Dear ChatGPT – can you teach me how to program an app for laboratory medicine?

Introduction

Historically, the fields of laboratory medicine and information technology have been closely intertwined [1]. Digital progress has been a driving force behind the digital transformation towards Lab 4.0 [2]. With an increasing focus on digital solutions, the need for expertise extending beyond the core competencies of laboratory medicine becomes increasingly prominent [3].

Thus, according to the literature, future laboratory professionals require not only a fundamental understanding of technology and basic digital skills but also programming abilities [4]. Correspondingly, the Association for Diagnostics and Laboratory Medicine issued to ‘Embrace the R Programming Language’ as a ‘Gateway to Laboratory Medicine’s Digital Future’ as early as 2020 [5].

However, among the ‘digital native’ ‘Young Scientists’ in laboratory medicine, only 23 % program daily in R and a mere 6 % in Python. Despite 96 % of young laboratory medicine scientists recognizing the necessity of digital education, only 20 % receive it. This reflects a disparity between the availability and demand for digital education in this field, with a call for learning resources tailored to various knowledge levels [4]. In this context, ChatGPT’s capabilities in code generation, optimization, debugging assistance, code documentation, and review, coupled with its accessibility and ease of use [6], present a potential opportunity for novices to develop programming skills [7, 8]. This case study, to the best of the authors’ knowledge, is the first to explore the application of ChatGPT in programming a laboratory medicine application by non-programming-experts, aiming to further bridge the ‘digital skills gap’ in laboratory medicine [4].

Methods

Results

Over the course of two weeks, we exchanged 135,615 words with ChatGPT, which facilitated the development of an application that met our intended specifications.

ChatGPT compensated for our lack of skills in areas such as extracting data from PDF files and creating user interfaces (Figure 1). It provided multiple solutions to open-ended questions. Knowledge gaps were bridged by the easy-to-understand code annotations and explanations provided by ChatGPT, concurrently improving our programming skills.

Figure 1:

User interface of the developed application for assessing the intra-assay, inter-device precision of immunophenotyping, Cologne 2023. Users can select folders and parameters for analysis through a drop-down menu, download extracted values in CSV format, and obtain inter-assay precision results as a PDF containing a regression plot with a correlation coefficient.

However, implementation without basic knowledge of programming languages and the structure of R would not have been feasible, even with the support of ChatGPT.

ChatGPT frequently produced erroneous code in response to vaguely formulated user input (“prompts”). We often had to experiment with the wording and language, switching from German to English, until ChatGPT correctly interpreted our intentions and translated them into code. Even with an adjusted input formulation, ChatGPT’s execution of solution strategies was often inadequate, requiring the development of new strategies for sub-problems within the process. In general, the code snippets written by ChatGPT often generated error messages, which in most cases could be resolved with the help of ChatGPT and further research into the vignettes (Table 1).

The authors are currently in communication with the IT department of the University Hospital Cologne to enable routine use of the shiny app in the future.

Discussion

Programming expertise is becoming increasingly vital in the daily practice of future medical laboratory professionals. Thus, to bridge the “digital skills gap” in laboratory medicine, the development and research of potential learning resources for individual skill acquisition are essential [4].

In the wider landscape of software development, social media users are already leveraging ChatGPT’s capabilities for developing and debugging code across 10 different programming languages [9]. This utility underscores ChatGPT’s superior problem-solving capabilities, as in Python, where it outperforms counterparts such as Bard and Claude [10], and demonstrates robustness in addressing tasks of varying complexity in Java and C++ [11].

This research further extends this narrative by demonstrating the suitability of ChatGPT for aiding the development of functional laboratory medicine applications by non-programmers. This is consistent with existing literature on the utilization of ChatGPT for code generation in pharmacometrics [7] and medical statistics [8], where it has enabled users with minimal programming knowledge to write code and develop operational programs [7, 8], indicating its potential applicability across various disciplines.

Furthermore, this study underscores ChatGPT’s cross-linguistic ability to generate code, a capability further emphasized by the diverse linguistic focus in prior studies [7, 8]. However, the observations indicate a preference for English inputs, a tendency probably stemming from the predominance of English in ChatGPT’s training dataset [12]. This linguistic inclination is consistent with trends noted in other domains [13], including medical exams [14], suggesting a broader pattern of language bias.

In harmony with findings from multiple studies [7, 8, 10, 11, 15], the necessity of repeated input refinement is critical for deriving accurate and operational code from ChatGPT also within the domain of laboratory medicine. Moreover, observations regarding the inconsistency in ChatGPT’s output for identical prompts are reaffirmed, highlighting a deficiency in reproducibility [7]. In addition to concerns regarding reproducibility, reliability, and accuracy, ChatGPT’s code-generation capabilities are also not immune to the more general criticisms of this type of Artificial Intelligence, including ethical concerns, over-reliance, and security risks [6]. For instance, over-reliance on ChatGPT’s programming capabilities might inhibit critical thinking and hinder the development of individual programming skills [6, 16], thus leading to “deskilling” and “automation-bias” [17]. In turn, inadequate proficiency in code reading and interpretation present a security hazard [6], with the potential for sensitive data leakage [18] and discriminatory algorithms [9] to pass undetected. This concern is especially pronounced in sensitive domains such as healthcare, where the integrity of patient data and network security are critical [19]. Consequently, instead of relying solely on automated code generation, adopting a cautious strategy alongside a basic grasp of programming principles is essential for ensuring safe usage.

Contrary to these concerns, research by Kazemitabaar et al. highlights that ChatGPT does not adversely affect novice programmers’ ability to modify or generate code [20]. In fact, it has been shown to improve performance, boost self-efficacy, decrease frustration, and promote skill retention over time [20, 21], thereby positioning ChatGPT as a potential educational aid and companion for novice programmers [16].

Overall, while ChatGPT holds the promise to streamline code development and debugging, as well as enrich educational experiences for learners, its outputs must be rigorously monitored and evaluated, especially within data-sensitive fields like healthcare. Therefore, ChatGPT should only be used as a complementary tool in laboratory medicine by users with basic programming knowledge. Future studies are needed in this context to investigate possible integration into corresponding academic training programs.

Conclusions

Overall, it is evident that ChatGPT is capable of assisting individuals with limited coding expertise in writing laboratory medicine programs using R. However, in light of valid criticisms regarding its accuracy and reliability, as well as concerns pertaining to security, over-reliance, and ethical implications, the outputs generated by ChatGPT should be subjected to rigorous scrutiny.

The easily comprehensible explanations and annotations provided by ChatGPT underscore its potential to support digital education in the field of laboratory medicine. Therefore, future research focusing on the successful integration of ChatGPT into academic digital education programs appears to be a worthwhile endeavor.

Learning points

1. ChatGPT can aid non-experts in programming medical laboratory applications.

2. Through clear explanations and detailed connotations, ChatGPT can help develop the programming skills of medical professionals.

3. ChatGPT’s code often generates error messages that can only be partially solved by input reformulations.

4. Due to justified criticism regarding reproducibility and accuracy as well as ethical and safety concerns, ChatGPT should only be used by trained personnel for programming support.