<p>Traditional psychiatric assessments often depend on self-reports, which may be biased, while real-world evaluations pose safety and feasibility challenges. Specific phobias are underdiagnosed and often comorbid, with affected individuals at elevated risk for severe outcomes. This study aimed to integrate the DSM-5 diagnostic framework, Virtual Reality (VR), and Machine Learning (ML) to identify multimodal features most predictive of phobia severity. Ninety-four adults completed 100 independent trials across four VR scenarios (neutral, Cynophobia, Astraphobia, combined), during which behavioral, physiological, demographic, and self-reported measures were recorded alongside DSM-5 severity ratings. Feature selection and ML analyses revealed that Cynophobia severity was best predicted by a multimodal subset comprising age, neutral-scenario task completion time, Cynophobic virtual distance, oxygen-level variations, and DSM-5 Astraphobia severity. The next most predictive subset comprised sense of presence and DSM-5 Astraphobia severity. The Naïve Bayes classifier achieved robust performance across all features, underscoring the complementary value of multimodal inputs. These findings suggest a potential comorbidity between Cynophobia and Astraphobia and demonstrate the diagnostic relevance of integrating VR-based behavioral and physiological measures with DSM-5 criteria. The study contributes an integrated and scalable framework for enhancing the objectivity and reliability of phobia assessment and highlights future potential for VR–AI applications in clinical diagnostics.</p> Graphical abstract <p></p>