라이프로그 데이터의 신뢰성 확보를 위한 베이지안 신경망 기반 결측치 처리 방안 연구

초록

Methods: Heart-rate measurements were collected from nine participants over 3 months using commercial smart watches. After conducting sliding-window segmentation and min-max scaling, seven imputation methods were evaluated across missing rates of 5%, 10%, 15%, and 20%. These included three traditional statistical methods (mean, median, and mode), Multi-Layer Perceptron models in deterministic and Bayesian forms, and a Transformer-based Self-Attention Imputation for Time Series (SAITS) model in deterministic and Bayesian (Monte Carlo dropout) variants. Model performance was evaluated using RMSE, MAE, and MAPE. Results: The conventional statistical approaches resulted in higher errors, highlighting their limitations in capturing complex temporal patterns with simple point estimators. Conversely, machine learning models significantly reduced errors, and SAITS frequently surpassed other methods. The Bayesian framework did not consistently outperform deterministic models in point-estimation accuracy, but it provided valuable uncertainty quantification, particularly in regions with abrupt heart-rate fluctuations. Conclusion: These findings suggest that incorporating Bayesian principles enhances imputation reliability while offering critical uncertainty estimates. This is particularly advantageous in healthcare contexts, where inaccurate predictions can lead to clinical risks.

키워드