Evaluation of classification criteria for the detection of cortisol pulses in repeated-measures designs

Abstract

Rationale : Hypothalamus–pituitary–adrenal (HPA) axis reactivity, which has been considered a potential endophenotype for psychiatric disorders, is commonly investigated by repeated-measures designs utilizing frequent sampling of salivary cortisol in temporal proximity to psychosocial stressors. To remove sources of cortisol variance, which are not related to HPA axis reactivity, researchers often utilize classification criteria to identify individuals who show no cortisol response (non-responders), for example, baseline-to-peak distances of 2.5 nmol/l. However, such classification criteria have not been systematically evaluated with regard to their classification performance. Methods : As a first step, we fitted an autoregressive latent trajectory model to cortisol data, which was obtained from longitudinally sampled saliva of 504 participants, of which 309 were exposed to the Trier Social Stress Test. Different sources of time-series variance were accounted for by modeling of initial cortisol levels, amplitude of the subsequently occurring secretory episodes, and continuous cortisol elimination. Assuming zero-amplitudes for individuals who show no stress response, a mixture distribution was implemented for secretory episodes, resulting in appropriate classifications of cortisol responders, or non-responders. Then, as a second step, we evaluated the classification performance of various proposed classifiers by constructing receiving operator characteristics. Results : Results reveal (a) that covariance and mean structure of cortisol time-series can be sufficiently accounted for by the proposed model, allowing to infer on endocrine parameters that can barely be extracted by conventional analyses and (b) that the 2.5 nmol/l criterion is suboptimal in terms of simultaneously minimizing false-positive and false-negative classifications and inferior as opposed to other classifiers. Conclusion : To maintain the low number of false positives, but to increase true-positive classifications, we suggest to lower the conventional baseline-to-peak classification threshold to 1.5 nmol/l. Furthermore, classification performance can be increased by adjusting baseline-to-peak differences for initial cortisol levels.

Keywords

• autoregressive latent trajectory model
• cortisol
• longitudinally
• Trier Social Stress Test
• time-series variance