A Hybrid Analytical Pipeline for Circadian Rhythm Modeling from Accelerometer Data: A Maned Wolf Case StudyJukan, Nermin (Avtor) Moškon, Miha (Mentor) Halužan Vasle, Ana (Komentor) Nogueira de Moraes, Rosana (Komentor) Potočnik, Hubert (Komentor) analytical pipelinePython packageopen-source softwarehidden Markov modelshidden semi-Markov modelscircadian rhythmsaccelerometrycosinor regressionlinear mixed modelsmaned wolfThis thesis presents a reproducible analytical pipeline for characterising circadian activity patterns from biologger time series, demonstrated on maned wolves (Chrysocyon brachyurus) in Brazil. The pipeline was developed during the process of the analysis of 408,171 observations from 15 subjects, recorded by implanted biologgers. Activity was quantified using the firmware-derived count of active minutes per 15-minute epoch; the accompanying per-axis mean accelerations were shown to describe body orientation rather than movement, and were reserved as an independent posture channel. Per-subject Gaussian Hidden Markov Models (HMMs) supported an interpretable four-state behavioural representation (deep rest, light rest, moderate activity, high activity), whose ordering was corroborated by the posture channel: mean posture magnitude declined monotonically with increasing activity. Circadian structure was quantified using multi-component cosinor regression. The five-harmonic model clearly outperformed the single-component baseline at population level (R2 = 0.993 vs. R2 = 0.878), confirming that multi-harmonic rhythmometry is required to capture the observed diel waveform. The fitted harmonics are ecologically coherent, with the fundamental peaking at 23:04 and the 12-hour component at 06:02 and 18:02. The framework was extended to Hidden Semi-Markov Models (HSMMs) with explicit Poisson duration distributions; comparison with HMM outputs showed substantial duration-related differences. For inferential modelling, sex was retained as the only group-level covariate in linear mixed models. Fixed sex effects were modest, while individual-level heterogeneity was substantial relative to the group-level sex effect (conditional R2 = 0.974-0.984; ICC = 0.954-0.971). Overall, the thesis contributes a coherent and extensible workflow that links behavioural-state inference, circadian rhythm modeling, and individual-variation analysis without requiring labelled behavioural ground truth.20262026-10-19 11:05:23Magistrsko delo/naloga189963VisID: 38576COBISS_ID: 303235331sl