Dr Rahul Khanna

Abstract

Background

Post-Traumatic Stress Disorder (PTSD) manifests heterogeneously across cognitive, affective, physiological, and behavioral domains. Current gold-standard assessment methods face scalability limitations, creating barriers to timely diagnosis and intervention. Digital assessment paradigms offer potential solutions, but their diagnostic accuracy and feasibility need validation, particularly for at-risk populations such as military and police personnel. Aims & Objectives: This study aimed to develop and validate a digital assessment paradigm for PTSD using a contemporary theoretical framework with scalable deployment potential. Secondary objectives included examining whether remote assessment approaches could replicate previously identified differences between PTSD and control groups, collectively determining which components provide the greatest diagnostic utility.

Method

We employed an exploratory cross-sectional multi-component design with 42 participants from Australian Defence Force and police backgrounds, validated against the CAPS-5 structured clinical interview. The paradigm incorporated cognitive tasks (dot probe, CRAFT), physiological assessments (heart rate, blood pressure, facial affect during trauma discussions and exposure), and longitudinal assessment (Ecological Momentary Assessment over four days). Data were analyzed using mixed-effects modeling to examine group differences across tasks, and machine learning classification to integrate multi-modal data for optimizing diagnostic accuracy. Models were evaluated with cross-validation techniques and interpretability analyses.

Results

Several components demonstrated significant diagnostic value. The CRAFT task revealed PTSD participants responded significantly faster to disgust and fear faces compared to controls. During trauma narrative discussions, those with PTSD showed distinct emotional expression patterns, particularly slower resolution of negative affect. Linguistic analysis showed PTSD narratives contained significantly more negative emotional content, first-person singular pronouns, and less lexical diversity. EMA data revealed consistently higher symptom levels in the PTSD group across all time points. The machine learning classifier achieved an AUC of 0.75 on the test set. EMA-derived features, particularly symptom volatility metrics, contributed approximately 60% of the model’s predictive power, followed by narrative (25%) and acoustic features (10%).

Discussion & Conclusions

This study demonstrates the feasibility and utility of a multi-modal digital assessment approach for PTSD. The strongest performing features were also those least burdensome participants. These findings suggest that scalable digital assessment paradigms may be able to effectively identify PTSD, potentially enabling earlier intervention. Future refinements should focus on optimizing the most diagnostically valuable components, particularly dynamic symptom monitoring and linguistic analysis, to develop lighter-weight assessment protocols suitable for frontline deployment.

It was a delight to be invited by the Australian Digital Health Agency to lead a discussion on what good clinical governance should look like in the artificial intelligence age. Drop me a line if you’d like to discuss it - things are moving so quickly in the field that it’d be silly to leave slides up without context!

Abstract

Computational psychiatry offers promising opportunities for understanding and treating mental health disorders, yet achieving ecological validity—the accurate reflection of real-world experiences—remains a critical challenge. This perspective examines how digital health technologies can enhance ecological validity in computational psychiatry while addressing barriers in data collection, participant representation, validation, engagement, and methodological integration. We review key approaches, including digital phenotyping and adaptive design optimization, that enable more naturalistic data collection. However, achieving representative sampling and mitigating algorithmic biases remain unresolved challenges, particularly in AI-driven assessments. We discuss how expert-by-experience collaboration, systematic validation efforts, and structured open science practices can improve model generalizability and clinical applicability. Additionally, we explore the role of federated learning and edge computing in balancing privacy with robust, scalable model development. The paper concludes by integrating these challenges and solutions within a broader methodological framework, emphasizing the need for interdisciplinary approaches that bridge computational precision with real-world psychiatric care.

It was lovely to be interviewed by Tim Elliott at Good Weekend Magazine on this thoughtful piece titled In Bot We Trust in the edition published 21st March 2025. I recommend checking it out here (paywall) if you’re curious about where our digital relationships may take us.