Assessing the utility and limitations of accelerometers and machine learning approaches in classifying behaviour during lactation in a phocid seal

Table 3 Comparison of behaviour classification across accelerometer mounts

Behaviour	2015 Head mount (50 Hz)				2016 Head mount (25 Hz)				2016 Torso mount (25 Hz)
Behaviour	Precision	Recall	F1	F1 (Indiv.) (n = 5)	Precision	Recall	F1	F1 (Indiv.) (n = 12)	Precision	Recall	F1	F1 (Indiv.) (n = 7)
Rest	0.754	0.765	0.759	0.738 ± 0.13	0.690	0.925	0.790 (4.1%)	0.843 ± 0.04	0.731	0.935	0.820 (3.7%)	0.846 ± 0.05
Alert	0.643	0.769	0.700	0.746 ± 0.04	0.514	0.301	0.379 (− 45%)	0.552 ± 0.13	0.625	0.398	0.486 (28%)	0.595 ± 0.12
Presenting/Nursing	0.720	0.560	0.630	0.687 ± 0.15	0.808	0.192	0.311 (− 51%)	0.647 ± 0.17	0.771	0.496	0.604 (94%)	0.657 ± 0.25
Locomotion	0.925	0.027	0.054	0.147 ± 0.27	1.000	0.004	0.008 (− 85%)	0.045 ± 0.06	0.000	0.000	0.000	0.029 ± 0.05
Comfort Movement	0.638	0.015	0.029	0.111 ± 0.10	0.707	0.004	0.009 (− 68%)	0.059 ± 0.03	0.999	0.017	0.034 (277%)	0.067 ± 0.05
Flippering pup	0.688	0.407	0.511	0.425 ± 0.21	0.640	0.228	0.336 (− 34%)	0.252 ± 0.19	0.759	0.303	0.433 (28%)	0.320 ± 0.24
Mean	0.728	0.423	0.447	0.476 ± 0.31	0.726	0.275	0.305 (− 31%)	0.400 ± 0.32	0.647	0.358	0.396 (29%)	0.419 ± 0.34

Precision (positive predictive value), Recall (sensitivity), and F1 (harmonic mean of Precision and Recall) for Rest, Alert, Presenting/Nursing, Locomotion, Comfort Movement, and Flippering pup behaviours as defined in Table 1 modelled using a random forest for 2015 head-mounted, 2016 head-mounted, and 2016 torso-mounted accelerometers using data pooled across all individuals. F1 values when random forests are fit to individuals (‘F1 (Indiv.)’) are included for comparison, though only a small subset of individuals had sufficient data in all behavioural categories to build a model. Values closer to 1 indicate better model performance for each behaviour category. Values in parentheses indicate the change in pooled F1 values between 2015 and 2016 as well as between accelerometer mount configurations for the 2016 head- and torso-mounted accelerometry data, respectively. Locomotion bouts were completely undetected in the 2016 torso-mounted accelerometers

ISSN: 2050-3385