JMIR Publications

ABSTRACT

Background:

Population ageing is accompanied by parallel declines in two closely related domains: postural balance and cognitive function. Rather than being independent, balance control and cognition share neural and attentional resources and increasingly appear to decline together in older adults, with downstream consequences such as falls. Yet the two are usually assessed separately, using resource-intensive tools that are difficult to deliver repeatedly in community settings. Exergames are increasingly used to engage older adults in physical activity and may also provide an unobtrusive medium for embedding objective, low-burden assessment of both balance and cognition within enjoyable play.

Objective:

As a proof of concept, this study developed and internally evaluated whether an AI-driven, video-based dance exergame (SinDance) could estimate concurrent postural balance status and cognitive performance in older adults from standard video, without wearable or specialised laboratory equipment.

Methods:

Eighty-one older adults from Lions Befrienders Service Association (Singapore)’s Active Ageing Centres completed the SinDance exergame. Body postural key points were extracted from standard video using the MediaPipe pose estimation framework and used to develop machine learning models estimating (1) concurrent postural balance status, defined by a composite force-platform stabilometry score (HUR Balance Platform), and (2) concurrent cognitive performance, measured by the Montreal Cognitive Assessment (MoCA). Internal validation was performed using Leave-One-Out Cross-Validation (LOOCV) and 5-fold cross-validation. For balance classification, accuracy, sensitivity, specificity, and F1-score were reported. For MoCA estimation, Mean Absolute Error (MAE), R², and predictive accuracy within MoCA score tolerances of ±1, ±2, and ±3 points were reported.

Results:

For balance classification (n=77), a weighted ensemble model combining a Balanced Random Forest and a Support Vector Machine, trained with Ensemble with SMOTE data augmentation, achieved the highest LOOCV accuracy of 62.34% (n=77), with a sensitivity of 0.556 (95% CI 0.381–0.721), a specificity of 0.683 (95% CI 0.519–0.819), and an AUC of 0.752. For cognitive estimation (n=81), two LOOCV regression models were developed. Model 1 (sociodemographic, physical performance, and dance movement features) explained 47.6% of MoCA variance with an MAE of 2.72 points and 69.2% of predictions within ±3 points. Model 2 (sociodemographic and dance movement features) explained 39.5% of variance but achieved a lower MAE (2.48 points) and a higher proportion within ±3 points (72.8%).

Conclusions:

As a proof of concept, the SinDance system showed that AI derived biomechanical features from a simple, follow along dance exergame carry a measurable but modest signal related to concurrent postural balance and cognitive performance. To our knowledge, this is the first study to explore a video based dance exergame for estimating both constructs from standard video without wearables or specialised laboratory equipment. The community based implementation supports feasibility; however, the present evidence is preliminary, and external validation against established clinical references is required before any clinical use. Clinical Trial: N.A