Key points are not available for this paper at this time.
Objective: To systematically evaluate the effectiveness and methodologic approaches of natural language processing (NLP) techniques for early detection of cognitive decline through speech and language analysis. Methods: We conducted a comprehensive search of 8 databases from inception through August 31, 2024, following Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines. Studies were included if they used NLP techniques to analyze speech or language data for detecting cognitive impairment and reported diagnostic accuracy metrics. Two independent reviewers (R.S. and A.B.) screened articles and extracted data on study characteristics, NLP methods, and outcomes. Results: Of 23,562 records identified, 51 studies met inclusion criteria, involving 17,340 participants (mean age, 72.4 years). Combined linguistic and acoustic approaches achieved the highest diagnostic accuracy (average 87%; area under the curve AUC, 0.89) compared with linguistic-only (83%; AUC, 0.85) or acoustic-only approaches (80%; AUC, 0.82). Lexical diversity, syntactic complexity, and semantic coherence were consistently strong predictors across cognitive conditions. Picture description tasks were most common (n=21), followed by spontaneous speech (n=15) and story recall (n=8). Crosslinguistic applicability was found across 8 languages, although language-specific adaptations were necessary. Longitudinal studies (n=9) reported potential for early detection but were limited by smaller sample sizes (average n=159) compared with cross-sectional studies (n=42; average n=274). Conclusion: Natural language processing techniques show promising diagnostic accuracy for detecting cognitive impairment across multiple languages and clinical contexts. Although combined linguistic-acoustic approaches appear most effective, methodologic heterogeneity and small sample sizes in existing studies suggest the need for larger, standardized investigations to establish clinical utility.
Shankar et al. (Wed,) studied this question.