Scale-based Alzheimer disease detection method in cross-language environment

Abstract

A scale-based Alzheimer disease detection method in a cross-language environment comprises the steps of firstly transcribing voice data by using a whisper model, then performing quantitative scoring on preprocessed text data from seven dimensions of structure vocabulary number, irrelevant vocabulary number and the like, finally calculating a corresponding Z value, and inputting the Z value into a downstream model for detection. The Z score specifically comprises indexes such as speech richness and fluency, and the indexes can truly and meticulously reflect the cognitive condition of the patient and have no direct relation with the language, so that the method has natural cross-language generalization. The limitation of lack of AD data sets is broken through by introducing the Z score, the bottleneck of a data driving method is broken through, and the cross-language AD detection precision is improved. Meanwhile, the method not only can be applied to AD screening under various language backgrounds, but also can directly provide valuable reference basis for clinicians in the medical field needing transparency and interpretability.

Description

Technical Field

[0001] The present invention belongs to the technical field of Alzheimer's disease detection, and in particular relates to a scale-based Alzheimer's disease detection method in a cross-language environment. Background Art

[0002] Alzheimer's disease (AD) is one of the most common neurodegenerative diseases in the world, especially in the elderly. The main characteristics of AD are the decline of patients' cognitive function and daily activities, involving multiple areas such as language, memory, attention and executive function. Discovering and identifying these cognitive defects is of great significance, because timely intervention and treatment can effectively delay the deterioration of cognitive impairment and help preserve the patient's cognitive function level, thereby improving the quality of life.

[0003] To date, there are few publicly available high-quality datasets for research, especially in non-English languages. To meet this challenge, researchers have explored and applied a series of innovative technical means, such as transfer learning, self-supervised learning, unsupervised learning, and data enhancement methods. According to clinical medical research, the language manifestations of Alzheimer's disease include incoherent speech, reduced content words, and increased function words. In past AD detection studies, a large number of researchers have tried to use language pre-training models to capture AD features in patients' languages. However, since most language pre-training models have not been trained on speech abnormality-related tasks, the pre-training model, as a data-driven method, has a strong dependence on AD training data in its ability to capture AD features. However, the existing AD data is very scarce, especially in minority languages, which makes the development of AD detection in minority languages ​​very difficult. In recent years, some researchers have begun to explore cross-language AD detection, that is, training models on English, where AD data is relatively abundant, in the hope that the models will perform better in the target minority language. On the ADReSSo dataset, the embedding classification method based on traditional acoustic features and pre-training proposed by Winterlight Labs et al. achieved an accuracy of 67.61%. RaghavendraPappagar et al. used speaker and speech recognition embedding and prosodic features to achieve an accuracy of 71.7%. On the ADReSS-M Greek dataset, Chen et al. used paralinguistic features and pre-trained features to detect Alzheimer's disease across languages ​​with an accuracy of 69.57%. Shah Z et al. used domain knowledge to explore language-agnostic speech representations to detect Alzheimer's disease with an accuracy of 69.57%. In addition, integrating automatic speech recognition (ASR) technology with speech enhancement technology can also improve processing capabilities, thereby improving the recognition of AD-related speech features. Since the amount of English AD data is also very small, the language pre-training model fine-tuned on English AD data, although it performs well on English datasets, the insufficient amount of data during its training process leads to overfitting in the language, making its cross-language generalization very poor. Chen Xuchu (Chen) et al.'s research on ADReSS-M shows that the Whisper-RoBERTa system, which can have an 85% accuracy rate on English datasets, has only a 65% accuracy rate when tested cross-language on Greek. In summary, in the attempt to obtain AD cross-language detection, data-driven methods have a natural disadvantage. Summary of the invention

[0004] The purpose of the present invention is to provide a scale-based Alzheimer's disease detection method in a cross-language environment. This method combines medical scale rules and z-score standards to effectively detect AD.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A scale-based Alzheimer's disease detection method in a cross-language environment, characterized by the following specific steps:

[0007] S1. Build a four-layer fully connected neural network structure for data modeling and prediction;

[0008] S2. We first used the text dataset (ADReSS20) from the Interspeech 2020 ADReSS challenge, which includes voice recordings and transcripts of the Cookie Theft picture descriptions from the Boston Diagnostic Aphasia Examination; the transcripts were annotated using the CHAT coding system; then we used the dataset (ADReSSo) designed for detecting Alzheimer's disease through speech features from the ADReSSo challenge; and finally, we used the ADReSS-M Greek dataset from the ICASSP 2023 signal processing challenge to verify cross-language accuracy;

[0009] S3. The specific method is:

[0010] S3-1. Data preprocessing

[0011] The goal of processing the text is to automatically and effectively segment and classify the continuous dialogue text, splitting the text content into a list of question sentences, a combined string of answer sentences, and a separate list of answer sentences;

[0012] First, the input dialogue text is segmented at the sentence level using regular expressions to ensure that each independent sentence is correctly recognized without being affected by specific contexts (such as abbreviations). Then, for each segmented text, a series of strategies based on grammatical rules and keyword matching are used to distinguish question sentences from answer sentences.

