A-021 Analysis of troponin turnaround times using machine learning and artificial intelligence

Abstract

Background: Acute Myocardial Infarction (AMI) can lead to significant cardiac tissue damage and mortality, making rapid identification and diagnosis a critical component of patient care and minimizing cardiac myocyte necrosis. High-sensitivity Troponins (hs-TnT and hs-TnI) are currently utilized as the primary clinical assay for the rapid diagnosis patients in the acute setting. Thus, the turnaround time (TAT) for high-sensitive troponin assays need to be as short as possible to facilitate detection or rule-out of an AMI. In this study, we utilized machine learning and artificial intelligence (AI) to evaluate whether changes in instrument workflows (e.g. increasing the number of measuring cells) or additional test ordering factors influence in-lab TATs for hs-TnT at a multi-site tertiary care healthcare institution. Methods: In laboratory hs-TnT TATs were collected across a 12-week period (6 weeks prior to the addition of a measuring cell and 6 weeks post-implementation). Troponin T results (n=10,297) were deidentified and placed into a data set containing test result, time and date sample was received at the laboratory, ordering clinic, time and date hs-TnT results were resulted in the LIS, post-measuring cell implementation, hs-TnT order code (e.g. random draw vs timed draw). These results excluded critical values of hs-TnT which require a provider readback, as in lab TATs were the primary metric that was being evaluated. Guided by ChatGPT-4, statistical analysis and subsequent python script and R-code was generated to evaluate trends in hs-TnT TATs. Results: The difference between TATs before and after the implementation of additional measuring cells was not determined to be statistically significant (p-value: 0.321). However, other patterns that impact hs-TnT TATs were identified by machine learning and AI. This enabled further investigations into additional variables that correlated with delayed hs-TnT TATs. Upon further refinement, statistically significant differences were observed by what hs-TnT order code was performed (p-value < 0.0001) by ANOVA. Conclusion: The utilization of ChatGPT-4 enable rapid analysis and pattern recognition for clinical laboratorians. Importantly, the generation of a clean dataset was instrumental in guiding ChatGPT to perform the appropriate analyses. The identified patterns which increased TAT highlighted areas of improvement to the workflow processes in the clinical laboratories and the broader institution in general. As advances in machine learning and AI evolve, clinical laboratorians should be aware of both the advantages and risks of utilizing this approach to solve complex problems.