Chris Mainey

Why is my model so bad?: Tales of a wandering PhD student

George Box is often credited with the phrase ‘…all models are wrong, but some are useful.’ Whilst this makes sense to me now, it was frustrating in my early career, when I didn’t know if I was wrong, my model was the wrong choice, I coded it badly, or if it was working as well as it could. This talk summarise part of my PhD work, attempting to build predictive models on NHS Incident Reporting data in R. Firstly, the data quality was terrible. This is to be expected when you understand the context in which it was generated, and I’ll discuss data generating mechanisms and some of our assumptions about them. To get around my poor data, I ‘borrowed strength’ from other datasets and used them at the same level of aggregation. This became a ‘panel’ dataset, and I was modelling count data. I will discuss our default assumption of Poisson regression (GLM), why it is not perfect with real-word data, and what overdispersion is. There are a variety reasons for overdispersion, and my models proceeded by testing different sources and assumptions including: clustering, aggregation, and ‘noise’ from poor data quality. I will discuss methods for fitting these assumptions including GEE, GLMM, Boostrapping / Cross-validation, GAMs with smoothers, and a Random Forest thrown in for good measure. The ultimate question here is: ‘is my model any good?’ I will briefly discuss how we can assess predictive model accuracy, and a few different schools of thought on it. All examples will be presented with a why, and a how, including R code made available on GitHub. I will focus both on the how I chose and learnt the methods, as well as practical tips for analysing count data with overdispersion.

Coffee

Coffee

Emma Vestesson

R for personalised medicine

Getting the dose right for chemotherapy drugs is important but can be hard. A dose that is too low is not enough to kill the cancer but a dose that is too high can result in severe side-effects. In addition to this, knowing when the drug has left the body allows clinicians to know when it safe to discharge a patient. Estimating the best dose for a patient also depends on factors such as age, sex, weight and kidney function. What happens to the drug in the body can be estimated using non-linear mixed effects model but these are complex models that can’t be calculated by hand by a clinician. In comes R Shiny, a way to give clinicians the power of modelling but without having to write any code. I’ve built a shiny app that predicts when the drug has left the body that I hope will change clinical practice. In my talk I will discuss some of the technical challenges I encountered and some nifty packages that helped me along the way. I will also cover design choices I had to consider when trying build an app that is useful and most importantly safe within a hospital setting. This includes technical problems such as making sure that old inputs and calculations are cleared when new patient data is entered or making a dynamic input widget based on user input. It also includes working with pharmacists when designing the user interface and what outputs where the most useful to them. Finally I will discuss the challenge of deploying an app within the NHS and what to consider when working with patient data. This talk is for anyone who wants to move from basic shiny apps to using R in production.

Julia Silge

Understanding Word Embeddings

Modern natural language processing frameworks (including word2vec, GloVe, fastText, ULMFIT, and more) depend on word embeddings, a way of statistically modeling language where words or phrases are mapped to vectors of real numbers. In this talk, understand word embeddings by investigating how we can generate them using count-based statistics and dimensionality reduction, then learn how to make use of pre-trained embeddings based on enormous datasets. Finally, explore the ethical issues involved in using word embeddings and how they can amplify systemic and historical bias.

Kenneth McLean

ImpactR: Deriving and quantifying impact of biomedical research publications.

Introduction: Demonstrating the impact of research is essential to evidence its value to stakeholders (whether the public, peers, research or funding organisations, etc). Furthermore, although traditional (citation- or journal-based) metrics of research impact remain highly regarded, the scope of research impact has broadened to emphasise engagement and direct patient benefit. While individual packages are available to explore the impact of research in R, these are often not designed for integration or optimised for further analysis. Methods: The ImpactR package aims to consolidate functionality from a range of existing packages (including RISMed, rcrossref, rAltmetric, scholar) into holistic and useful quantifications of biomedical research impact. Essential publication information and research metrics (altmetric, citations, journal impact factor) can be extracted from online repositories into a single dataframe using either PubMed identification numbers (PMID) or Digital Object Identifiers (DOI). This can either be used directly for data analysis or visualisation, or passed to further functions to provide a formatted citation list (“cite_publication”); additional altmetric (“impact_altmetric”) and citation metrics (‘impact_cite’); or data on co-authorship networks (‘impact_auth’).Applications: The ImpactR package facilitates research groups and/or academics to better understand and demonstrate the impact of their research to stakeholders. These functions have a range of applications from generating publication lists for curricula vitae, or quantifying research impact through traditional metrics (citations, journal impact factor), authorship networks, or social media. However, it should be emphasised this package does not encompass all measures of impact from a research output, including direct patient benefit. The package is available on GitHub (https://github.com/kamclean/impactr) alongside vignettes on its use, and is designed to be user-friendly and applicable to all publications in biomedical journals with a PMID or DOI.

