Acoustic features of voice as a measure of cognitive load during performance of serial subtraction in a remote data collection context

Director of Research & Innovation, Francesca Cormack, presented acoustic features of voice as a measure of cognitive load during performance of serial subtraction in a remote data collection context at AAIC 2021.

Read on for the key findings and full poster.

Background

Cognitive load is the mental demand a task imposes for a specific person. Performance declines when demand exceeds capacity; therefore, increase in mental effort may precede measurable cognitive decline.
Physiological indices of load (e.g. heart rate, skin conductance etc.) are sensitive to task demand (e.g. subtracting three vs seven), show increased cognitive load with ageing and in MCI compared to healthy ageing.
Voice features have promise as non-invasive and scalable indicators of mental effort. Here, we aim to classify serial subtraction at high and low cognitive load using voice recordings captured using an automated remote data collection system.

Methods

Participants (aged 17-86 years) completed serial subtraction via the NeuroVocalix web-app (Figure 1) on their own devices. From a pool of 5,742 participants, 100 were randomly selected for manual review. Seven participants were excluded for audio or performance issues. Demographics are presented in Table 1.

Participants completed one minute each of subtraction by three and subtraction by seven, starting from 300. Subtraction by three was classed as low load and by seven as high load.

Analysis

Responses were transcribed at the start and end of each subtraction attempt marked, producing 3,398 attempts for analysis.
Low-level acoustic features were extracted and aggregated over each attempt, then normalized within participant.
Gradient Boosting classifiers were trained and evaluated using Leave-One-Subject-Out-Cross-Validation (LOSOCV) to predict high vs low load. LOSOCV repeatedly splits the dataset by subject, with attempts from one participant at a time used for testing, and the remainder used for training the model. This produces model predictions for each participant and attempt.

Results

At the participant level, average cross-validation accuracy was 0.81 (95% CI 0.78 to 0.84), with an average area under the curve (AUC) of 0.87 (95% CI 0.85 to 0.89).

At the attempt level (Figure 3) AUC was 88.49% and classification accuracy was 82.9%.

Conclusions

The results demonstrate that acoustic features can distinguish between utterances generated under conditions of high and low cognitive load, adding a novel, independent and sensitive outcome measure to a cognitive task with established utility in the context of neurodegeneration.

View poster