Skip to main content
Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

< Back to Article

Table 1.

Number of patients in the autumn and number protocols.

More »

Table 1 Expand

Table 2.

Demographics and clinical characteristics.

More »

Table 2 Expand

Fig 1.

Illustration of the pre-processing process.

Top: entire waveform (blue line) and threshold (black line). Bottom: pre-processed waveform where only the values above the threshold were extracted.

More »

Fig 1 Expand

Fig 2.

Overview of the proposed PWSI-AI-AC.

(a) Baseline scheme: The network receives the entire waveform as an input and performs learning and inference for the binary classification. (b) Proposed scheme: Unlike the baseline method, the entire waveform is divided into N patches, and the network receives the individual patch waveform as input and learns to perform binary classification for each patch. In the inference phase, N results are synthesized through majority voting to obtain the final binary classification result for the target patient. Through this diversity gain, performance improvement can be achieved.

More »

Fig 2 Expand

Fig 3.

The log-Mel spectrogram of the waveform according to the selected dysarthria types (no evidence of dysarthria, hypokinetic dysarthria, and ataxic dysarthria) for each autumn and number protocol.

More »

Fig 3 Expand

Fig 4.

Configuration of CNN14.

CNN14 is a representative network developed for audio-based classification, which converts input waveform into a log-Mel spectrogram (i.e., converts one-dimensional information into two-dimensional image type information) and makes the CNN appended to the spectrogram. The CNN is composed of convolution, batch normalization, and FC layers. In this study, the output size of the last layer was resized as two for binary classification, and transfer learning was adopted.

More »

Fig 4 Expand

Table 3.

Macro-average AUCs for the proposed PWSI-AI and baseline model.

More »

Table 3 Expand

Fig 5.

Case 1: Hypokinetic dysarthria vs. others.

Comparison of the performance between the proposed and baseline schemes in terms of AUC. (a) Autumn + proposed, (b) Autumn + baseline, (c) Number + proposed, (d) Number + baseline.

More »

Fig 5 Expand

Fig 6.

Case 2: Ataxia vs. others.

Comparison of the performance between the proposed and baseline schemes in terms of AUC. (a) Autumn + proposed, (b) Autumn + baseline, (c) Number + proposed, (d) Number + baseline.

More »

Fig 6 Expand

Fig 7.

Case 3: Hypokinetic dysarthria vs. ataxia.

Comparison of the performance between the proposed and baseline schemes in terms of AUC. (a) Autumn + proposed, (b) Autumn + baseline, (c) Number + proposed, (d) Number + baseline.

More »

Fig 7 Expand

Table 4.

AUCs, accuracy, and precision for the proposed PWSI-AI.

More »

Table 4 Expand

Table 5.

Confusion matrices for the proposed PWSI-AI.

More »

Table 5 Expand

Fig 8.

Comparison of the performance between the proposed PWSI-AI and doctors in terms of the AUC.

(a) and (c) Autumn protocol, (b) and (d) Number protocol.

More »

Fig 8 Expand

Table 6.

Comparison between the proposed PWSI-AI and doctors.

More »

Table 6 Expand