Neurophysiological and cognitive markers for panic disorder: a multimodal predictive modeling study

This is the authors' abstract. We don't add key points for this paper.

Background Panic disorder (PD) is characterized by heightened physiological arousal and disrupted cognitive control, yet its assessment still relies predominantly on subjective self-report scales. Resting-state EEG has been suggested as a promising physiological marker of PD, but findings across frequency bands have been inconsistent, and few studies have integrated EEG with other relevant modalities such as heart rate variability (HRV) and cognitive performance. Moreover, limited work has systematically evaluated these modalities within a unified predictive framework. Establishing reliable multimodal biomarkers that integrate neurophysiological, autonomic, and cognitive domains is therefore needed to support more accurate and biologically grounded assessment of PD. Aims & Objectives This study aimed to determine which multimodal features, and in which combinations, best distinguish individuals with PD from healthy controls (HC), with a focus on identifying EEG frequency bands that contribute meaningful discriminative value. Method Participants included 28 individuals with PD and 25 age-matched HC. All participants completed resting-state EEG recording, HRV assessment, the Cambridge Neuropsychological Test Automated Battery (CANTAB), and clinical evaluations including the Beck Anxiety Inventory (BAI) and the Panic Disorder Severity Scale (PDSS). Resting-state EEG data were preprocessed and log-transformed, after which band-specific absolute power and asymmetry measures were computed. HRV indices and CANTAB cognitive metrics were also obtained. We first conducted unimodal receiver operating characteristic (ROC) analyses to screen candidate predictors (AUC ≥ 0.65). Subsequently, least absolute shrinkage and selection operator (LASSO) logistic regression was applied to integrate EEG, HRV, cognitive, and clinical measures within a unified feature-selection framework. Final models were estimated using logistic regression, and performance was assessed using AUC, 2,000-sample bootstrap confidence intervals, cross-validated AUC, and calibration curves. Sensitivity analyses excluding participants receiving psychotropic medication were also performed. Results Among all EEG frequency bands examined, high-beta and gamma activity yielded the most robust discriminative models for distinguishing PD from healthy controls. In the high-beta band, the LASSO-selected model retained fronto-central asymmetry, central absolute power, frontal asymmetry, and cognitive indices reflecting verbal memory (VRM immediate recall) and inhibitory control (Stop Signal Task reaction time), demonstrating strong classification performance (AUC = 0.89; bootstrap 95% CI: 0.80–0.97; cvAUC = 0.89). Gamma-band models exhibited a similar multimodal signature, with contributions from fronto-central asymmetry, frontal power, VRM recall, and SST reaction time, yielding comparable performance (AUC = 0.86; bootstrap 95% CI: 0.74–0.99; cvAUC = 0.888). The discriminative accuracy of these multimodal models approached that of established clinical symptom scales, including the PDSS (AUC = 0.98) and BAI (AUC = 0.96). Model calibration was acceptable, and sensitivity analyses excluding medicated participants yielded similar AUCs. Discussion & Conclusions High-beta and gamma activity in fronto-central and frontal regions reliably distinguished patients with PD from HC. When combined with cognitive measures of verbal memory and inhibitory control, these EEG features produced accurate and stable classification models. Importantly, their discriminative performance approached that of established symptom scales, highlighting their potential as objective, physiology-based markers. These findings suggest that high-frequency fronto-central dysfunction, together with cognitive inefficiency, reflects core neurophysiological characteristics of PD. Multimodal models integrating EEG and cognition may offer a pathway toward more objective, biologically grounded diagnostic tools.