Truace
Subscribe

Lower limb prosthetic users' activity classification through real-world inertial sensing

The prescription of lower limb prosthetics relies heavily on the assigned K-level. K-level assessment has been criticised for subjectivity and variability, potentially leading to inappropriate component selection and poorer functional outcomes. This study presents an original methodology for objectively determining activity levels from inertial sensor data, developed through rigorous data collection on 20 amputees who walked on various types of outdoor terrain (flat ground, stairs, ramp, grass/uneven/unstable su…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1671· cite & verify
Lower limb prosthetic users' activity classification through real-world inertial sensing
Sprinting performance of individuals with unilateral transfemoral amputation - compensation strategies for lower limb coordination by Mingyu Hu, Toshiki Kobayashi, Genki Hisano, Hiroto Murata, Daisuke Ichimura and Hiroaki Hobara. CC BY 4.0 · https://creativecommons.org/licenses/by/4.0

In brief

Researchers collected inertial sensor data from 20 people with lower limb amputations walking over flat ground, stairs, ramps and uneven surfaces such as grass, with and without a walking aid. They trained machine learning models to recognise these daily activities outside the laboratory, aiming to replace the subjective K-level assessment used to guide prosthetic component choice.

If the approach matures, prescribers could rely on measured daily activity rather than judgement alone when selecting prosthetic components, which might improve mobility over time. The authors stress that accuracy on unseen users is modest and that the sample is small, so the tool is not yet ready to change prescribing practice.

Main points

  1. The study addresses criticism that K-level assessment is subjective and variable, which can lead to inappropriate component selection.
  2. A random forest classifier reached 85.65% cross-validation accuracy but only 65.70% mean leave-some-out accuracy.
  3. Per-class sensitivity in leave-some-out analyses ranged from roughly 44% to 78%, and was lower for ramp ascent and descent.
  4. A preliminary clinical evaluation involved 3 amputees monitored longitudinally and interviews with 4 prescribing health care professionals.

The gain

A random forest classifier trained on real-world sensor data from 20 amputees classified outdoor activities with 85.65% cross-validation accuracy, pointing toward more objective activity assessment for prosthetic prescription.

The rundown

The data came from 20 amputees walking on varied outdoor terrain, including flat ground, stairs, ramps and grass or other uneven or unstable surfaces. Walking was recorded with and without a walking aid using inertial measurement units, so the model was trained on real-world rather than laboratory movement.

The gap between cross-validation and leave-some-out accuracy was read by the authors as participant-specific overfitting. They treat the lower leave-some-out figure as the conservative estimate for new users and the higher figure as potential accuracy with a more robust dataset.

A small clinical evaluation tested an unsupervised, longitudinal monitoring protocol on 3 amputees. Exploratory interviews with 4 prescribing health care professionals gave preliminary views on how clinically significant the outcomes would be.

Sources

  1. Peer-reviewedPLOS One2026-10-09

The debate