Sleep position monitoring is key when attempting
to address posture triggered sleep disorders. Many studies have
explored sleep posture detection from a dedicated physical
sensing channel exploiting optimum body locations, such as
the torso; or alternatively non-contact approaches. But, little
work has been done to try to detect sleep position from a body
location which, whilst being suboptimal for that purpose, does
however allow for better extraction of more critical biomarkers
from other sensing modalities, making possible multi-modal
monitoring in certain clinical applications. This work presents
two different approaches, at varying levels of complexity, for
detecting 4 main sleep positions (supine, prone, lateral right
and lateral left) from accelerometry data obtained by a single
wearable device placed on the neck. An ultra light-weight
threshold-based model is presented in this work, in addition
to an Extra-Trees classifier. The threshold-based model was
able to achieve 95% average accuracy and 0.89 F1-score on
out-of-sample data, showing that it is possible to obtain a
moderately high classification performance using a simple rule-
based model. The ExtraTrees classifier, on the other hand, was
able to achieve 99% average accuracy and 0.99 average F1-
score using only 25 base estimators with maximum depth of
20. Both models show promise in detecting sleep posture with
high accuracy when collecting the signals from a neck-worn
accelerometer sensor.