AI predicts agitation in dementia patients using sleep and physiological data

Sleep disturbances are a key contributor to agitation. Poor sleep quality, characterized by frequent awakenings, insomnia, and fragmented rest, affects up to half of dementia patients in advanced stages. Although clinical research has long documented this connection, few predictive systems have integrated sleep parameters into automated models

Dementia cases are surging globally and the stakes for innovative caregiving tools are higher than ever. With over 55 million people affected worldwide and millions more expected in the coming decades, solutions that ease caregiver strain and enhance patient safety are urgently needed.

A team of researchers has demonstrated how machine learning can accurately predict agitation in dementia patients by analyzing sleep and physiological signals. Their study, “Machine Learning Prediction of Agitation in Dementia Patients Using Sleep and Physiological Data,” published in Applied Sciences (2025), presents a novel approach to one of dementia care’s most pressing challenges.

Agitation, marked by restlessness, aggression, and emotional distress, represents one of the most common and difficult symptoms in people living with dementia (PLwD). It worsens patient well-being, increases caregiver stress, and often leads to institutionalization. The research underscores how artificial intelligence can support caregivers by forecasting agitation episodes in advance, thereby enabling early interventions.

Can sleep patterns provide reliable signals of agitation risk?

Sleep disturbances are a key contributor to agitation. Poor sleep quality, characterized by frequent awakenings, insomnia, and fragmented rest, affects up to half of dementia patients in advanced stages. Although clinical research has long documented this connection, few predictive systems have integrated sleep parameters into automated models.

To address this gap, the team utilized the Technology Integrated Health Management (TIHM) dataset, which captures in-home monitoring data from dementia patients. This dataset provided detailed information on sleep states, vital signs, and caregiver-verified agitation events. For their analysis, the researchers focused on 17 patients with sleep data and expanded the sample with physiological and behavioral data from a larger group of 56 participants. After preprocessing and aligning datasets, they produced 525 usable records.

The study tested whether sleep quality indicators such as time in bed, total sleep time, sleep onset latency, and wake after sleep onset could predict agitation the following day. Results confirmed that these sleep measures, when combined with physiological data, offered valuable signals for predicting agitation risk.

Which machine learning models deliver the best predictions?

The researchers compared five machine learning algorithms: Random Forest, LightGBM, Extra Trees, XGBoost, and Gradient Boosting. Each was evaluated on its ability to balance accuracy, sensitivity, and reliability in identifying agitation versus non-agitation cases.

Among these, the LightGBM model consistently outperformed others, achieving a weighted F1 score of 93.6%. Importantly, it produced very few false negatives, meaning real agitation cases were rarely missed. In dementia care, minimizing missed detections is critical because overlooked agitation can escalate into harm for patients or caregivers. False positives, though present, were considered less concerning since early alerts allow for preventative action.

The team also applied explainable AI techniques, specifically Shapley values, to clarify which features drove predictions. This transparency is vital in healthcare, where clinicians must understand the reasoning behind algorithmic outputs. Key predictors included mean heart rate during sleep, variability in respiratory rate, and time in bed. Higher nighttime heart rates were strongly associated with increased agitation risk, while greater variability in respiratory rate tended to reduce the likelihood of agitation.

These insights align with medical literature showing that circadian rhythm disturbances and cardiovascular stress contribute to agitation in dementia patients. By validating machine learning outcomes against established medical knowledge, the study reinforced the clinical credibility of its findings.

What are the practical implications for dementia care?

The research focuses on minimal, non-invasive sensor inputs such as sleep mats and wearables. Unlike previous studies that required multiple sensor arrays and extensive data collection, this streamlined system reduces complexity for caregivers while still delivering actionable insights.

From a clinical standpoint, the ability to predict agitation a day in advance can transform dementia care. Caregivers could adjust routines, implement calming interventions, or seek medical support before an episode escalates. This not only improves patient safety but also eases caregiver burden, which is a significant factor in long-term dementia management.

The study acknowledges limitations, including the relatively small dataset and lack of demographic diversity, as most participants were older Caucasian individuals with confirmed dementia diagnoses. Missing data and imputation methods may have reduced the model’s ability to capture clinically meaningful variations. The authors recommend expanding future research with larger, more diverse populations and refining data preprocessing methods.

The study further lays the groundwork for real-world applications. Integrating predictive models into home monitoring systems could enable scalable, cost-effective solutions across healthcare systems worldwide, particularly in regions facing rising dementia prevalence.