💻 computer science

Adaptive Multi-Agent Feature Selection for Personalized Fall Risk Prevention

This paper introduces PAFIR, a reinforcement learning-based framework that performs adaptive, personalized feature selection on longitudinal multimodal health data to dynamically identify evolving fall risk factors and improve prevention strategies for older adults.

Original authors: Chang Liu, Ladda Thiamwong, Yanjie Fu, Rui Xie

Published 2026-08-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chang Liu, Ladda Thiamwong, Yanjie Fu, Rui Xie

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Falls among older adults are not merely accidents; they are often the result of a complex, shifting interplay between the body's physical capabilities and the mind's perception of safety. For decades, medical science has understood that risk factors like muscle weakness, poor balance, and fear of moving are central to preventing these injuries. However, a persistent challenge has been that these factors do not remain static. A person's risk profile changes over time, influenced by daily activities, gradual physical decline, or sudden shifts in confidence. Traditional methods for identifying who is at risk often rely on a single snapshot of a patient's health or a fixed list of warning signs, failing to account for how these signals evolve and interact across different types of data, from clinical exams to continuous movement tracking.

To address this, researchers have developed a new approach called PAFIR, a system designed to learn which risk factors matter most for each individual as time passes. The system was tested using data from a large real-world study involving 341 older adults living in senior communities in Central Florida. These participants were monitored over several months, providing a rich stream of information that included structured medical assessments, survey responses about their confidence and habits, and continuous, minute-by-minute recordings of their movement captured by wearable sensors. The goal was to see if a computer could learn to pick out the most relevant warning signs from this flood of data, adapting its choices as the participants' lives unfolded, rather than sticking to a rigid, pre-determined checklist.

The researchers found that by treating the selection of risk factors as a learning process, the system could identify patterns that static methods missed. The system works by constantly comparing different groups of data—such as physical strength versus psychological confidence—to see which combination best predicts a fall or a decline in balance. It uses a method similar to how a coach might adjust a training plan: it tries a new set of factors, observes the outcome, and then refines its choices based on whether the prediction improved. Crucially, because actual falls are rare events in a study of this size, the system was also taught to listen to "proxy" signals, such as a patient's self-reported fear of falling or their performance on a balance test, which provide continuous feedback even when no accident occurs. This allowed the system to learn effectively without waiting for a rare, negative event to happen.

When applied to the data from the PEER study, the system revealed distinct differences between two groups of participants: those receiving a specialized intervention that combined physical exercise with cognitive and psychological support, and those in a control group receiving standard care. For the intervention group, the system consistently identified a broader range of risk factors, including psychological measures like anxiety, attention control, and behavioral regulation, alongside physical metrics. This suggests that when the mind and body are engaged together, the risk of falling becomes tied to a wider set of interconnected factors. In contrast, for the control group, the system focused almost exclusively on traditional physical indicators like frailty, muscle function, and chronic health conditions. This divergence indicates that the type of care a person receives can fundamentally change which risk factors are most relevant to their safety.

The study also tracked how risk factors shifted for individuals leading up to and following a fall. For one specific participant, the system showed that before a fall, the most significant warning sign was a decline in hand grip strength. However, shortly after the fall occurred, the most prominent signal shifted to the amount of time the person spent lying down, while psychological factors like mindfulness and behavioral screening measures also became highly relevant. This illustrates that the warning signs for an individual are not fixed; they evolve as the person's condition changes. The system successfully captured these transitions, moving from a focus on physical capacity to a broader view that included psychological and behavioral changes in the aftermath of an injury.

In testing the system against other advanced methods, the researchers found that PAFIR was more accurate at recovering the true set of risk factors and more stable in its selections over time. It was particularly effective at preserving the relationships between different factors, recognizing that a decline in physical ability often goes hand-in-hand with a change in psychological state. By removing the system's ability to compare different groups of data or to use the continuous proxy signals, the researchers confirmed that these specific features were essential for its success. Without them, the system's ability to identify the correct factors dropped significantly, proving that the adaptive, comparative nature of the approach is what allows it to navigate the complexity of real-world health data.

The findings suggest that personalized fall prevention does not require a single, universal list of warning signs. Instead, the most effective strategies may need to adapt to the individual's current state, the type of care they are receiving, and the specific way their risk factors interact over time. By learning from the continuous flow of data that modern technology can provide, systems like PAFIR offer a way to move beyond static checklists toward a dynamic understanding of risk. This approach allows for earlier and more tailored interventions, potentially helping older adults maintain their independence by addressing the specific, evolving factors that put them in danger. While the system was tested in a controlled research setting, the results point toward a future where fall prevention is as responsive and individualized as the people it aims to protect.

Technical Summary: PAFIR – Personalized and Adaptive Feature Selection for Fall Risk Identification and Prevention

Problem Statement
Falls among older adults constitute a critical public health challenge driven by complex, time-varying interactions across multiple risk domains. Existing approaches to fall risk identification are largely static, failing to adaptively model evolving, individualized risk factors across diverse modalities (e.g., structured clinical surveys, wearable sensor streams) and time. Key challenges include:

Data Heterogeneity: Integrating sparse, delayed fall outcomes with continuous, high-frequency wearable data and structured clinical assessments.
Dynamic Interactions: The misalignment between physical capacity and psychological risk perception (e.g., fear of falling) evolves over time, yet most models assume static relationships or treat factors in isolation.
Sparse Rewards: Fall events are rare and irregular. In the PEER study, only 69 fall events were observed across 1,364 visit-level observations, creating a sparse reward signal that makes learning effective feature selection policies difficult for standard reinforcement learning (RL) agents.

