AbstractAnxiety and stress-related disorders are highly prevalent and debilitating conditions that impose an enormous burden on society. Sensitive measurements that can enable early diagnosis could mitigate suffering and potentially prevent onset of these conditions. Self-reports, however, are intrusive and vulnerable to biases that can conceal the true internal state. Physiological responses, on the other hand, manifest spontaneously and can be monitored continuously, providing potential objective biomarkers for anxiety and stress. Recent studies have shown that algorithms trained on physiological measurements can predict stress states with high accuracy. Whether these predictive algorithms generalize to untested situations and participants, however, remains unclear. Further, whether biomarkers of momentary stress indicate trait anxiety – a vulnerability foreshadowing development of anxiety and mood disorders – remains unknown. To address these gaps, we monitored skin conductance, heart rate, heart rate variability and EEG in 39 participants experiencing physical and social stress and compared these measures to non-stressful periods of talking, rest, and playing a simple video game. Self-report measures were obtained periodically throughout the experiment. A support vector machine trained on physiological measurements identified stress conditions with ~96% accuracy. A decision tree that optimally combined physiological and self-report measures identified individuals with high trait anxiety with ~84% accuracy. Individuals with high trait anxiety also displayed high baseline state anxiety but a muted physiological response to acute stressors. Overall, these results demonstrate the potential for using machine learning tools to identify objective biomarkers useful for diagnosing and monitoring mental health conditions like anxiety and depression.
A machine learning approach to identifying objective biomarkers of anxiety and stress
