Classification in postural style

Antoine Chambaz, Christophe DenisHAL 2012 · DOI

TL;DR

This paper introduces a novel statistical framework for classifying individuals based on their postural style, aiming to identify deficits in postural control, particularly in seniors. By leveraging advanced machine learning techniques like Targeted Maximum Likelihood Estimation (TMLE) and Super-Learning, they achieve a remarkable 87% classification accuracy using just the most informative postural protocol.

What Problem Does This Paper Address?

Maintaining stable posture is crucial for everyday activities and is influenced by a complex interplay of sensory information (visual, proprioceptive, vestibular). Deficits in postural control, especially in older adults, can lead to falls and reduced quality of life. While identifying these deficits is important for rehabilitation, current methods can be time-consuming and unpleasant for participants. This paper tackles the challenge of efficiently classifying individuals into "normal" and "deficit" groups based on their postural responses to various experimental perturbations, aiming to minimize the number of protocols needed for accurate classification.

Methods

The core of the methodology involves a two-step classification procedure. First, they rank four different experimental protocols based on their informativeness for distinguishing between postural styles. This ranking is achieved using Targeted Maximum Likelihood Estimation (TMLE), a robust statistical inference technique. TMLE, in turn, relies on Super-Learning, a machine learning approach that intelligently combines multiple estimators to create a more accurate and stable prediction model. Second, they use this ranked information to build plug-in classifiers. These classifiers leverage baseline covariates (age, gender, etc.) and summary measures extracted from the postural trajectories. The paper explores using the single most informative protocol, the two best, the three best, and all four protocols for classification.

Datasets

The study utilizes a dataset collected from 54 participants at the Center for the study of sensorimotor functioning (CESEM, Université Paris Descartes). This dataset includes 22 hemiplegic subjects (with proprioceptive deficits) and 32 normal subjects. Each participant undergoes four experimental protocols designed to perturb different sensory systems. The paper also includes a simulation study with three different scenarios to evaluate the classification procedure's performance under various conditions.

Key Insights and Ideas

  • Postural Style as a Classifiable Trait: The paper demonstrates that "postural style" can be quantitatively defined and used for classification, moving beyond qualitative descriptions.
  • Protocol Ranking is Crucial: Not all experimental protocols are equally informative. Identifying the most relevant ones significantly improves classification efficiency and accuracy.
  • TMLE and Super-Learning Synergy: The combination of TMLE for robust inference and Super-Learning for flexible model aggregation proves powerful for this complex classification task.
  • Efficiency Gains: The ability to achieve high accuracy with fewer protocols is a significant practical advantage, reducing participant burden.
  • Summary Measures are Sufficient: Extracting low-dimensional summary measures from complex postural trajectories is effective for classification, avoiding the need to process raw, high-dimensional data directly.

Key Results

The classification procedure, when applied to the real dataset, achieved a remarkable 87% rate of correct classification (47 out of 54 subjects correctly classified). This high performance was achieved using only the single most informative protocol, identified through the TMLE ranking. When using the two most informative protocols, the accuracy increased to 80% (43/54) with a smaller set of estimators in the super-learner, and further improved to 85% (46/54) when using all four protocols with a reduced set of estimators. The simulation study also showed good performance, with accuracies around 80-90% depending on the scenario and noise level.

Limitations and Future Work

The authors acknowledge that the sample size (n=54) is relatively small, and larger cohorts would likely yield even better results. They also note that the theoretical properties of Super-Learning are asymptotic, meaning they are most reliable with larger datasets. Future work could involve exploring more sophisticated summary measures derived from the original trajectories to minimize information loss. Additionally, the simulation study, while informative, was not exhaustive and could be expanded to cover a wider range of scenarios, including variations in conditional distributions and the dependence of noise on experimental conditions. The authors also suggest exploring the use of stochastic differential equations for simulating trajectories to better understand information loss from summary measures.

Summary updated

Abstract

This article contributes to the search for a notion of postural style, focusing on the issue of classifying subjects in terms of how they maintain posture. Longer term, the hope is to make it possible to determine on a case by case basis which sensorial information is prevalent in postural control, and to improve/adapt protocols for functional rehabilitation among those who show deficits in maintaining posture, typically seniors. Here, we specifically tackle the statistical problem of classifying subjects sampled from a two-class population. Each subject (enrolled in a cohort of 54 participants) undergoes four experimental protocols which are designed to evaluate potential deficits in maintaining posture. These protocols result in four complex trajectories, from which we can extract four small-dimensional summary measures. Because undergoing several protocols can be unpleasant, and sometimes painful, we try to limit the number of protocols needed for the classification. Therefore, we first rank the protocols by decreasing order of relevance, then we derive four plug-in classifiers which involve the best (i.e., more informative), the two best, the three best and all four protocols. This two-step procedure relies on the cutting-edge methodologies of targeted maximum likelihood learning (a methodology for robust and efficient inference) and super-learning (a machine learning procedure for aggregating various estimation procedures into a single better estimation procedure). A simulation study is carried out. The performances of the procedure applied to the real data set (and evaluated by the leave-one-out rule) go as high as an 87% rate of correct classification (47 out of 54 subjects correctly classified), using only the best protocol.
BibTeX
@inproceedings{chambaz2012classification,
  author = {Antoine Chambaz and Christophe Denis},
  title = {{Classification in postural style}},
  booktitle = {HAL},
  year = {2012},
  doi = {10.48550/arxiv.1209.6221},
}
Figure