Fill in the gaps - applying polynomial-based imputation techniques for heart rate data
Publication date
2026-06-01
Document type
Conference paper
Author
Organisational unit
Conference
Conference on Artificial Intelligence 2026 (IEEE CAI 2026) ; Granada, Spain ; May 8–10, 2026
Publisher
IEEE
Book title
2026 IEEE Conference on Artificial Intelligence (CAI)
Is supplemented by
Part of the university bibliography
✅
Language
English
Keyword
Bezier Imputation
Evaluation metrics
Hypoglycemia
PCHIP Imputation
Type 1 Diabetes
Abstract
Advances in clinical time-series modeling increasingly rely on AI-driven anomaly detection to identify the occurrence of adverse physiological events. The integration of multi-modal wearable sensor data with modern machine learning (ML) architectures has substantially improved the detection of extreme events in healthcare, surpassing the capabilities of conventional analytical methods. For instance, in individuals with type 1 diabetes (T1D), recent AI-based predictive models have demonstrated notable gains in forecasting critical states such as hypoglycemia (blood glucose below 70 mg/dL). Research on hypoglycemic prediction typically uses a combination of blood glucose (BG) readings and heart rate (HR) data to predict hypoglycemic events. Given that these features are collected through wearable sensors, they can have missing values, necessitating efficient imputation methods for improved prediction performance. This work makes significant contributions to the current state-of-the-art by introducing two novel polynomial-based imputation techniques for imputing HR values over short-term horizons: Controlled Weighted Rational Bezier Curves (CRBC) and Controlled PCHIP with Mapped Peak and Valleys of Control Points (CMPV). To evaluate pattern fidelity and error with respect to original values for short-term gaps, we employ RMSE and introduce two complementary pattern capture metrics and propose a combined metric (CM) that integrates them. Across all time gaps, CMPV achieves the best average CM score (0.33), followed by CRBC (0.48), outperforming alternative approaches. We then impute the missing HR values of the D1NAMO dataset to predict short-term hypoglycemia (before 30 min) as a binary classification task, using Support Vector Machines (SVM) and Linear Discriminant Analysis (LDA) classification models. Results show that our imputation methods both accurately reconstruct missing heart rate values and improve downstream prediction performance, underscoring their value for real-time hypoglycemia risk assessment.
Version
Published version
Access right on openHSU
Metadata only access