Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Infants Neural Speech Encoding Forecasts Language Outcomes in Preschool Years: Convergence of Prediction and Validation
There is a Poster PDF for this presentation, but you must be a current member or registered to attend SNL 2024 to view it. Please go to your Account Home page to register.
Poster D36 in Poster Session D, Saturday, October 26, 10:30 am - 12:00 pm, Great Hall 4
Shaoqi PAN1,2, Ching Man Lai1, Peggy H.Y. Chan1,3, Gangyi Feng1,2, Hugh Simon Lam3, Tak Yeung Leung4, Nikolay Novitskiy1,2, Ting Fan Leung3, Patrick C.M. Wong1,2; 1Brain and Mind Institute, The Chinese University of Hong Kong, Hong Kong SAR, China, 2Department of Linguistics and Modern Languages, The Chinese University of Hong Kong, Hong Kong SAR, China, 3Department of Paediatrics, The Chinese University of Hong Kong, Hong Kong SAR, China, 4Department of Obstetrics and Gynaecology, The Chinese University of Hong Kong, Hong Kong SAR, China
Introduction: Understanding the neurobiological basis of language development is crucial for early diagnosis and intervention of childhood neurodevelopment. Early neural speech encoding (NSE), as reflected by early-latency EEG responses such as frequency following responses (FFR), provides a snapshot of auditory processing in the hearing brain (Kraus et al., 2017). These responses have shown to be neural markers of language functions in early childhood (Wong et al., 2021), as measured by the Chinese Communicative Development Inventory-Cantonese version (CCDI-C). Objectives: Building on the initial predictive results, the current study examined whether NSE can predict additional language developmental measures as well as whether the performance of the original model remains high when validated by unseen data. Methods: A cohort of 534 Cantonese-speaking children (240 females) received language assessments using the CCDI-C and/or the language subscale of the Bayley Scales of Infant and Toddler Development, third edition (Bayley-III). All children underwent EEG testing, with 486 evaluated by the CCDI-C (mean age, months: 5.3155±3.7756, 0-24) and 419 by the Bayley-III (mean age, months: 5.5482±3.623, 0-24). EEG measured neural encoding of three speech stimuli: native Cantonese Tones 2 and 4, and non-native Mandarin Tone 3, within the /ga/ syllable, extracting both early-latency responses (FFRs) and long-latency responses (LLRs). A predictive model was constructed to validate the generalizability of the NSE-based prediction model by forecasting language outcomes from 118 Propensity Score Matching (PSM)-matched unseen data. In addition, predictive models were developed to analyze the association between early NSE data and later language outcomes (CCDI-C: mean age, months: 20.2783±5.6006, 11-32; Bayley-III: mean age, months: 20.0764±7.4084, 7-36). All models were constructed by Random Forest with out-of-bag validation, in which children were classified as below or above 16th percentile (1 SD below mean) of the language outcomes. Results: Validation using unseen data showed area under curve (AUC) and sensitivity over 0.88, demonstrating the original models' robustness in predicting language outcomes. The NSE-based models accurately forecasted language outcomes assessed by CCDI-C and Bayley-III up to 36 months. The EEG->CCDI-C model achieved an AUC of 0.888±0.0078 and a sensitivity of 0.811±0.0155, while the EEG->Bayley-III model attained an AUC of 0.947±0.0047 and a sensitivity of 0.898±0.01298. Conclusion: High AUC and sensitivity values were found in predictive models constructed with NSE data, regardless of whether CCDI-C or Bayley-III were the outcome measures. When tested with unseen data, the performance of the predictive models remained high. As a language diagnosis can usually be made after 3 or 4 years of age, these findings underscore the high potential of NSE and AI analytics in predicting future language development at least 3 years earlier, which could enable prescription of preemptive intervention. Future research should focus on refining these predictive models and exploring their applications in broader neurodevelopmental and language contexts to further understand the neurobiology of language development.
Topic Areas: Language Development/Acquisition, Computational Approaches