본문 바로가기
  • Home

A Case Study of ECG Analysis Using LLM-Based Vibe Coding

  • Industry Promotion Research
  • Abbr : IPR
  • 2026, 11(4), pp.131~139
  • DOI : 10.21186/IPR.2026.11.4.131
  • Publisher : Industrial Promotion Institute
  • Research Area : Interdisciplinary Studies > Interdisciplinary Research
  • Received : September 14, 2026
  • Accepted : September 28, 2026
  • Published : October 31, 2026

Hyun Mok Park 1,  Woojin Cho 1,  Seok-Oh Bang 1,  Hyun-Seok Choi 1,  Ki-Taek Kwon 1,  Sang-Wuk Shin 1,  Jin-Sung Rho 1,  Jong-Min Lim 1

1키엘연구원

Accredited

ABSTRACT

This study implemented three ECG analysis pipelines through LLM-based vibe coding and examined the requests, generated code, and evaluation results. Using GPT-5.6 Sol in ChatGPT and the MIT-BIH Arrhythmia Database, we requested a complete analysis in one request (E1), sequential implementation steps (E2), and sequential steps with explicit methodological requirements (E3). We executed the generated code locally and examined data partitioning, normalization, class composition, and prediction metrics. E1 implemented the conventional DS1/DS2 record partition but assigned records 201 and 202 from the same patient to training and testing. E2 predicted all 455 test beats from record 100 as normal; its accuracy of 0.9846 matched an always-normal baseline. E3 implemented patient separation and normalization using internal training statistics. However, 2,062 of its 2,066 Q-class test beats came from record 104 with lead V5, and its F-class test set contained only 16 beats. From these cases, we derived verification items concerning data scope, partition units, preprocessing, class composition, and evaluation metrics. The findings illustrate the need to specify methodological requirements and examine the resulting data and evaluation structures when developing analysis pipelines through natural-language requests.

Citation status

* References for papers published after 2025 are currently being built.