Detection of daily activities from ear-worn accelerometer signals: a pilot study with topological data analysis

The purpose of this research was to develop and investigate a method for real-time lifestyle assessment, with the long-term goal of associating detected behaviours withmenstrual cycle phases, eating habits and hydration. This was achieved by designing a system for detecting health-related activities -- eating, drinking and smoking -- using a custom ear-worn device equipped with an accelerometer. To account for confounding behaviours, speaking and a generic ``other activities'' category were also included. Fifteen prototype devices were built and tested in the laboratory, and nine were worn by healthy volunteers for data collection. Raw accelerometer signals were segmented into overlapping windows and processed using topological data analysis (TDA) to extract shape-aware features. Time-delay embeddings were applied to transform signals into point clouds, from which persistence diagrams were computed using Vietoris-Rips and lower-star filtrations. Statistical descriptors derived from these representations included entropy measures, lifetime distributions, and topological complexity metrics. On a dataset of 47658 labelled windows collected from participants, a random forest classifier achieved 89% balanced accuracy on a held-out test split. Results demonstrate that TDA enables effective discrimination of mandibular and head movements captured by an ear-worn sensor. The source code and feature definitions are publicly available to support reproducibility.