CROSS-LINGUAL TRANSFORMER-BASED SCREENING OF POST-TRAUMATIC STRESS DISORDER BASED ON COMPARATIVE ANALYSIS OF BERT AND XLM-ROBERTA WITH MACHINE TRANSLATION ADAPTATION FOR UKRAINIAN LANGUAGE

Abstract

The paper presents a comprehensive study on automated screening of post-traumatic stress disorder (PTSD) using transformer-based natural language processing models in a cross-lingual setting. The research aims to evaluate the feasibility of deploying an intelligent PTSD detection system in Ukraine under conditions of limited, localised training data. A balanced corpus of 4,822 text records was constructed by aggregating publicly available PTSD-related datasets, including 2,042 PTSD-positive texts and 2,780 control texts representing neutral content and other psychological conditions. The study compares the performance of the English-language BERT (bert-base-uncased) model and the multilingual XLM-RoBERTa (xlm-roberta-base) model applied to a Ukrainian-language corpus generated via machine translation using the Google Translate API and large language models for complex structures. Word cloud visualisation and semantic analysis confirmed preservation of core psychological markers during translation. Experimental results demonstrate high predictive performance for both architectures. The English-language model achieved an Accuracy of 0.90 and an ROC-AUC of 0.962. In contrast, the Ukrainian-language model achieved an Accuracy of 0.85 and an ROC-AUC of 0.940, significantly outperforming existing Ukrainian multi-class stress detection models (Accuracy ~0.45) and exceeding standard multilingual mental health benchmarks (0.78–0.82), establishing a robust state-of-the-art baseline for Ukrainian clinical NLP. Importantly, Recall remained identical (0.88) across both language settings, indicating strong sensitivity to PTSD markers despite translation-induced lexical noise. The minimal AUC degradation (2.3%) confirms the robustness of transformer architectures to cross-lingual adaptation. The findings validate the viability of combining machine translation with multilingual transformers for the rapid deployment of mental health screening systems in low-resource language environments. The proposed pipeline enables scalable and cost-effective digital PTSD monitoring while maintaining clinically relevant diagnostic sensitivity.