Machine Learning Tool May Effectively Identify Agitation in Dementia

A novel monitoring and analysis tool has identified nonpharmacologic, personalized interventions for agitation in dementia according to the findings of a study published in the journal EClinicalMedicine. 

Researchers conducted a longitudinal study from 2020 through 2023 using in-home monitoring devices to determine the association of real-world factors and agitation in an in-home setting. The researchers used data from the Minder study to develop a machine learning (ML) framework that identified episodes of agitation.

Participants aged 50 years or older with a clinically ascertained diagnosis of dementia or mild cognitive impairment and current or previous treatment at a psychiatric unit were included in the study. Most participants had study partners or caregivers who also attended study assessments. The researchers collected in-home monitoring data using passive infrared sensors (PIR), sleep mats, door sensors, and kitchen appliance sensors. Outdoor weather, temperature, and light data were also used.

The researchers contacted participants and/or study partners weekly to report the presence or absence of behavioral symptoms including agitation, delusions, hallucinations, depression, and anxiety. The presence of agitation during an 8-day period was identified using several different ML models, including Light Gradient-Boosting Machine (LightGBM).

A traffic-light–based system was developed to group probability estimates for each instance of agitation (green=low probability; amber=medium probability; red=high probability). Shapley Additive Explanation (SHAP) values were employed for statistical analysis.

A total of 63 participants were included (age group: 80-90 years, n=31; men, n=41; White race, n=23). Compared to weeks without reported agitation, weeks where agitation was reported had a significantly higher average awake ratio. Moreover, the average minimum respiratory rate was significantly lower in agitation vs non-agitation weeks. Indoor and outdoor illumination were also significantly higher during weeks of agitation.

Compared with other ML models, the LightGBM tool demonstrated significantly higher sensitivity and higher scores in most performance metrics. The addition of the traffic-light stratification improved all LightGBM metrics. Using SHAP values, the most important features for identification of agitation were:

• low respiratory rate,
• high awake ratio,
• extreme values of visibility and indoor illuminance (high and low),
• low illuminance ratio,
• high temperature ratio, and
• low indoor temperatures.

The researchers designed an interactive tool to improve clinical applicability of models by simulating changes in features that increase the likelihood of agitation. They found indoor lighting and temperature adjustments were the most promising and feasible intervention options to decrease the odds of agitation among participants.

Study limitations included potential confounding via study partners and a lack of diversity in the dataset.

“By employing an agitation monitoring model in real-world settings, we could enhance the detection of missed agitation instances, ultimately improving patient care and contributing to the development of more precise definitions for agitation episodes,” the researchers concluded.