Abstracts

Rationale: Patients with epilepsy (PWE) and Alzheimer’s dementia report difficulty with episodic and semantic memory. Current clinical tests have limited ecological validity and likely miss subtle changes in memory over time. The Famous Faces (FF) Task has previously measured poorer name and remote recall in left temporal lobe epilepsy (LTLE) patients, compared to right TLE (RTLE) patients and healthy controls (HC). We explored the use of Natural Language Processing (NLP), the application of computational methods to text to analyze linguistic structure, to measure the depth of memory impairment during spontaneous recall of FF. NLP applied to spontaneous recall could become a sensitive, objective, and automated means of measuring memory impairment.

Methods: PWE and HCs were recruited from a single center. Epilepsy diagnosis was determined by seizure semiology, EEG, and/or MRI. Twenty FF in politics, sports, and entertainment publicly (active 2008-2017) were displayed by laptop, then subjects were asked to spontaneously recall as much biographical detail as possible. We used the Cookie Theft (CT) task from the BDAE to control for speech output. Subjects’ responses to FF and CT tasks were transcribed, then analyzed via NLP using Python (with spaCy package), yielding the number of total words, content words, interjections, interruptions, repetitions, and pauses for each response. Two keyword dictionaries were compiled - one with biographical details extracted and verified by a human rater; the second with the top 30 keywords generated automatically using TF-IDF (term frequency - inverse document frequency). We calculated descriptive and ANOVA statistics on the NLP-derived linguistic measures and biographical details by group.

Results: Forty-eight (48) adults were included in this study: 20 LTLE (70% F, 17 Right-handed, mean 33.2±10.4 y), 12 RTLE (58% F, 9 RH, 33.9±7.8 y) and 16 HC (68% F, 16 RH, 26.3±7.3 y) (Table 1). When scored by the human-generated keyword dictionary, the number of biographical details generated between the 3 groups differed (p = 0.045, Fig 1). LTLE patients recalled fewer key biographical details (41.5±17.2) compared to RTLE patients (62.7± 28.3, p=0.038) (Fig 1). There was a trend toward group-level differences when FF responses were scored by the automated dictionary (p=0.077, Fig 1). There was a trend toward group differences in FF total words (p=0.061) and content words (p=0.08). There was a trend toward group differences in CT total words (p=0.057), but not content words (p =0.146)(Fig 1).

Conclusions: LTLE patients generated fewer biographical details compared to RTLE patients during spontaneous recall of FFs. NLP applied to spontaneous spoken recall can replicate expected group level differences. Several automated approaches, including FF term frequency and word counts, may capture episodic memory impairment and depth of semantic knowledge. This novel method presents an ecologically valid and semi-automated approach to quantify depth of remote memory impairment in a patient cohort.

Funding: Please list any funding that was received in support of this abstract.: K23-NS 104252-03 (Liu)
NYU Finding A Cure for Epilepsy and Seizures (FACES).