How ActiNet turns wrist movement into a daily activity profile

A wrist accelerometer records motion, not a diary. ActiNet supplies an interpretation layer: it assigns 30-second windows to sleep, sedentary behaviour, light activity or moderate-to-vigorous activity. In its October 2025 preprint, a self-supervised deep-learning approach improved classification against a random-forest baseline.[1]

The central result, with its denominator

The comparison used CAPTURE-24 data from 151 participants, containing 318,686 labelled windows. Nested five-fold cross-validation separated people across the training and test groups. Both the baseline and ActiNet used temporal smoothing, so this was not a comparison between a smoothed sequence and an entirely raw one.[1]

The accuracy difference is 0.039, or 3.9 percentage points. Balanced accuracy averages recall across the four classes, while macro F1 gives each class equal weight when combining precision and recall. Those summaries help prevent a model from looking strong merely because it labels a common activity well.

For example, in an invented dataset containing mostly sedentary time, a classifier that repeatedly says “sedentary” might achieve an attractive overall hit rate while missing most exercise. The equal-class summaries ask a different question: how well does it do across the activity categories, including less frequent ones?

From movement to a sequence

ActiNet uses a pretrained movement feature extractor and then imposes temporal structure. Its post-processing also converts sleep blocks shorter than one hour to sedentary time. That rule is consequential for interpreting short naps; it is part of the pipeline, not a directly observed fact about the wearer.[1]

A sequence model can make an activity profile more coherent because adjacent windows belong to the same unfolding day. But coherence is supplied by assumptions about plausible transitions. A brief unusual period may be genuine, so the smoothing rule belongs in any account of how a daily total was produced.

What the daily numbers would mean

Here is a worked example, not participant data. A complete 24-hour record contains 2,880 half-minute windows. Suppose the final labels contain 960 sleep windows, 1,080 sedentary windows, 720 light-activity windows and 120 moderate-to-vigorous windows. Dividing each count by 120 gives eight, nine, six and one hours respectively, totalling 24.

A difference of 96 windows is already 48 minutes. That arithmetic is why seemingly small classification changes can matter when a study compares time spent in different behaviours. It is not an estimate of ActiNet’s daily error; it is a way to interpret the units.

The software provides time-series outputs and summary metrics. Its documentation distinguishes crude estimates from adjusted estimates that fill missing periods using the same time of day on other available days. An adjusted 24-hour profile can therefore contain imputed time as well as observed recording time.[2]

This is a research-data tool, rather than a promise that an ordinary watch export will work without preparation. The repository describes supported accelerometer files and a CSV route; the Oxford group also directs non-academic users to licensing arrangements.[2][3]

The useful result is a better-supported way to turn wrist-motion recordings into comparable activity profiles. To understand any resulting number of minutes, keep the classification model, temporal rules and missing-data treatment together with the total.

Report an error or suggest a correction