Fig 1.
Illustration of the smartphone-based ambulatory voice monitor that uses a neck-surface accelerometer attached to the skin halfway between the thyroid prominence and the suprasternal notch of a female subject.
Fig 2.
Representation of the subglottal system.
(a) Accelerometer position and sub1 and sub2 system parts. (b) A mechano-acoustic analogy of the subglottal system including load impedance from skin. Reproduced with permission.
Table 1.
Occupations and mean age of adult females with PVH and matched-control participants analyzed (48 pairs).
Table 2.
Frame-based glottal airflow measures estimated from the ambulatory neck-surface accelerometer signal using impedance-based inverse filtering.
Fig 3.
Example of ambulatory IBIF analysis.
(A) Estimated glottal airflow waveform and (B) its derivative, showing how time-domain measures were derived per glottal cycle. Measures were then averaged over all cycles to yield a single value per frame for each time-domain measure.
Fig 4.
Spectrum of the frame in Fig 3(A).
Fig 5.
Feature extraction and classification process for 96 subjects.
Table 3.
Top 11 week-long summary statistics (from a total of 77) sorted by p-value from the 48 paired t-tests.
Statistically significant differences (*) were found by applying the Benjamini-Hochberg method using a false discovery rate of 0.1.
Table 4.
Classification performance of L1 logistic regression (L1-LR) and support vector machine (SVM) approaches for 96 subjects using IBIF features.
Mean (standard deviation) is reported for the performance metrics. Previous results using 51 pairs [3] and 20 pairs [16] are also shown. It is worth noting that the distribution of metrics such as AUC, across all models, may be non-normal and may benefit from other summary statistics such as median (IQR).
Fig 6.
Performance results across subject pairs with L1-logistic regression.
Area Under the ROC Curve (AUC), Accuracy, Sensitivity, Specificity. The red crosses indicates the average value for each performance metric.
Fig 7.
Classification results from L1-logistic regression.
The threshold (blue line) at 0.57 classifies correctly 79 from 96 subjects (82.3%).
Fig 8.
F-score distributions from Table 5.
From all 26 features (rightmost box plot) to only one feature (H1-H2 95th%, leftmost box plot).
Table 5.
Association count of Beta (weight) variables that were included in all 48 models.
These 26 features were present in each logistic regresion model.
Table 6.
Mean and (standard deviation) performance metrics from L1-logistic regression for different group of features from Table 5, starting with the whole set of 26 features.
Iteratively, the following group is obtained by taking out the feature with the smallest absolute Beta value.
Fig 9.
Odds ratio association with phonotraumatic subjects.