Methodology: The PAFIR Framework
The authors propose PAFIR (Personalized and Adaptive Feature selection framework for fall risk Identification and pRevention), which formulates fall risk factor identification as a multi-agent reinforcement learning problem over longitudinal multimodal health data. The framework consists of three core innovations:

Hierarchical Multimodal State Representation:
- Structural Encoding: Clinical features are organized into clinically meaningful groups (e.g., body composition, psychological status) based on the PEER assessment domains. Within each group, features form a fully connected graph to capture intra-domain correlations. An inter-group topology captures co-variation in importance across different domains.
- Temporal Encoding: Wearable-derived time-series data (e.g., step counts, vector magnitude, posture states) are processed via a time-series encoder (Transformer-based) to capture fine-grained behavioral dynamics between clinical visits.
- State Fusion: A fusion mechanism aligns the structural feature representations with temporal embeddings to create a comprehensive state representation $s_{i,t}$ for each participant at each update $t$ .
Adaptive Comparison-Driven Feature Selection Mechanism:
PAFIR employs a three-agent RL framework to decompose the feature selection process into sequential, coordinated decisions:
- Candidate Agent: Proposes a feature group for evaluation based on potential outcome relevance.
- Reference Agent: Selects a competing feature group from the remaining space for comparison.
- Operation Agent: Decides whether to add, remove, or retain the candidate group based on the comparison.
  This mechanism iteratively refines the active feature set $F_{i,t}$ by comparing candidate groups against reference groups, allowing the model to assess incremental contributions rather than evaluating features in isolation.
Sparse-Aware Reward Design with Clinical Proxies:
To address the scarcity of fall events, PAFIR utilizes a dynamic reward mechanism:
- Fall Events: When a fall occurs, the reward is adjusted immediately to reflect the contribution of selected features.
- Clinical Proxies: In the absence of falls (scheduled visits), the reward is derived from clinically validated proxy measures known to correlate with fall risk, specifically the Falls Efficacy Scale-International (FES-I) and the BTrackS Balance Tracking System score (BBS).
- Reward Function: The reward $r_{i,t}$ is calculated as the improvement in performance (using fall outcomes or proxies) relative to the best previously observed performance, ensuring continuous feedback even when fall events are absent.

Key Contributions

Hierarchical Structure in RL: The paper embeds a topology-aware hierarchical structure into the RL framework to capture multi-level dependencies among features, reflecting the natural organization of clinical assessments (e.g., questionnaire items within a "fear of falling" domain).
Comparison-Driven Selection: The proposed mechanism decomposes selection into candidate generation, reference-based comparison, and iterative refinement, enabling more reliable identification of risk factors in complex, correlated healthcare settings.
Sparse-Aware Rewards: By augmenting sparse fall-event signals with clinical proxies, the framework captures broader fall-risk patterns and improves the robustness of feature selection in low-event scenarios.

Experimental Results
The framework was evaluated using data from the Physio fEedback Exercise pRogram (PEER) cluster-randomized trial (341 participants, 1,364 visit-level observations, and over 2.5 million time-stamped wearable observations).

Performance: PAFIR outperformed state-of-the-art baselines (including TAR, SAT, TTG, Group LASSO, and LASSO) in Feature Recovery Rate (FRR) (0.874 vs. 0.795 for the next best) and Selection Stability (0.990). It also achieved superior Correlation Recovery (0.974), indicating better preservation of feature dependencies.
Ablation Studies: Removing any core component (hierarchical structure, three-agent mechanism, or sparse-aware rewards) resulted in consistent performance degradation, confirming the necessity of each element.
Clinical Insights:
- Community-Level: PAFIR identified that the intervention site (Kinneret) showed consistent selection of psychological and behavioral factors (e.g., anxiety, mindfulness) alongside physical metrics, whereas the control site (LCA) relied predominantly on traditional physical and functional factors. This suggests the intervention successfully engaged psychological domains.
- Temporal Dynamics: The model captured shifting feature importance relative to fall events. Pre-fall, physical activity and muscle strength were prominent; post-fall, the model expanded to include psychological, behavioral, and clinical screening features (e.g., sleep patterns, chronotype), reflecting a multidimensional recovery and risk profile.
- Personalization: The framework demonstrated the ability to generate individualized, time-evolving feature priorities, adapting from baseline sedentary behavior indicators to post-fall psychological and clinical screening needs.

Significance and Claims
The paper claims that PAFIR offers a significant advancement over static or single-modal approaches by enabling dynamic, subject-specific feature selection that adapts to evolving risk factors. By integrating multimodal data and utilizing proxy signals to overcome data sparsity, PAFIR supports more timely and personalized fall prevention strategies. The authors posit that this approach is generalizable to other healthcare settings requiring the identification of clinically relevant risk factors from heterogeneous, multimodal, and longitudinal data, particularly where outcomes are sparse or delayed. The study emphasizes that effective prevention requires not just detecting risk signals but adaptively determining which features are most informative and when to intervene for each individual.

Technical Summary: PAFIR – Personalized and Adaptive Feature Selection for Fall Risk Identification and Prevention

More like this