Machine learning based identification of anxiety disorders using autonomic biomarkers: a diagnostic cross-validation study

Background

This study aimed to evaluate the validity and reliability of autonomic biomarkers derived from heart rate variability (HRV) to correctly classify cases with and without anxiety disorder symptoms.

Methods

This multi-center diagnostic cross-validation study involved 124 participants at two research labs, Sussex and Sheffield. Participants were exposed to a three-phase anxiety induction paradigm lasting 9 minutes (3 min. rest, 3 min. induction, 3 min. recovery). Participants completed measures of anxiety (GAD-7) and depression (PHQ-9) prior to the lab task. Their HRV was measured continuously using: [1] an industry-standard electrocardiogram (ECG) and [2] a miniaturized photoplethysmography (PPG) biosensor. Correlations between [1] and [2] were examined. A machine learning classifier was developed using ECG and PPG-derived features in a training sample from Sussex (n = 88) to identify cases with anxiety disorders (GAD-7 ≥ 8). Classification accuracy was evaluated using the area-under-the-curve (AUC) in the validation sample from Sheffield (n = 36). Correlations using dimensional scores were used to evaluate the machine learning model’s convergent (GAD-7) and discriminant validity (PHQ-9).

Results

HRV features from PPG and ECG were strongly correlated across all three phases (r ≥ .88, p < .001). An ECG-based machine learning classifier generalized to an external validation sample (AUC= .71), correctly classifying anxiety disorder cases with adequate sensitivity (88%) and specificity (61%) and outperforming a PPG-based classifier. As expected, the classifier had marginally higher correlations with GAD-7 (r= .51) than PHQ-9 (r= .47).

Conclusion

Biomarker data processed using a machine learning algorithm converges with an established measure of anxiety. This supports the potential use of HRV-derived biomarkers to monitor changes in real-time, which could be highly informative for exposure-based therapies.