Smartwatch data predicted emotional health with just 5-10% error over ten months

Can a device on your wrist detect that your brain health is slipping before you notice it yourself? Researchers at the University of Geneva set out to answer that question by strapping smartwatches on 88 volunteers for ten months and letting an AI figure out what the data meant. The answer, published in npj Digital Medicine, is a qualified yes, with some important caveats about what qualifies as prediction and what remains aspiration.

Passive data, no behavior change required

The study enrolled participants aged 45 to 77 and equipped them with a dedicated smartphone app and a smartwatch. Over the monitoring period, the devices continuously collected what the researchers call passive data: measurements that require no action from the wearer. Heart rate, physical activity levels, sleep patterns, step counts, and environmental factors including weather conditions and air pollution exposure. In total, 21 indicators were tracked.

Every three months, participants also completed active assessments: standardized questionnaires about their emotional state and cognitive performance tests. These served as the ground truth against which the passive data predictions were measured.

12.5% average error, but emotions are easier than cognition

The AI models, developed specifically for this project, were trained to predict the questionnaire and test results from the passive data alone. The average prediction error across all measures was 12.5 percent, which the researchers describe as opening new possibilities for continuous, non-invasive brain health monitoring.

But the averages obscure an important distinction. Emotional states were predicted far more accurately, with error rates generally between 5 and 10 percent. Cognitive performance was harder to forecast, with errors ranging from 10 to 20 percent. In practical terms, the AI was substantially better at predicting how someone would answer a mood questionnaire than how they would perform on a memory or attention test.

This discrepancy makes biological sense. Emotional states are tightly coupled to physiological signals that wearables capture well: heart rate variability, sleep disruption, and activity levels all have established relationships with mood. Cognitive performance depends on a broader set of factors, many of which, such as medication effects, social engagement, and acute stressors, are not captured by a smartwatch.

Which passive signals mattered most

For predicting cognitive fluctuations, the most informative passive indicators were air pollution exposure, weather conditions, daily heart rate patterns, and sleep variability. For emotional states, the strongest predictors were weather, sleep variability, and heart rate during sleep.

The air pollution finding is particularly interesting. Mounting evidence links particulate matter exposure to both acute cognitive decrements and long-term neurodegeneration, but this is among the first studies to show that ambient pollution data, passively captured through smartphone location and environmental databases, can predict individual-level cognitive performance changes over time.

What counts as brain health monitoring

The study sits at the intersection of two major public health trends. According to the World Health Organization, more than one in three people worldwide live with neurological disorders, and more than one in two will experience a mental disorder at some point. As populations age, these numbers are climbing. The appeal of continuous, passive monitoring that could flag declining brain health before clinical symptoms emerge is obvious.

But the gap between predicting questionnaire scores with 12.5 percent error and detecting early neurological disease is enormous. The study participants were generally healthy adults, not patients with diagnosed conditions. The emotional and cognitive fluctuations being predicted were within the normal range of day-to-day variation, not pathological declines.

Limitations the numbers do not capture

Eighty-eight participants is a small sample by clinical standards, and the demographic was narrow: adults aged 45 to 77, presumably from the Geneva area. Whether the AI models would perform similarly across different age groups, cultural contexts, or health conditions is unknown.

The active assessments occurred only every three months, providing just four data points per participant across the study period. This sparse ground truth limits the ability to validate predictions at finer time scales. The AI might be capturing seasonal trends or broad lifestyle patterns rather than the week-to-week or day-to-day fluctuations that would be most clinically useful.

The study does not report individual-level variation in prediction accuracy. An average 12.5 percent error could mask wide variation, with excellent predictions for some individuals and poor predictions for others. Understanding which personal characteristics are associated with better or worse model performance is critical for any eventual clinical application.

The next phase of research, already underway, will extend data collection to 24 months and examine individual differences in model performance. That longer timeline and more granular analysis should provide a clearer picture of whether passive wearable monitoring can deliver on its clinical promise or remains a research curiosity with appealing statistics.