The potential of freely available Artificial Intelligence tools in cardiology-related diagnosing based on medical letter information

Abstract

This study assesses the diagnostic capacities of four top ranking AI chatbots in Romanian medical letter interpretation: Gemini, Copilot, DeepSeek, and Qwen. Extending earlier studies on AI interpretation of Holter investigation images, these tools’ capacity to produce credible diagnoses from medical texts, with and without images from Holter investigations, were analyzed. In order to rate diagnostic predictions, a weighted scoring system was put in place and performed five repetitions per instrument version and case for a more rigorous approach. The findings show that all four AI systems can identify key diagnostic elements that often match the real medical diagnoses, though with varying degrees of accuracy and consistency ranging from 81.2% to 92.44%, with Qwen achieving the highest consistency after incorporating imaging data. Response times varied significantly (31-84 seconds), with Gemini demonstrating the fastest average response. The study highlighted the importance of prompt engineering, as structured prompts with clear instructions produced more organized and relevant responses. Some limitations were identified, such as occasionally providing contradictory diagnoses and not always recognizing certain diseases. These findings suggest there is no universally superior AI tool for cardiology diagnosis, as optimal performance depends on the specific clinical presentation and input format. The multidisciplinary team, which included a cardiologist and computer scientists, made review possible from two points of view: medical accuracy and consistency across multiple test cases. The research aims to provide valuable insights for healthcare professionals regarding the promising capabilities and current limitations of freely available AI systems in medical contexts.

Ähnliche Arbeiten