AI-Assisted Authoring of an Open-Access Portuguese Biostatistics Book in Quarto and R

Hello everyone. I'm Andre Banel. I'm a professor at Ceará University. And here to deliver a talk AI system outcome of open access biostatistics book in R. So at the university I teach, I teach a course in biostatistics and we teach R for health science students. So most students don't have any practical programming skills, so they need to learn from scratch. And we have a gap where students need a good statistical foundation, practical programming skills, and real-world health health data examples, but we also miss open access resource in Portuguese. Uh we also need examples in Brazilian health data. And in we also need a personalized book for the course. Uh if you would write a book from scratch, it would take more weeks or months of full-time work. And our objective to see if AI could assist in that process. So the book is called Biostatistics Aplicada - Saúde e Nutrição. And in English, Applied Biostatistics in Health using R. This book teaches uh from introduction to R and RStudio until data wrangling in Tidyverse, exploratory and data visualization, hypothesis testing, uh regression models, also teaches systematic analysis and text mining using LLMs. So this book was built for promotion at promotion in the problem at the Cesumar University. Uh so, what we used to write this book? We used scripts from used at the course. So, the scripts were already made and classroom presentations. So, the PDFs on the lesson. And we all wrote the data sets also that were used. And our workflow we used existing R scripts with the PDFs. We had a similar AI tool called code. We didn't talk the book. The AI uh drafted the book and the we authors corrected and reviewed all both the code and the text in the book. To make it more interesting to use the health uh care examples using real data from the Brazilian public health system. In Brazil, we have a a data source called DataSUS. It's from the Brazilian Ministry of Health and the open health records platform. Uh one of the resource service uh you can download in this uh from DataSUS is called SIH. It's a hospital information system. So, it's admission and patient records from Tariff Brazil. And we use from uh the city where our university's located is in Maringá, Paraná, Brazil from the year 2024. All this data was downloaded with the microdata source R package. It's is that was made uh make it easier to download data. So, all AI uh architectures that used to uh assisting authoring used Cloud Sonnet uh for 0.5 uh 0.6. It's not the best model, but it also worked very, very good. Cloud Code is a CLI uh agent called do. For our IDE is Positron. I teach we use RStudio, but for coding the repository. And for our free hosting website is GitHub Pages. So, Cloud Code accelerated the products uh code examples. It also used Quarto for the book structure. So, it formatted the all the code and scripts uh using Quarto. And our workflow consisted I defined the objectives for the book. Cloud Code written the uh scripts and PDFs. Drafted the code and the prose for uh the book. I reviewed the and drafted the code. Tested with real world data, and this was many directions until the book was publication ready. Ready. So, where AI helped the most? They have a lot in part chunk of information options. Uh boilerplate code, for example, is a plot or some other R packages like for data wrangling using tidyverse. It can also help to maintain consistent code styles across chapters. The structure of the scripts also into prose was uh very good using uh AI. And for the Portuguese person there. So, it made my teaching materials and I can allow to improve and create the book. But also, there's a lot of domain expertise was needed. For example, statistical interpretation. You need to verify uh what the AI would make of some statistical results. The clinical context. For example, a lot of observations in Brazil, there's ICD code finishing >> [clears throat] >> uh the last digit is nine. So, there's a non-specified disease. It's a characteristic of our health care system. So, if you don't know this, you don't have this context, it makes interpretation hard. For example, analyze some cancer. And all the cancer was not certified for example for location on the body. Also, what was needed is sequence, a logical sequence for the PhD student so they could know what they could learn first. But there was there a lot of challenges along the way. For example, the CI/CB pipeline had a lot of memory failures and caused some Like exactly one uh they had some problems compiling all of the code. Uh there's isn't enough memory in GitHub for using our free hosting book. Also needed a lot of manual configuration of the system and a lot of workflows iterations before I could reach a stable version. We also needed acting expertise. So, to configure the CLI tool, we need to install Node.js, the same party scripts, the stand on the Git and GitHub actions. And now also a lot of the content required significant manual revisions. So, even because the book is is drafted, we need to correctly review all the the code and the text. Here's a QR code, so our viewers can see the book. It's in Portuguese. I plan to add English translation. And here I propose a robotics model. So, this is very important for non-English speaking research communities. Uh so, for example, an individual who has subject matter expertise, it can use it existing teaching materials, open data and required tools, and AI assisted drafting, for example, our code of subject matter tool, and create open educational resources in languages. That's a very very interesting uh opportunity we have now. Some languages don't have a lot of content, and you can make it uh books that are specific to our domains and are very uh personalized for our course. That's what I wanted to share. I love our medicine thousand thousand uh six, and thank you for this opportunity.