Context-Aware Alerting in Elderly Care Facilities: A Hybrid Framework Integrating LLM Reasoning with Rule-Based Logic

Area of Science:

• Gerontology and Health Informatics
• Artificial Intelligence in Healthcare
• Nursing and Patient Safety Technology

Background:

• Conventional nurse alert systems in long-term care (LTC) facilities suffer from excessive false alarms, leading to alarm fatigue and reduced care quality.
• Nursing shortages exacerbate the challenges of managing patient needs and alert systems effectively.
• Existing systems lack the contextual awareness to differentiate critical alerts from non-urgent ones, impacting caregiver efficiency and patient safety.

Purpose of the Study:

• To develop and evaluate a hybrid, context-aware nurse alerting framework for LTC facilities.
• To minimize redundant alarms and reduce alarm fatigue by intelligently suppressing, delaying, and validating alerts.
• To enhance patient safety and caregiver balance in multi-person care scenarios, such as mealtimes.

Main Methods:

• Integration of rule-based logic with Large Language Model (LLM)-driven semantic reasoning for intelligent alert management.
• Utilized video-derived skeletal motion, resident care-level annotations, and dynamic nurse-elderly proximity data for decision-making.
• Conducted an experimental study in a real-world LTC environment with 28 residents and 4 nurses over seven days, employing statistical analyses (F1 score, accuracy, FPR, FNR) for performance evaluation.

Main Results:

• The proposed framework reduced average alarm load by 72.5% (from 100% baseline to 27.5%) compared to traditional systems.
• Achieved significant improvements in performance metrics: macro F1 score increased from 0.18 to 0.97, and accuracy rose from 0.21 to 0.98.
• Drastically reduced error rates, with false positive rate (FPR) dropping from 0.20 to 0.005 and false negative rate (FNR) from 0.79 to 0.023, outperforming baseline and rule-based methods.

Conclusions:

• The hybrid alerting framework effectively minimizes false alarms and maintains operational efficiency in LTC settings.
• Integration of LLM-based contextual reasoning with rule-based systems enhances alert accuracy and mitigates alarm fatigue.
• The system promotes safer, more sustainable, and human-centered care practices, demonstrating suitability for practical deployment in real-world LTC environments.