Machine learning analysis of posturography in panic disorder: a pilot study for objective physiological biomarker identification

TL;DR

This study explores using machine learning on posturography data to identify objective biomarkers for panic disorder, showing high accuracy in distinguishing patients from healthy controls.

Contribution

First application of machine learning to posturography for identifying physiological markers of panic disorder.

Findings

Logistic Regression achieved 93.8% accuracy in classifying panic disorder patients and controls.

PD patients showed reduced mediolateral sway compared to controls under various sensory conditions.

Static posturography outperformed clinical screening tools like PHQ-PD and PDSS.

Abstract

Panic disorder (PD) is linked to subtle abnormalities in postural control, which are inadequately captured by traditional statistics. Machine learning (ML) techniques applied to stabilometric data may enhance the detection of PD-related postural patterns.

Evaluate static postural control in PD patients and determine if ML analysis of multivariate stabilometric data can improve differentiation from healthy controls.

In this cross-sectional case-control study, 12 adults diagnosed with DSM-5 PD and 21 matched healthy volunteers (total n = 33; 341 force platform trials) underwent stabilometry under five sensory conditions. Classical statistics used repeated-measures ANOVA on baseline trials only (to preserve independence). ML models (Decision Tree, k-Nearest Neighbors, Linear Discriminant Analysis, Logistic Regression, and Random Forest) were trained under stratified, subject-grouped, fourfold cross-validation (StratifiedGroupKFold), ensuring that all trials from each participant were confined to a single fold to prevent leakage. For the explainability of the model, Local Interpretable Model-Agnostic Explanations (LIME) was accessed.

ANOVA revealed a significant group and condition interaction for mediolateral center of pressure (CoP) displacement (p < 0.01), with PD patients exhibiting consistently reduced mediolateral sway. No significant between-group differences emerged for anteroposterior sway. Using an optimized decision threshold (Youden), Logistic Regression achieved an accuracy of 93.8% and area under the receiver operating characteristic curve (AUC) = 96%; Linear Discriminant Analysis presented the highest specificity (91.7%).

This is the first study applying ML to posturography for identifying physiological markers of panic disorder. Using ML for stabilometric data improves classification accuracy, highlighting static posturography as superior to clinical screening tools like the Patient Health Questionnaire for Panic Disorder (PHQ-PD) and Panic Disorder Severity Scale (PDSS). Larger, externally validated cohorts and portable measurement solutions are needed to translate these findings into routine clinical assessment.

Figures