AIMO
Help Download
← Knowledge

Basics

How the AIMO Score measures

How does a camera rate a squat – and how well does that match experts? The foundations of the AIMO Score from six years of research.

Award-winning

The foundational paper “Towards an Automated Assessment of Musculoskeletal Insufficiencies” received the Best Paper Award at KES-IDT 2019 in Malta and was published by Springer.

Why the squat?

The overhead squat is a classic of movement analysis. It requires mobility in the ankle, hip and thoracic spine, plus strength and control. That makes it representative of many everyday movements: sitting down and standing up, bending, lifting, climbing stairs. The National Academy of Sports Medicine (NASM) has described exactly how experts rate it: do the knees go inward? Does the upper body tilt forward? Do the arms drop? Do the heels lift?

The problem: this rating depends on the eye of the person watching. Two experts don't always reach the same result, and even the same person rates differently after the hundredth recording than after the first.

Step 1: translating the experts' view into an algorithm

In the first paper (Dressler, Liapota, Löwe 2019), the NASM assessment was translated into an algorithm. A 3D camera records the squat, the algorithm detects the typical compensatory movements, weights them and calculates a score. For this, a new method was developed that aligns movement recordings in time while ignoring the start, pauses and end that are not part of the movement.

The result: experts rated 81 movement recordings according to the NASM standard. The automatic score agreed well with this (correlation r = 0.86). Analysing the deviations showed that a large part was due to differences between the experts themselves. When the most experienced expert carefully re-rated the same recordings, the agreement rose to r = 0.96.

The authors also name the limitations: the sample was not representative, and the validation should be strengthened further with larger groups and trained experts.

Step 2: learning instead of writing rules

Writing rules by hand for every new exercise is laborious. The second paper (Dressler, Liapota, Löwe 2019) presents five data-driven methods that learn from example recordings what a good execution looks like. One of them agreed well with the tested method from step 1 and requires hardly any expert knowledge of the exercise. Further work compared which learning methods work best (Hagelbäck et al. 2019). A random forest approach was the most accurate. A second paper shows how to compare fast and slow executions fairly (dynamic time warping).

Step 3: many values, one fair score

A movement recording produces dozens of individual values with different units and distributions: angles, distances, times. Simply adding them up would be arbitrary. Welf Löwe's research group developed a method based on joint probabilities (Ulan et al. 2018, 2021). Put simply, the score describes how a movement performs compared with many others. That makes it comparable and traceable. The method was tested on very large datasets from software quality, with the same mathematical foundations as the AIMO Score.

Step 4: from special device to smartphone

The first studies used 3D cameras such as the Kinect. A study of three commercially available 3D cameras showed that they deliver such different data that trained models are not easily transferable (Hagelbäck et al. 2019). For the move to the smartphone this meant that the models had to be retrained and tested for the new camera. In addition, the “jitter” in the camera data had to be smoothed without distorting the movement (Gauss et al. 2021). The appropriate filter runs in the app today.

What this means today

The AIMO Score (0–100 points) is based on methods published in journals and at international conferences.

With a 3D camera, it agreed well with expert ratings.

Validating the smartphone camera measurement under everyday conditions is an ongoing research goal. Individual scans can vary; the trend over several scans is what's meaningful.

The scan measures movement quality. It does not make a diagnosis.

Sources

[1], [2], [3], [4], [5], [7], [9], [10], [11] of the publication list

Why the squat? →All publications →
Try it yourselfStart with the movement scan – free with many health insurers.