Ziel

Accurate assessment of activity-induced energy expenditure (AEE) is required to study its relationship with health outcomes and enable individual feedback-systems on energy balance. While accurate measurement methods exist within the laboratory (e.g. doubly-labeled water or indirect calorimetry), the challenge is to provide accurate AEE estimations in free-living conditions. This is typically done through wearable and unobtrusive systems such as accelerometry, but these systems are not currently accurate enough within research and clinical applications.
The proposed research project therefore aims to train and validate a machine learning model to estimate AEE using raw acceleration data captured from the thigh. Activity classification and stride segmentation algorithms will be used as additional model inputs.

Planung

To support efficient and robust training of the machine-learning model a diverse population sample will be recruited. All sample participants will undergo a standardized activity protocol in the laboratory while AEE will be measured using indirect calorimetry. Raw triaxial acceleration of the thigh will be measured and used for subsequent modelling. Activity classification and stride segmentation will be performed on the raw acceleration data using existing algorithmic approaches and subsequently used as model inputs. Model performance will be evaluated against indirect calorimetry as well as previous regression models.
In collaboration with the AUT Human Potential Centre, we will implement existing and validated activity classification algorithms. Further, we plan to merge prospective, co-existing datasets to enhance our possibilites for training and testing the modelling approach.

Ergebnisse

Based upon previous research using activity classification and stride segmentation, we expect the novel modelling approach to perform with greater accuracy compared to existing thigh-based estimation models as well as other wearable systems (such as a commercial smartwatch).