Objective
The objective of this project is to develop intelligent imputation methods for reconstructing missing values in multivariate sensor time-series data, with a primary focus on healthcare applications. The project investigates how different missing-data mechanisms and gap lengths influence imputation quality and aims to design adaptive, gap-aware reconstruction strategies that improve both data quality and the performance of downstream machine learning models for early event prediction
Description
Sensor data collected from wearable devices, physiological monitors, industrial sensors, and IoT systems has become a cornerstone of modern artificial intelligence applications. In healthcare, continuous monitoring of physiological signals such as heart rate, glucose, and physical activity enables early detection of critical conditions, including hypoglycemia, cardiac abnormalities, and patient deterioration. Similarly, industrial sensor networks support predictive maintenance and fault detection.
However, real-world sensor data is frequently affected by missing values caused by sensor malfunction, battery limitations, communication failures, motion artifacts, environmental interference, or human error. These missing segments reduce data quality, introduce bias into machine learning models, and may obscure clinically significant events. The problem is particularly severe in healthcare, where inaccurate reconstruction of physiological signals can directly impact predictive performance and clinical decision support.
Missing data varies considerably in both origin and duration. It can occur as Missing Completely at Random (MCAR), Missing at Random (MAR), or Missing Not at Random (MNAR), with gap lengths ranging from a few seconds or minutes to several days. Despite these differences, current state-of-the-art research often applies a single imputation technique, most commonly linear interpolation, to all missing gaps. Such generic approaches fail to capture the complex temporal dynamics and physiological characteristics of sensor signals.
This project proposes an intelligent, gap-aware imputation framework that selects suitable reconstruction techniques based on missingness characteristics, gap length, signal properties, and contextual information. By combining statistical methods, polynomial-based models, machine learning, and domain knowledge, the project aims to reconstruct missing sensor data more accurately while preserving meaningful physiological patterns.
Ultimately, the research seeks to improve both reconstruction quality and downstream predictive performance, thereby enabling more reliable AI systems for healthcare and other sensor-driven applications.
Current Research
The current research focuses on developing a comprehensive imputation paradigm that adapts reconstruction methods to different missing-gap scenarios instead of relying on a single universal technique. The framework investigates collections of complementary imputation algorithms tailored to different gap lengths and missing-data mechanisms and validates them using real-world healthcare datasets.
A major research direction is the evaluation of how imputation quality influences downstream machine learning tasks rather than considering reconstruction error alone. This is particularly important for early hypoglycemia prediction using wearable sensor data, where improved data quality directly translates into better predictive performance.
Recent contributions include the development of two novel polynomial-based imputation methods, CRBC and CMPV, for reconstructing missing heart rate data. These methods consistently outperform conventional approaches across varying short-gap lengths and significantly improve hypoglycemia prediction up to 30 minutes before onset.
Letzte Änderung: 6. August 2026