V-068 | NLP signatures of episodic memory difficulties in persons with human immunodeficiency virus

Cognition, Behavior, and Memory
Author: Lucas Federico Sterpin | Email: sterpinlucas@gmail.com

Lucas Sterpin^1°, Jeremías Inchauspe^1°, Camilo Avendaño^1°;4°, Gonzalo Pérez^1°;2°;5°, Franco Ferrante^1°;2°;5°, Lucía Amoruso^1°;3°, Lorena Abusamra^6°, Bárbara Sampedro^7°, Valeria Abusamra^8°, Adolfo M. García^1°;9°;10°

^1° Cognitive Neuroscience Center, Universidad de San Andrés, Argentina
^2° National Scientific and Technical Research Council, Argentina
^3° Basque Center on Brain, Language and Cognition, Spain
^4° Universidad de Santiago de Chile, Chile
^5° School of Engineering, Universidad de Buenos Aires, Argentina
^6° Hospital Dr. Diego Thompson, Buenos Aires, Argentina
^7° School of Linguistics, Universidad de Buenos Aires, Argentina
^8° Centro Interdisciplinario de Investigaciones en Psicología Matemática y Experimental “Dr. H. Rimoldi” (CIIPME- CONICET), Buenos Aires, Argentina
^9° Global Brain Health Institute, University of California, USA
^10° School of Linguistics and Literature, Universidad de Santiago de Chile, Chile

HIV-associated neurocognitive disorders are a rising cause of morbidity in people living with HIV (PLWH). Among other symptoms, PLWH face verbal episodic memory (EM) deficits, even under treatment, affecting quality of life. Retelling tasks are often used to assess EM, but standard measures rely on decontextualized stimuli and manual scoring based on predefined correct responses. Here we introduce an automated, granular, ecologically valid NLP approach for assessing EM in PLWH. We asked 50 PLWH and 42 matched controls to complete a validated retelling task. The original text and each participant’s retelling were run through NLP algorithms to extract key features for each content word (CW, namely, nouns, verbs, adjectives, adverbs). These comprised (i) the ratio of each CW; (ii) the semantic distance for each CW class (cosine similarity between the average embedding of words in each class in the original text and retellings); and (iii) the topological distance for each CW class via differences between the original text and each retelling in relevant speech graph measures tapping on text connectivity, repetitions, and global structural properties. Robust ANOVAs showed that PLWH were characterized by fewer nouns, larger semantic distance across CW classes, and larger topological distances in specific connectivity, repetition, and global structural measures. These results suggest that our NLP approach can reveal fine-grained differences in EM that escape traditional methods.