Lunch

Mine Çetinkaya-Rundel

R 4 All: Welcoming and inclusive practices for teaching R

With the rising interest in data science and programming, those of us teaching (with) R can now reach a larger audience than we ever thought possible. As educators, it is our responsibility to ensure that while we are building interesting and challenging curricula for these students, we also do it in a way that is attractive and engaging for a diverse audience as well as supportive enough to minimize the number of students who fall through the cracks. Adopting welcoming and inclusive practices can enable these students, whatever their background and circumstances, to achieve their potential and grow and engage with the larger R and data science community. In this talk, we highlight a collection of pedagogical considerations, tips, and tricks for designing a welcoming and inclusive curriculum for teaching (with) R. In addition, we demonstrate tooling and infrastructure solutions for making it as straightforward and painless as possible to put these approaches into practice in the classroom.

Newcastle University

Pablo De Juan Bernabeu

Open Data and Reproducibility v2.0

Scientific articles have been including more and more supplementary materials, which are ever more advanced. This has been facilitated by burgeoning technologies and available training. Another contributor was the replication crisis in psychology. As a result, nowadays, researchers customarily examine the data and code for studies, and the same may be done by members of the public, too. Standard materials such as data sets and code scripts could be referred to as Version 1.0. Excitingly, software developers and contributors have released Version 2.0 already! Let’s look at three interconnected applications. 1. Binder environments — Enable public access to your data and code in RStudio, directly on the Internet. We’ll look at requirements and further possibilities, such as hosting Shiny apps. 2. R Markdown documents — Set your input and output in stone with R Markdown documents. They may be enriched with website features (CSS, ‘pagedown’), and published as websites and PDF or Word documents. 3. Online data dashboards — A further step in open data is offered by online, interactive dashboards. These all-reproducible visualisations are published as websites, and therefore can include hyperlinks and downloadable files. Some of the R packages used are: knitr, reactable, ggplot, plotly, rmarkdown, flexdashboard, shiny. The aim is to look at different forms of dashboards (Flexdashboard, Shiny and Flexdashboard-Shiny), and the suitable platforms to host them online (personal websites, RPubs, Binder, Shinyapps and custom servers). A great thing is that dashboards may be made very simple, but they also offer the possibility of notable improvement. Real examples will be shown, common challenges flagged, and further resources provided. Background: https://github.com/pablobernabeu/Data-is-present/blob/master/README.md

Registration

Russ Hyde

Your Code As Data

The code you write this month won’t stay still. You’ll eventually want to add extra features to that package or extra analyses to that project. You might find a bug or try to speed things up. Your tests might start to fail when dependencies change or your report might not build on a different system. As it evolves to accommodate these changes, it is worth reappraising your code. R has a range of tools to help assess your current code-base encompassing static analysis (style / syntax / structure: lintr, goodpractice, cyclocomp, pkgnet) and dynamic analysis (test coverage / benchmarking / profiling: covr, bench, profvis). For version-controlled projects, the git history can provide an archaeological and social view of the code base that complements analysis of the code itself. In this talk (inspired by Adam Tornhill’s book ‘Software Design X-Rays’), I show how existing R tools for working with git can be combined with code analysis tools to provide a rich dataset about your code.

Theo Roe

Improving the Communication of Environmental Data Using Shiny

Countries are big. Rivers are long. Countries have a lot of rivers. The government-run environment agency has a wealth of data relating to the health of rivers in the UK. The problem being that it’s difficult to explore the data and thus hard to build any tools using the data. We’ve built a Shiny app that allows for visual exploration of the river health in the UK. We want to talk about some of the harder parts of the the data prep and the cool parts of the Shiny app.

Welcome

Close

Zoë Turner

Using R in Mortality Surveillance within a mental and general health secondary care NHS provider

In 1991, the so called ‘Angel of Death’ was convicted of the deaths of 4 children, attempting to murder 3 other children, and causing grievous bodily harm to a further 6 who had been inpatients at Grantham and Kesteven Hospital, Lincolnshire. In 2000, Harold Shipman was convicted of the murder of 15 patients with a later estimate putting the figure at 215 or more. In 2015, Victorino Chua was convicted of the murder of 2 patients and poisoned as many as 19 other patients at Stepping Hill Hospital, Stockport. Of the 190,000 patients seen every year at my NHS Trust around 4.7% have a recorded death. How do we know that these deaths are natural, unavoidable and not related to major systemic problems or even murder? Using data collected from 6 clinical systems, I: 1) use an RMarkdown report to show seasonality and age patterns of deaths; 2) track variation by using qicharts2 (based on ggplot2) for SPC (statistical process control charts) and Pareto charts; 3) use dplyr to get the details of the specific patients behind the points for focused checks.