Applications of generative artificial intelligence to augment clinician’s capability for medical data analysis in RStudio

Abstract

Dear Editor, In the ever-evolving healthcare landscape, the integration of artificial intelligence (AI) has emerged as a transformative force, offering unprecedented opportunities to enhance clinical decision-making and patient care. While traditional AI is a rule-based, preset algorithmic program, among the diverse array of AI techniques, generative AI (GenAI) is an innovative technology capable of producing new and original content – a promising tool for augmenting clinicians’ capabilities in analysing medical data.[1-3] RStudio [RStudio Team (2021) RStudio: Integrated Development for R; RStudio, Inc., Boston, MA, USA] is a robust integrated development environment (IDE) designed specifically for statistical computing and data analysis in the R statistical programming language [R Core Team (2020), R: A language and environment for statistical computing; R Foundation for Statistical Computing, Vienna, Austria], a programming language popularly used by statisticians for handling massive amounts of data. As an open-source IDE, RStudio provides a user-friendly interface and a comprehensive suite of tools to facilitate data manipulation, visualisation and statistical modelling. However, the primary drawback and barrier to using an IDE is the requirement for knowledge of coding in the R programming language. Leveraging the GenAI capabilities of Chat Generative Pre-Trained Transformer (ChatGPT) in tandem with Rstudio can allow a clinician with a basic understanding of RStudio to create intricate and interactive graphical content to visualise data effectively and efficiently. There are two ways in which ChatGPT can aid clinicians in analysis in RStudio: one, by generating an appropriate code in the R programming language based on prompts provided by the clinician and second, by analysing and suggesting additional corrections in a code written by the clinician as the prompt. For example, if a clinician desires an analysis of the correlation between multiple factors and an outcome, a one-sentence prompt, ‘Create a correlation plot in RStudio using the circle as shape’, in ChatGPT would provide the corresponding R programming code for generating correlation plots [Figure 1]. Similarly, ChatGPT can also be prompted to modify the same (or different) code further to yield customised graphs [Figure 1b–e]. As can be observed, a single-line prompt allows one to generate the entire program in R for analysis [Figure 2].Figure 1: Creating a correlation matrix between various demographic factors and clinical parameters of patients admitted to ICU. (a) A simple command ‘Write a code for correlation matrix for various factors in patients admitted to ICU’ generates this output in ChatGPT. (b) Entering the same code in RStudio along with the data set yields this correlation matrix in which the positive correlation is presented in blue, with the darker colour and larger size of the circle representing a stronger positive correlation, and the negative correlation is presented in red, with the darker colour and larger circle representing a strong negative correlation. The corresponding numerical correlation coefficients are also depicted in the graph. (c) The resultant modified graph after the code for displaying numerical correlation coefficients is deleted from the previous code. (d) Another modification of the same correlation matrix after replacing ‘circle’ for ‘square’ in the code. (e) Correlation matrix using ‘ellipse’ as the shape. ChatGPT=Chat Generative Pre-Trained Transformer, ICU=intensive care unitFigure 2: A few examples of the RStudio output using the codes generated by ChatGPT. The codes generated by ChatGPT with colloquial English prompts and the graphical output obtained: (a) Use of ‘ggplot2’ and ‘reshape2’ packages to generate heatmap to represent the PaO2/FiO2 ratio of 10 patients for the first 10 days of ICU admission. Colour coding the data represents improving or deteriorating trends of the patients in terms of PaO2/FiO2. The ‘reshape2’ package was used to arrange the data in the long format, and the ‘gglopt2’ package was used to create the graph, both suggested by ChatGPT. (b) Bland–Altman plot for studying the agreement between propofol plasma concentration as estimated by a target-controlled infusion pump and that measured directly from a plasma sample. (c) Violin plots of a group before and after an event with a trend line joining the corresponding data points – SOFA scores before and after the incidence of barotrauma in patients in an ICU. (d) Grouped boxplot for pulmonary function tests in two groups of patients. The data-specific names of the data file and various parameters have been entered in each code obtained from ChatGPT. ChatGPT=Chat Generative Pre-Trained Transformer, ICU=intensive care unit, SOFA=Sequential Organ Failure Assessment, PaO2/FiO2=partial pressure of oxygen/fractional inspired oxygenAnother way to utilise these large language models (LLMs) is to learn programming. If a clinician types a program code, such as an R program, as a prompt, LLM technology can return suggested corrections. LLMs are deep learning AI techniques that recognise, process and generate natural human language.[4] These capabilities enhance efficient learning and help optimise the input to RStudio. One can also use these models to understand programs written by others easily. These tools may be used in daily practice for enhanced efficiency and productivity.[5] However, caution must always be exercised when interpreting the output generated by such technologies to avoid inherent biases or inaccuracies. The training methodologies associated with these models make them susceptible to such biases and errors[6] as the quality of the input data is vital for their effectiveness, and since much data goes into the training, fact-checking each piece of data is complicated and sometimes not feasible. Nevertheless, in conjunction with human intellect, these technologies can augment one’s capabilities and unleash unprecedented productivity, efficiency and effectiveness. Financial support and sponsorship Nil. Conflicts of interest There are no conflicts of interest.

Ähnliche Arbeiten