Speech token detection and discrimination in individual infants using functional near-infrared spectroscopy

Abstract

Speech detection and discrimination ability are important measures of hearing ability that may inform crucial audiological intervention decisions for individuals with a hearing impairment. However, behavioral assessment of speech discrimination can be difficult and inaccurate in infants, prompting the need for an objective measure of speech detection and discrimination ability. In this study, the authors used functional near-infrared spectroscopy (fNIRS) as the objective measure. Twenty-three infants, 2 to 10 months of age participated, all of whom had passed newborn hearing screening or diagnostic audiology testing. They were presented with speech tokens at a comfortable listening level in a natural sleep state using a habituation/dishabituation paradigm. The authors hypothesized that fNIRS responses to speech token detection as well as speech token contrast discrimination could be measured in individual infants. The authors found significant fNIRS responses to speech detection in 87% of tested infants (false positive rate 0%), as well as to speech discrimination in 35% of tested infants (false positive rate 9%). The results show initial promise for the use of fNIRS as an objective clinical tool for measuring infant speech detection and discrimination ability; the authors highlight the further optimizations of test procedures and analysis techniques that would be required to improve accuracy and reliability to levels needed for clinical decision-making.

Conflict of interest statement

The authors declare no competing interests.

© 2021. The Author(s).

Figures

Figure 1

Illustration of experimental set-up on the infant. Left: an image of the cap on a participant’s head. Right: The 8 × 8 source-detector montage that was placed over the temporal and prefrontal regions of the infant participants, resulting in four regions of interest. Sources are colored red, detectors are blue, and channels are indicated by purple lines.

Figure 2

Sensitivity profile and sensor channel locations for the left temporal and prefrontal regions. The sensitivity of each probe to detecting brain hemodynamics is represented on a logarithmic color scale ranging from − 2 to 0 (arbitrary units; higher number indicates higher sensitivity). Source and detector positions are indicated by the red and blue circles and corresponding labels respectively, and yellow lines show channels. This figure was generated with AtlasViewer.

Figure 3

Representation of the sound presentation protocol used in this study. (A) Details of the test run sequence and timing for each infant in the study. Each two-color segment represents a test run of a specific speech contrast pair. Note that the final 6 infants received only one contrast pair, and the test run was modified to replace the “Post-Nov” condition with 1 min of silence. (B) Representation of the stimulus protocol for a single experimental test run. One run comprised four sections labeled as “Hab 1”, “Hab 2”, “Novel” and “Post-Nov”, each containing 5 stimulus blocks. The habituation speech sound was presented in the blocks labeled “Hab 1”, “Hab 2” and “Post-Nov”, while the novel speech sound was presented in the blocks labeled “Novel”.

Figure 4

An example of TDDR and wavelet denoiser’s effect on the data. (A) The raw data processed with only optic density and modified beer-lambert law conversion, demonstrating spike-like artefacts from movements. (B) The data processed with only optical density, modified beer-lambert law and a bandpass filter (0.01 to 0.25 Hz), with the spikes still present and smoothed out from the filtering. (C) Data cleaned first by TDDR and wavelet denoising before standard processing as in (A), with visible reduction in artefacts.

Figure 5

Grand average responses acquired at each recording channel. Each source and detector is indicated by red and blue text (S1–S8, D1–D8) respectively. Grey bars show when the sound stimulus was presented (five seconds duration), and shaded bars show one standard deviation across participants. Each region of interest is indicated in black text (LF, LT, RF and RT for left prefrontal, left temporal, right prefrontal, right temporal respectively) and consists of groups of four channels in their corresponding locations.

Figure 6

Group HbO responses at each ROI and to the Hab 1, Hab 2 and Novel stimulus conditions. Each ROI’s responses are in a different panel, with the ROI indicated by highlight on the head montage illustration. Each trace represents the response averaged across participants.

Figure 7

Difference between “Novel” and “Hab 2” conditions for each speech token contrast pair. (A) Shows the group average responses for each speech token (one color each); the “Hab 2”and “Novel" responses are represented by the broken and solid traces respectively. (B) Shows the calculated response size difference as a bar graph; error bars are one standard error of the mean across participants. Labels on x-axis indicates the number of participants who were delivered that particular speech token pair stimuli. There is no significant difference between discrimination response sizes.

Figure 8

Individual infant detection and discrimination analysis. (A,B) Show examples from one participant SD011 of detection responses averaged across all channels and conditions for true (A) and control (B) stimulus blocks (shaded areas are 1 SEM). (C,D) Show the true and control discrimination responses respectively. (E) Shows the ROC curve, calculated across all 23 participants, for detection and discrimination response discovery respectively.

Figure 9

Comparison of responses to two sound presentation strategies. (A) Shows the “alternating” protocol, where two speech token pairs are presented in alternating runs. (B) Shows the “repeating” protocol, where the same speech token pair was repeated in five sequential runs. Shaded error bars show one standard deviation across participants. (C) Shows the calculated response sizes for both protocols as bar plots, with error bars showing one standard error of the mean; test statistics were calculated with these response sizes.