Explainable CNN-attention network (C-attention network) architecture for automated detection of Alzheimer's disease

Patent metadata

Field	Value
Publication number	US-12390149-B2
Application number	US-202118022981-A
Country	US
Kind code	B2
Filing date	Aug 24, 2021
Priority date	Aug 24, 2020
Publication date	Aug 19, 2025
Grant date	Aug 19, 2025

Abstract

Three artificial intelligence (AI) linguistic processing architectures are proposed for early detection of Alzheimer's Disease based entirely on a patient's language abilities. Three C-Attention network architectures are presented: one that uses only PoS features, one that uses only the latent features (e.g., language embeddings) and a unified architecture, which uses both features.

First claim

What is claimed is:

1. A method of speech analysis for medical diagnostics, comprising the steps of:
a) obtaining a speech sample from a patient;
b) classifying said speech sample with a plurality of parts of speech tags;
c) applying a self-attention module to said plurality of parts of speech tags;
d) applying an attention layer to said plurality of parts of speech tags;
e) applying a convolution layer to said plurality of parts of speech tags;
f) applying a softmax layer to said plurality of parts of speech tags;
g) generating an intra-class explanation of said speech sample from data derived from the performance of steps c) through f); and
h) determining a diagnosis of a medical condition.

2. The method of claim 1 , wherein said self-attention module comprises a multi-head attention layer.

3. The method of claim 2 , wherein said multi-head attention layer comprises six layers.

4. The method of claim 2 , wherein said multi-head attention layer is configured to apply scaled dot product attention.

5. The method of claim 1 , wherein said medical condition is Alzheimer's disease.

6. The method of claim 1 , further comprising the step of applying a class weight correction to said speech sample.

7. The method of claim 1 , wherein said steps d) and e) are performed simultaneously.

8. The method of claim 1 , wherein said diagnosis is based on said intra-class explanation.

9. A method of speech analysis for medical diagnostics, comprising the steps of:
a) obtaining a speech sample from a patient;
b) representing said speech sample as universal sentence embeddings;
c) applying a self-attention module to said universal sentence embeddings;
d) applying an attention layer to said universal sentence embeddings;
e) applying a convolution layer to said universal sentence embeddings;
f) applying a softmax layer to said universal sentence embeddings;
g) generating an intra-class explanation of said speech sample from data derived from the performance of steps c) through f); and
h) determining a diagnosis of a medical condition.

10. The method of claim 9 , wherein said self-attention module comprises a multi-head attention layer.

11. The method of claim 10 , wherein said multi-head attention layer comprises six layers.

12. The method of claim 10 , wherein said multi-head attention layer is configured to apply scaled dot product attention.

13. The method of claim 9 , further comprising the step of applying a positional encoding module to said universal sentence embeddings.

14. The method of claim 9 , wherein said medical condition is Alzheimer's disease.

15. The method of claim 9 , further comprising the step of applying a class weight correction to said speech sample.

16. The method of claim 9 , wherein said steps d) and e) are performed simultaneously.

17. The method of claim 9 , wherein said diagnosis is based on said intra-class explanation.

18. A method of speech analysis for medical diagnostics, comprising the steps of:
a) obtaining a speech sample from a patient;
b) classifying said speech sample with a plurality of parts of speech tags;
c) representing said speech sample as universal sentence embeddings;
d) applying a first self-attention module to said universal sentence embeddings to obtain first data;
e) applying a second self-attention module to said plurality of parts of speech tags to obtain second data;
f) applying a first attention layer to said first data;
g) applying a first convolution layer to said first data;
h) applying a first softmax layer to said first data;
i) applying a second attention layer to said second data;
j) applying a second convolution layer to said second data;
k) applying a second softmax layer to said second data;
l) applying a third attention layer to said first data and said second data to obtain third data;
m) applying a dense layer to said third data;
n) applying a third softmax layer to said third data;
o) generating an intra-class explanation from said third data from data derived from the performance of steps l) through n);
p) generating an inter-class explanation from said third data from data derived from the performance of steps l) through n); and
q) determining a diagnosis for a medical condition.

19. The method of claim 18 , further comprising the step of determining relative importance between said plurality of parts of speech tags and said universal sentence embeddings.

20. The method of claim 18 , wherein said first attention module comprises a multi-head attention layer.

21. The method of claim 20 , wherein said multi-head attention layer comprises six layers.

22. The method of claim 20 , wherein said multi-head attention layer is configured to apply scaled dot product attention.

23. The method of claim 18 , further comprising the step of applying a positional encoding module to said universal sentence embeddings.

24. The method of claim 18 , wherein said medical condition is Alzheimer's disease.

25. The method of claim 18 , wherein said steps f) and g) are performed simultaneously.

26. The method of claim 18 , wherein said steps i) and j) are performed simultaneously.

27. The method of claim 18 , wherein said diagnosis is based on said intra-class explanation and said inter-class explanation.

Assignees

• Stevens Institute Of Technology

Inventors

• Subbalakshmi Koduvayur P
• Chen Mingxuan
• Wang Ning

Classifications

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• A61B5/7264

  Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems · CPC title

• A61B5/4803

  Speech analysis specially adapted for diagnostic purposes · CPC title

• G06N3/045

  Combinations of networks · CPC title