Abstract
Falls remain a leading cause of injury-related morbidity and mortality among adults aged 65 and older, and these human and economic costs keep rising as populations age worldwide. Most existing risk assessment tools rely on periodic clinical evaluations or single-sensor wearable devices, neither of which tracks the dynamic, multifactorial nature of fall risk throughout daily life. In this paper we describe the design, implementation, and pilot-level field validation of an integrated system that pairs a multi-sensor wearable unit (triaxial accelerometer, gyroscope, insole pressure array, and wrist-worn photoplethysmography) with a CNN-LSTM deep learning backbone enhanced by eight-head attention. The network ingests temporally aligned signals from all four modalities, extracts hierarchical spatiotemporal features, and outputs a continuous risk score that a rule-based engine maps onto four graded alert tiers. We collected data from 120 community-dwelling older adults (mean age 72.4 years) over six months and evaluated the system on a held-out test set of 18 participants (12,540 windows, 6 fall events). The model reached 94.2% overall accuracy and an AUC of 0.967, surpassing random forest, SVM, single-modality CNN and LSTM, a Transformer encoder baseline, and a late-fusion variant. We stress, however, that the extreme class imbalance (fall windows constitute only 0.05% of total windows, with a high-risk PPV of merely 6.25%) severely limits the stability and interpretability of sensitivity estimates; a single misclassification among the six fall events shifts sensitivity by roughly 17% points. End-to-end pipeline latency averaged 138 ms on a smartphone, satisfying real-time operational requirements. In a subsequent three-month deployment with 80 participants, the group receiving active alerts recorded 62% fewer falls than the passive-monitoring control group (0.82 vs. 2.14 falls per person-month; p < 0.01). This deployment, however, was conceived as a technical field test of system operability rather than a clinical efficacy trial; the fall incidence difference is an exploratory, secondary observation confounded by concurrent daily risk summaries and caregiver notifications, and it should not be interpreted as evidence of causal preventive effect. Importantly, the window-level classification PPV (6.25%) and the deployment-level alert PPV (87.8%) differ because the latter applies episode-level aggregation through the rule-based warning engine, and readers should note this distinction when comparing metrics across sections. User satisfaction averaged 4.3/5 on a Likert scale, and wearing compliance reached 91.3% of waking hours. Taken together, these results suggest that multimodal sensor fusion combined with attention-based deep learning offers a technically viable path toward continuous, community-based fall risk monitoring, while also underscoring the need for larger-scale, prospectively registered randomized controlled trials to establish whether such systems can causally reduce fall incidence.