← Cloud analytics

Posture and movement profile

Second-by-second classification based on multi-sensor data.

κ = 0.93 agreement with human annotators on posture, against 0.95 between two annotators
13 categories: six postures and seven movement types
1.15 s time resolution across the whole recording

The posture and movement profile is the base layer of every MAIJU analysis. The classifier labels each moment of a recording — a new frame every 1.15 seconds — with one of six postures and one of seven movement categories, and flags the moments when the infant was carried (Figure 1A). The result is a timeline of what the infant actually did, rather than a count of how much they moved.

Summed over a recording, the timeline becomes a profile: how the infant's free playtime divides between postures, and what kind of movement happens within each. Repeated over months, the profiles show a child's motor repertoire changing shape (Figure 1B).

AHow the classifier labels a recording

BOne infant's posture mix, recording by recording

Figure 1. A The classifier reads all four sensors together, in 2.3-second frames that overlap by half. An encoder network fuses the 24 signal channels of each frame (three accelerometer and three gyroscope axes per sensor), and a second stage models how frames follow one another before assigning a posture, a movement category and a carrying label. Postures are shaded from lying to standing, as in B. B Share of free playtime in each posture at seven home recordings of one infant from the external clinical cohort of the growth-chart study — the same infant as in the BIMS figure. Up to 10.4 months the infant spent most of their playtime lying, on their back or front, while a clinical AIMS assessment placed them below its 5th centile (filled points). By 12.2 months sitting and crawl posture had taken over, and from 15.5 months standing took about half of playtime or more. Carried is time held or carried by an adult, which the analysis separates from the infant's own postures. Data: Airaksinen et al., Science Translational Medicine 2026, data file S4 (participant id_09).
Show the data for panel B
Age (months)SupineSideProneCrawlSittingStandingCarriedAIMS at the time
7.076.5%8.5%11.4%0.3%2.9%0.4%1.8%below 5th centile
8.438.4%7.9%42.9%1.0%9.7%0.1%2.9%below 5th centile
10.426.6%4.2%57.0%0.7%10.2%1.2%11.3%below 5th centile
12.20.0%0.0%0.3%31.9%61.2%6.6%6.1%at or above 5th centile
13.94.1%2.8%0.5%14.7%47.2%30.7%2.1%at or above 5th centile
15.50.7%1.1%0.1%8.0%27.8%62.4%2.2%at or above 5th centile
18.30.0%0.3%0.4%4.5%48.2%46.6%0.5%at or above 5th centile

What you get

Per recording
The labelled timeline — posture, movement and carrying for every frame — and the distributions derived from it: time in each posture, movement type within each posture, and the transitions between postures.
Per child
Profiles across repeated recordings, as in Figure 1B, and each metric positioned on its gross motor growth chart.
Postures
Prone, supine, side, crawl position, sitting, standing.
Movements
Still, roll, pivot, proto-movement, elementary movement, fluent movement, and transitions between postures.

What you need

Garment
MAIJU suit with a fitting size.
Sensors
Four Movesense Flash or Movesense MD sensors.
Successful recording
Every frame is labelled whatever the length; for stable distributions, from 30 minutes of detected free playtime with at least three sensors working, as in the published growth charts.

How it is derived

The classifiers are trained on second-level human annotations of synchronised video: 29.3 hours of recordings from 41 infants, each annotated independently by two or three of five trained annotators with a background in infant health care or research. Where annotators disagreed, their labels were combined with the classifier's own probabilistic decision in an iterative annotation refinement, to give more consistent training targets (Communications Medicine 2022). The first published classifier was a convolutional network; the pipeline now uses transformer-based classifiers pre-trained on unannotated recordings (IEEE EMBC 2023), the version behind the normative growth charts (Science Translational Medicine 2026).

A separate classifier of the same design detects when the infant is carried or held, so that being moved by an adult is not counted as the infant's own movement — 96% accuracy against video annotation (Scientific Reports 2024). The method builds on earlier work tracking infant posture and movement with wearable sensors (Scientific Reports 2020), and on a comparison of network architectures and data augmentation for the task (Sensors 2023).

Against human annotators

Posture is the easier task, for people and for the classifier. Movement categories have softer boundaries — fluent movement shades into elementary movement — and trained annotators agree on them less often, so the classifier is judged against the level people themselves reach.

Across posture categories the classifier agreed with human annotators at an average κ of 0.93 (prone and supine 0.97, standing 0.85), against 0.95 between the annotators themselves. For movement, where annotators agreed at κ = 0.60 overall, agreement between classifier and humans was substantial for the most distinct categories: still 0.68, roll 0.62, pivot 0.61 and fluent movement 0.73 (Communications Medicine 2022).

What the configuration changes

A later study retrained the classifier on reduced configurations to find what the result depends on. With the full four-sensor setup at 52 Hz, posture classification reached a Cohen's κ of about 0.90. A single sensor made the task infeasible (posture κ below 0.75, movement below 0.45), while the sampling rate could be reduced as far as 6 Hz with almost no loss. The minimum workable setup was one arm and one leg sensor at 13 Hz or more (JMIR mHealth and uHealth 2025). This is why MAIJU uses four sensors rather than one wrist device.

What the profiles show

In typical development the profile changes shape quickly. Lying falls from 92% of playtime at 6 months to 3% at 11 months; prone rises for a while before the infant moves up into crawl posture and sitting; sitting levels off at around a third of playtime from the first birthday; and standing keeps growing, to more than half of playtime at 19 months. Crawl posture peaks around the end of the first year but does not disappear once the infant walks — still about 7% of playtime at 16 months (Science Translational Medicine 2026). Each of these proportions has its own normative growth chart.

The profile also relates to development beyond movement: in 107 infants, more time in independent movement — crawling, standing and walking — went with more advanced prelinguistic and social abilities (Pediatric Research 2025).

Available on

Publications

All research →

Request a quote