Using a smartwatch to detect changes in brain health

Can smartphones or smartwatches help detect early signs of neurological or mental illness? Researchers at the University of Geneva (UNIGE) monitored a group of participants wearing connected devices, and used artificial intelligence to analyze data such as heart rate, physical activity, sleep and air pollution. Their findings show that connected devices can accurately predict emotional and cognitive fluctuations, opening new avenues for the early detection of changes in brain health. The study has been published in npj Digital Medicine.

Brain health, encompassing both cognitive and emotional functions, is one of the major public health challenges of the 21st century. According to the World Health Organization (WHO), more than one in three people worldwide live with neurological disorders such as stroke, epilepsy or Parkinson’s disease, while more than one in two individuals will experience a mental disorder — including depression, anxiety disorders or schizophrenia — at some point in their lives. As populations age, these figures continue to rise.

Even in healthy adults, brain health fluctuates over time, reflecting interactions between multiple factors, including environmental influences and individual lifestyle habits. Analysing day-to-day or week-to-week changes in cognitive and emotional functioning is therefore essential to enable proactive and individualized prevention strategies.

A team at the University of Geneva (UNIGE) set out to determine whether wearable and mobile technologies could be used to monitor brain health continuously and non-invasively. To this end, 88 volunteers aged between 45 and 77 were equipped with a dedicated smartphone app and a smartwatch. Over a ten-month period, these devices collected “passive” data — without any intervention or change in participants’ daily habits — including heart rate, physical activity, sleep patterns, as well as weather conditions and air pollution levels. In total, 21 indicators were analyzed.

Every three months, participants also provided “active” data by completing questionnaires on their emotional state and undergoing cognitive performance tests.

AI-analyzed data

Once data collection was complete, the passive data were analysed using artificial intelligence developed as part of the project.

The aim was to determine whether AI could predict fluctuations in participants’ cognitive and emotional health based on these data.”

Igor Matias, doctoral assistant at the Research Institute for Statistics and Information Science at the Geneva School of Economics and Management (GSEM) at UNIGE and lead author of the study

The AI-based predictions were then compared with the results of the questionnaires and tests. “On average, the error rate was just 12.5%, opening up new possibilities for the use of connected devices in the early detection of abnormalities or changes in brain health,” the researcher adds.

Emotional states are the easiest to predict

Emotional states were the most accurately predicted by the artificial intelligence, with error rates ranging generally between 5% and 10%. Cognitive states, in contrast, were predicted less precisely, with error rates ranging from 10% to 20%. In other words, AI performs better at forecasting responses to emotional questionnaires than cognitive tests.

Regarding the relevance of passive indicators, air pollution, weather conditions, daily heart rate, and sleep variability emerged as the most informative factors for cognition. For emotional states, the most influential factors were primarily weather, sleep variability, and heart rate during sleep.

This research, supervised by Prof. Katarzyna Wac of the Research Institute for Statistics and Information Science at GSEM and Prof. Matthias Kliegel of the Cognitive Aging Laboratory at the Faculty of Psychology and Educational Sciences, is part of the joint faculty project Providemus alz. The next phase is already underway. It aims to collect the same types of data over a 24-month period, while examining individual characteristics of participants associated with the highest- and lowest-performing AI models, in order to better understand their applicability in real-world individualised scenarios.