미래 예측과 사건 기준 사전 예후 감지 방법의 비교: ECG 기반 AI 알고리즘 당뇨 예후 감지

초록

The guiding research question of this study is whether event-anchored pre-diagnostic prediction outperforms forward prediction, which estimates the probability of future disease occurrence. The answer provided by this study is affirmative. This is because forward prediction inherently involves a high level of uncertainty, whereas pre-diagnostic prediction focuses on detecting latent physiological signals that already exist prior to clinical diagnosis. Although both approaches utilize the same set of input features, they differ fundamentally in the definition of the target variable and the temporal structure of the data. Investigating how these differences affect predictive performance constitutes the primary objective of this study. To this end, we constructed a large-scale ECG-based dataset consisting of 20,000 samples, including 10,000 patients and 10,000 non-patients. Each record was designed to include temporal information relative to the diagnosis event. ECG-derived features, such as corrected QT interval (QTc), heart rate, and signal statistical characteristics, were extracted to form a consistent set of input variables. Based on these features, forward prediction models were developed across multiple temporal horizons (3, 5, 7, 10, and 15 years), while pre-diagnostic models were aligned relative to the diagnosis event. Model performance was evaluated using stratified 5-fold cross-validation and the area under the receiver operating characteristic curve (AUC). The results consistently demonstrate that pre-diagnostic models outperform forward prediction models across all temporal horizons, with the performance gap being more pronounced at shorter temporal distances. Furthermore, the findings remain robust across different model types, feature configurations, and random seeds. Additional analyses reveal that ECG signals contain detectable latent disease-related information several years prior to diagnosis. These findings suggest that the observed performance differences are not driven by model choice, but rather by fundamental differences in problem structure. This study highlights the importance of problem formulation in medical artificial intelligence and provides important insights into how temporal alignment influences predictive performance in long-term disease risk modeling.

키워드