Project OverviewMy project focuses on developing and applying temporal-modelling approaches to accelerometry to extract interpretable dynamical features of human activity and relate them to health outcomes in population datasets (mainly NHANES). Secondments: Leiden University (Netherlands) – Collaboration on causal representation learning for physical activity using foundation models with My Little Moves data from Amsterdam UMC (0–4 y accelerometry and behaviour data); EpiAgeing Team (Paris) – Planned work on time-series clustering based on physical-activity datasets.First StudyIn this project I introduce metrics derived from Hawkes point processes, which are well-suited to capture self-exciting and bursty structures in accelerometry data. Accelerometer signals recorded at different body locations (hip, wrist) are transformed into point processes from which Hawkes parameters and summary measures are obtained. These new metrics are then related to BMI and mortality using NHANES to test whether dynamical features of activity, beyond total volume, contain epidemiologically relevant information.Second StudyThe baseline Hawkes formulation is extended to better reflect real-world physical activity rhythms. Two extensions are considered: (1) regularization strategies to ensure numerical stability and generalizability, and (2) switching baseline intensities to model distinct behavioural regimes such as sleep and wake states. These extended models aim to more faithfully capture the temporal structure of human movement while still allowing health-linked inference. This work also uses NHANES data.Third StudyIn this project I use fractional Brownian motion (fBm) as a motivating interpretive framework to study scale-dependent regulation in physical activity. I will assess the concordance and divergence between time-domain (DFA) and frequency-domain (PSD) scaling estimates under real-world nonstationarity, to better understand long-range correlations and fluctuation structure in accelerometer data. I will then evaluate whether frequency-domain features provide added value beyond time-domain measures by testing their associations with demographic and health outcomes in NHANES, which serves as the empirical test-bed.Fourth StudyI apply matrix profiling to discover recurrent temporal patterns (“motifs”) in physical-activity time series. The goal is to classify and interpret common behavioural motifs and examine their epidemiological correlates using NHANES. In addition, these motif-based summaries will be integrated into a toolbox for multi-scale time-series visualisation, enabling interpretable inspection of behavioural structure across different temporal resolutions. |