posted on 2021-08-12, 17:17authored byPatrick C. M. Wong, Ching Man Lai, Peggy H. Y. Chan, Ting Fan Leung, Hugh Simon Lam, Gangyi Feng, Akshay R. Maggu, Nikolay Novitskiy
Purpose: This study aimed to construct an objective and cost-effective prognostic tool to forecast the future language and communication abilities of individual infants.
Method: Speech-evoked electroencephalography (EEG) data were collected from 118 infants during the first year of life during the exposure to speech stimuli that differed principally in fundamental frequency. Language and communication outcomes, namely four subtests of the MacArthur–Bates Communicative Development Inventories (MCDI)–Chinese version, were collected between 3 and 16 months after initial EEG testing. In the two-way classification, children were classified into those with future MCDI scores below the 25th percentile for their age group and those above the same percentile, while the three-way classification classified them into < 25th, 25th–75th, and > 75th percentile groups. Machine learning (support vector machine classification) with cross validation was used for model construction. Statistical significance was assessed.
Results: Across the four MCDI measures of early gestures, later gestures, vocabulary comprehension, and vocabulary production, the areas under the receiver-operating characteristic curve of the predictive models were respectively .92 ± .031, .91 ± .028, .90 ± .035, and .89 ± .039 for the two-way classification, and .88 ± .041, .89 ± .033, .85 ± .047, and .85 ± .050 for the three-way classification (p < .01 for all models).
Conclusions: Future language and communication variability can be predicted by an objective EEG method that indicates the function of the auditory neural pathway foundational to spoken language development, with precision sufficient for individual predictions. Longer-term research is needed to assess predictability of categorical diagnostic status.
Supplemental Material S1. Supplemental methods and results.
Supplemental Material S2. Neural recording age (EEG age, black circles) and age at which language outcome was collected (Outcome age, black squares) for each of the 118 participants. The age in months is plotted on the abscissa axis. The length of each line corresponds to the gap between the neural data recording and outcome assessment.
Supplemental Material S3. Experimental setup. During the experiment the infant participants were typically sleeping and being held by the caregiver. The auditory stimuli were delivered via plastic tubes to the infant’s ears and the EEG was recorded from the electrodes gently attached to its scalp.
Supplemental Material S4. Schematic of the machine-learning algorithm used to construct and validate our predictive models. Support-vector machine (SVM) models were trained and tested in 10-fold cross-validation. The statistical significance of the model was estimated by 10000 iterations of bootstrapping and permutation.
Supplemental Material S5. Adding socioeconomic status (SES) to the predictive models did not improve predictive performance.
Wong, P. C. M., Lai, C. M., Chan, P. H. Y., Leung, T. F., Lam, H. S., Feng, H., Maggu, A. R., & Novitskiy, N. (2021). Neural speech encoding in infancy predicts future language and communication difficulties. American Journal of Speech-Language Pathology. Advance online publication. https://doi.org/10.1044/2021_AJSLP-21-00077
Funding
We thank participants of the Chinese University of Hong Kong (CUHK) Stanley Ho Developmental Cohort Study for participating in this study as well as the Dr. Stanley Ho Medical Development Foundation for their funding support. This study was supported by the Innovation and Technology Fund of the Hong Kong SAR Government (Grant ITS/067/18), including funding from its research talent program (Grants InP/285/19, InP/286/19, PiH/030/19, and PiH/034/19).