Memory Assessment in Temporal Lobe Epilepsy Patients Using Natural Language Processing Applied to Spontaneous Speech
Abstract number :
3.342
Submission category :
11. Behavior/Neuropsychology/Language / 11A. Adult
Year :
2023
Submission ID :
1013
Source :
www.aesnet.org
Presentation date :
12/4/2023 12:00:00 AM
Published date :
Authors :
Presenting Author: Eden Tefera, – NYU Langone Health
William Barr, PhD, ABPP – Professor, Neurology, NYU Langone Health; Kristie Bauman, MD – Neurology – NYU Langone Health; Helen Borges, MA – NYU Langone Health; Amadou Camara, MD – NYU Langone Health; Hunaid Hassan, MD – NYU Langone Health; Simon Henin, PhD – Associate Research Scientist, Neurology, NYU Langone Health; Stephen Johnson, PhD – Professor, Population Health, NYU Langone Health; Joshua Larocque, MD PhD – NYU Langone Health; Anli Liu, MD, MA – Associate Professor, Neurology, NYU Langone Health; Haley Peters, BS – NYU Langone Health; Peem Teerawanichpol, BFA – New York University Arts and Science; Charlotte Blewitt, MS – New York University Arts and Science; Aaqib Mansoor, N/A – New York University Arts and Science
Rationale: Epilepsy patients rank memory problems as their most significant cognitive comorbidity, impacting daily function. Current clinical assessments are laborious to administer/score and may be an insensitive measure of subjective memory changes over time. Previously, left temporal lobe epilepsy (LTLE) patients, compared to right TLE (RTLE) patients and healthy controls (HC) have poorer name and remote recall as evaluated by the Famous Faces Task (FF). Natural Language Processing (NLP) uses computational methods to analyze linguistic structure and has potential as a precise, objective, and automated tool in assessing memory impairment. We applied NLP to spontaneous recall of FF to measure the depth of memory impairment in LTLE, RTLE, and HCs.
Methods: We recruited Epilepsy patients and Healthy Controls (HCs) ages 18-60 from a single center. Epilepsy diagnosis was determined by a combination of EEG, MRI, and/or seizure semiology. Twenty FF in politics, sports, and entertainment (active 2008 through 2017) were displayed by laptop, we asked subjects to spontaneously recall as much biographical detail as possible. To control for speech output, we administered the Cookie Theft (CTT) task from the BDAE. Subjects’ responses to FF and CT tasks were transcribed and total & content word count were determined using the Google Colab (SpaCy NLP library). The human-generated dictionary was compiled by two human raters who independently extracted keywords. The automated dictionary was trained on transcripts of a random subset of HCs and generated using TF-IDF (term frequency - inverse document frequency). We calculated descriptive and ANOVA statistics on the demographic details and outcome measures by group using Stata.
Results: Seventy-three (73) adults were included in this study: 28 LTLE (60% F, 23 RH, mean 31.1±9.1 y), 18 RTLE (61% F, 14 RH, 33.9±7.3 y) and 27 HC (67% F, 25 RH, 28.2±9.0 y) (Table 1). During recall of FF, LTLE patients generated fewer total words (798.36± 430.07) than RTLE patients (1394.17± 1292.47 (p=0.0286) (Fig 1). LTLE patients spoke fewer content words (291.36± 174.08) than RTLE patients (497.94± 478.55, p=0.0425) (Fig 2). When scored by the human-generated keyword dictionary, LTLE patients recalled fewer keywords (24.09± 12.26) compared to HCs (31.03± 11.80, p=0.0371) (Fig 3). When scored by the automated keyword dictionary, the percentage of keywords identified between the 3 groups differed (p = 0.046, Fig 1). LTLE patients recalled a smaller percentage of keywords (24.40± 11.76) compared to HCs (32.11± 12.89, p=0.0245) (Fig 4). There were no group differences in CTT total words (p=0.0799) (Fig 5) or content words (p= 0.099) (Fig 6).
Conclusions: LTLE patients recalled fewer details than HCs, and generated fewer words compared to RTLE patients during recall of FFs. These results replicate and extend clinically expected group-level differences by applying NLP to spontaneous speech. Automated approaches applying NLP to spontaneous spoken recall may capture remote memory impairment.
Funding: K23-NS 104252-03 (Liu)
NYU Finding A Cure for Epilepsy and Seizures (FACES).
Behavior