlgebraicEpiModels

Looks pretty great: cdcgov.github.io/AlgebraicEpi...

Sam Abbott reposted

Andrei R. Akhmetzhanov (吳亞克)

@aakhmetz.bsky.social

Replying to @aakhmetz.bsky.social

Data, code and all Stan models are on GitHub, with a link to the full model output: github.com/aakhmetz/Bun...

Sam Abbott reposted

IDDjobs

@iddjobs.org

Postdoc (London, UK) Develop new methods for pathogen genomic epidemiology with @ceciletk.bsky.social at Imperial College London @mrc-outbreak.bsky.social More details: http://iddjobs.org/jobs/2596

journals.plos.orgBest practices for estimating and reporting epidemiological delay distributions of infectious diseasesEpidemiological delays are key quantities that inform public health policy and clinical practice. They are used as inputs for mathematical and statistical models, which in turn can guide control strat...

A few years ago we (Kelly Charniga @sangwoopark.bsky.social et al.) did some thinking on how we should estimate and report epidemiological delay distributions (i.e incubation periods etc. etc.). journals.plos.org/ploscompbiol...

medrxiv.orgHow bursty infectiousness shapes epidemic dynamicsAn epidemic’s expected course is determined by the magnitude and timing of a typical person’s infectiousness — captured, in turn, by the basic reproduction number and the generation-time distribution....

Who also has this really nice preprint (which I am honestly still digesting) that looks like it could have viral loads shoved into it as a bursy mechanism.... bsky.app/profile/skis...

Stephen Kissler@skissler.bsky.social

I'm excited to share a preprint that's been years in the making: the idea took shape during the COVID pandemic. Anecdotally, people would sometimes spread COVID at work but not at home, and vice versa. Why? One explanation would be if infectiousness is fleeting...

arxiv.orgInferring infectiousness: a joint model of the within-host viral kinetics of SARS-CoV-2

i.e. "This decomposition mirrors within-host viral dynamics: stochastic viral replication early in infection creates variation in the onset of detectable virus (the latent period), while subsequent shedding kinetics are more consistent across individuals (the burst)"

Sam Abbott reposted

Stephen Kissler

@skissler.bsky.social

Some really nice recent work linking these ideas from Christopher Boyer, Stephen Kissler and @mlipsitch.bsky.social arxiv.org/html/2605.20...

skissler.bsky.socialStephen Kissler (@skissler.bsky.social)Infectious disease epidemiologist / computer science professor @ CU Boulder. Trying to figure out how to keep us all healthy and fed.

woops skissler.bsky.social not doing a good job of finding people today!

An obvious question from all of this and one we are still exploring is what happens when you jointly model with viral load data. That is still in the works but for now here is the impact of different prior choices (i.e. informed by viral load data and not).

This has been a big collaborative effort with a huge amount of work from Nyall in particular. Some cool follow ups to this to come including thinking about applying the idea to othe diseases and joint fitting with viral load data. Please reach out if interested in being involved!

yes you can recover true parameters under simulation

We also found that in a few case studies a viral load based distribution fit the data better than the distribution used as the best fit in the original study and that this changed the estimated parameters in ways that might have impacted downstream modelling and hence decisions.

So we did! www.medrxiv.org/content/10.6...

medrxiv.orgmedrxiv.org

The TLDR is that yes you can do this, yes it is fairly elegant (in my opinion), yes you can connect this specialised distribution to common ones with different assumptions,

In conversations (via Epiengange part of the CDC funded insight net project) Nyall, Laren Meyer and I realised that maybe if we thought about the generation time (and its component distributions) as being driven by within host processes in the same way Nyall had been thinking about the incubation

period we could develop an epidemiologically motivated model for the serial interval, generation time, and all its component delays.

Coming back to the Charniga work it would have also enabled us to give some advice beyond try a few things and cross your fingers (i.e your a distribution motivated by the within host mechanisms of your disease).

So nothing else to do? Not quite. Alongside all of this a big gap that remains in using common outbreaks models is understanding what the generation time (i.e time from infection to infection) is and how to model it.

As someone who doesn't love Burrrs I did not initially find this that exciting but after reading more closely I realised that the main motivation for the modified Burrr here was an argument that the incubation period is an emergent property of within host viral kenetics which got me very excited!

journals.plos.orgCombined analyses of within-host SARS-CoV-2 viral kinetics and information on past exposures to the virus in a human cohort identifies intrinsic differences of Omicron and Delta variantsThe emergence of successive SARS-CoV-2 variants of concern during 2020-22 created a need to understand the drivers of such growth. This study uses a Bayesian model to reveal how a set of key covariate...

Why? Well we ( Timothy Russell, @adamjkucharski.bsky.social and co) had just been modelling viral kenetics jointly with sypmtom onsets (assuming a lognormal distribution) so a distribution that was viral load motivated would have allowed for very elegant linkage. journals.plos.org/plosbiology/...

The lack of mechanism in that approach is potentially problematic both in that it will struggle under sparse data but also that it can struggle in settings with compound and complex delays such as generation times.

journals.plos.orgThe Burr distribution as a model for the delay between key events in an individual’s infection historyAuthor summary In public health, it is important to know key temporal properties of diseases (such as how long someone is ill for or infectious for). Mathematical characterisation of properties requir...

At about the same time as we were doing our thinking Nyall Jamieson, Ian Hall, and co were thinking deeply about how much they loved Burrr distributions for modelling incubation periods. journals.plos.org/ploscompbiol...

A lot of that work solid and I hope has moved the field forward a bit. However, there was a gap when it came to how to select a distribution for fitting to your data. We said try a few, report all of them, and give an indication of which fits best (via best fit metrics etc. etc.)