D1NAMO-HR+

D1NAMO-HR+: Quality-Enhanced D1NAMO Dataset with Polynomial-based Heart Rate Imputation Description D1NAMO-HR+ is an enhanced version of the publicly available D1NAMO dataset in which missing heart rate (HR) values of 9 T1 diabetes patients have been reconstructed using two novel polynomial-based imputation techniques proposed in our work: Controlled Weighted Rational Bézier Curves (CRBC) and Controlled PCHIP with Mapped Peak and Valleys of Control Points (CMPV). The enhanced dataset is designed to improve the quality and completeness of wearable physiological data for machine learning applications in hypoglycemia prediction. Polynomial-based Imputation Methods The dataset contains HR values imputed using two complementary methods: Controlled Weighted Rational Bézier Curves (CRBC) reconstructs missing HR segments using weighted Rational Bézier curves. The method utilizes control points extracted from the observed HR values before and after each missing interval. Peaks and valleys within these control points are assigned higher weights, allowing the Bézier curve to preserve important physiological fluctuations while generating smooth and realistic transitions across the missing region. Controlled PCHIP with Mapped Peak and Valleys (CMPV) employs Piecewise Cubic Hermite Interpolating Polynomials (PCHIP) together with an extremum mapping strategy. Instead of simply interpolating between the endpoints, CMPV first maps the peaks and valleys from the surrounding HR sequences into the missing interval. These mapped control points provide a physiological template that preserves the local rhythm and variability of heart rate. PCHIP interpolation is then applied to generate a continuous HR signal while maintaining monotonicity and avoiding overshooting. Both methods constrain the imputed HR values to clinically plausible ranges (40–160 bpm), ensuring physiologically meaningful reconstructions. Dataset Purpose The primary objective of D1NAMO-HR+ is to provide a quality-enhanced version of the D1NAMO dataset for developing and evaluating machine learning models that rely on wearable physiological signals. By replacing missing HR values with physiologically consistent estimates, the dataset reduces information loss while preserving temporal patterns that are critical for clinical prediction tasks. Experimental Validation The proposed imputation methods were evaluated using a comprehensive framework that measures both numerical reconstruction accuracy and preservation of physiological signal patterns. Beyond reconstruction quality, the imputed dataset was used for short-term hypoglycemia prediction (30-minute prediction horizon) using Support Vector Machine (SVM) and Linear Discriminant Analysis (LDA) classifiers. Experimental results demonstrate that HR imputation using CRBC and CMPV improves downstream hypoglycemia prediction performance compared with conventional interpolation methods. In particular, the proposed methods achieve improved recall and balanced classification performance for hypoglycemic events while preserving the natural dynamics of heart rate, making D1NAMO-HR+ a valuable resource for wearable healthcare research and AI-based diabetes management