This paper proposes a method for providing reliable uncertainty quantification, specifically adjusted prediction intervals (PIs), to enhance clinicians' confidence and interpretation when applying deep learning models to predict biosignals such as heart rate and blood pressure. Based on the Reconstruction Uncertainty Estimate (RUE), we propose two methods: a parametric approach assuming a Gaussian copula distribution and a non-parametric approach using k-nearest neighbors (KNN). Experimental results using minute- and hourly-sampled data demonstrate that the Gaussian copula method outperforms low-frequency data, while the KNN method outperforms high-frequency data. This suggests that the RUE-based PI can provide interpretable and uncertainty-aware biosignal predictions.
