Proactive behavioural monitoring in nursing homes: LSTM-based insights for intention out of bed
Master of Applied IT
Project description
This project explores how AI can enhance elderly care in nursing homes by predicting out-of-bed intentions and mitigating risks such as pressure ulcers. Using data from BedSense, a sensor system that monitors pressure, vibration, and bed angles, the study develops predictive models based on Long Short-Term Memory (LSTM) networks to analyse time-series patterns.
The goal is to assist caregivers with early warnings, enabling proactive interventions to prevent falls and manage pressure ulcers risks. Advanced preprocessing methods, such as noise filtering and dynamic thresholding, address sensor variability from different mattress types, improving data reliability.
This research is conducted in collaboration with healthcare providers and technology experts, using anonymized data compliant with GDPR standards. By combining real-world insights with predictive analytics, the project aims to improve resident safety, enhance caregiver efficiency, and provide a foundation for integrating AI into elderly care practices.
Context
This project explores how AI can enhance elderly care in nursing homes by addressing challenges like staff shortages and increasing demand. The research uses data from BedSense, a sensor system that captures pressure, vibration, and bed angle information, to predict out-of-bed intentions and mitigate risks such as pressure ulcers.
The study employs Long Short-Term Memory networks, a machine learning model designed for analysing time-series data, to identify transitions between behavioural states. These predictions allow caregivers to act proactively, reducing falls and addressing immobility-related risks that contribute to pressure ulcers. Advanced preprocessing methods, including noise filtering and dynamic thresholding, improve the quality and consistency of sensor data by accounting for factors like mattress variability and environmental conditions.
Collaboration with healthcare providers, nursing homes, and technology experts ensures the project aligns with real-world caregiving needs. All data is anonymized and processed according to GDPR regulations to uphold privacy and ethical standards.
By integrating predictive analytics into care settings, the project aims to enhance caregiver efficiency, improve resident safety, and provide a foundation for applying AI-driven solutions in elderly care practices.
Results
The results of this project suggest that AI-driven predictive models have the potential to improve elderly care in nursing homes by analysing data from BedSense, a sensor system that captures information on pressure, vibration, and bed angle. The models were designed to predict out-of-bed intentions and identify patterns of prolonged immobility, offering insights that may support caregiver interventions.
The Long Short-Term Memory (LSTM) network used in the study provided predictions for transitions between in-bed and out-of-bed states. The model showed reasonable recall and precision rates, with F1 scores of 0.87 for out-of-bed predictions and 0.84 for in-bed predictions. Probabilistic metrics such as a Brier score of 0.0884 and an RMSE of 0.2973 indicate that the model’s predictions generally aligned with observed outcomes.
Preprocessing methods, including noise filtering, resampling, and dynamic thresholding, improved the consistency of the data by addressing variability caused by mattress types, weight distribution, and environmental factors. However, challenges such as dataset imbalance emerged, as memory foam mattresses were overrepresented compared to other mattress types like static air mattresses. This limitation impacts the generalizability of the findings and highlights areas for further refinement.
The system’s predictions provided insights into resident behaviours, specifically on transitions between in-bed and out-of-bed states, which may help caregivers anticipate and respond to potential risks more effectively. Some predictions near the decision threshold reflected areas where further tuning and validation could enhance model performance.
The findings indicate that the approach could be useful for reducing unnecessary physical checks, predicting out-of-bed transitions, and identifying prolonged immobility. However, further work is needed to expand the dataset, improve the handling of uncertain states, and include more contextual data to refine the system’s applicability in diverse caregiving environments.