[0013] S3-2. Feature extraction

[0014] Inspired by the scale rules, we conducted a lexical feature analysis during an in-depth analysis of the "Cookie Theft" image description task in the ADReSS20 dataset. Based on these lexical analyses, we converted them into seven statistically significant Z-score indicators, which are key characteristic parameters for measuring the patient's performance in this task.

[0015] S3-2-1. Analysis of vocabulary characteristics

[0016] First, the original text is segmented, a vocabulary list is generated, and the total vocabulary is calculated; then, an exclusion vocabulary library is established to filter out words that distort the original meaning of the text or reduce the efficiency of expression;

[0017] When analyzing the relevance of content, we explored the vocabulary clusters closely related to the preset topic through the original relevant vocabulary list, and combined with regular expressions for further verification and statistical analysis, revealing the relevance of the text and the description content;

[0018] In the structural vocabulary analysis phase, a set of key syntactic structural units such as conjunctions and prepositions is preset, and regular matching rules are used to sort out text vocabulary and count the frequency of use of structural vocabulary;

[0019] In assessing the performance of the text in describing details and the richness of the content, part-of-speech tagging technology was adopted, with particular emphasis on the use of adjectives and adverbs; the number of modifying words was counted to quantitatively judge the participants' level of delicacy and thoroughness in describing details;

[0020] Finally, after completing the above-mentioned screening and classification, other words that fail to fall into the above-mentioned categories are identified as non-key content words;

[0021] S3-2-2. Z-score calculation

[0022] The general calculation formula for a Z-score (also called a standard score) is as follows:

[0023]

[0024] In this formula:

[0025] X represents a specific observation or data point;

[0026] It is the average of the entire group or sample data;

[0027] S is the standard deviation of the population or sample data, which measures the dispersion of the data points around the mean;

[0028] For each different language feature indicator x, treat it as an observation and subtract the mean value of the corresponding feature from this observation Then divide it by the standard deviation S of the feature to get the corresponding Z-score Z; these Z-scores are used as features extracted from the scale;

[0029] S3-3. Feature Analysis

[0030] First, SHAP is used to analyze the seven features, based on the Shapley value theory, by decomposing the model prediction results into the contribution of each feature;

[0031] S4. Experimental design

[0032] Traditional machine learning models including but not limited to K-nearest neighbor algorithm (K-NN) and the newer neural network KAN are used to predict the task; in KAN, the input is 7-dimensional, the output is 2-dimensional, and there is a hidden layer in the middle containing 16 hidden units; the number of network intervals and the order of the piecewise polynomial are both 3; LBFGS is used as the optimizer to minimize the loss function; the training process will execute 20 steps, and the parameters of the model will be updated at each step to minimize the loss function; the constructed 4-Layer FCNN model includes four interconnected fully connected layers, each layer is followed by a ReLU nonlinear activation function; the input layer of the network accepts 7 feature variables, and after layer-by-layer nonlinear transformation, the final output layer produces a 2-dimensional prediction result; a stochastic gradient descent optimizer with a learning rate of 0.001 is used, and a momentum term is introduced in the stochastic gradient descent optimizer.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention combines the case of using Z scores in medical research, uses the statistical Z score standard, bypasses the data-driven model learning process, and directly measures the degree of speech abnormality of the speaker. Z scores specifically include indicators such as speech richness and fluency. These indicators can truly and meticulously reflect the patient's cognitive status, and have no direct relationship with the language itself, so the present invention has natural cross-language generalization. The present invention first uses the whisper model to transcribe the speech data, and then quantitatively scores the pre-processed text data from seven dimensions such as the number of structural words and the number of irrelevant words, and finally calculates the corresponding Z value, and inputs it into the downstream model for detection. By introducing the Z score, the limitation of the lack of AD data sets is broken, the bottleneck of data-driven methods is broken, and the accuracy of cross-language AD detection is improved. At the same time, the present invention can not only be applied to AD screening in various language backgrounds, but also directly provide valuable references for clinicians in the medical field that requires transparency and explainability. DETAILED DESCRIPTION

[0035] The present invention is further described in detail below in conjunction with embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and provides a detailed implementation method and a specific operation process, but the protection scope of the present invention is not limited to the following examples.

[0036] In order to verify the effectiveness of the present invention, the present invention first tests and analyzes the ADReSS20 text data, and then further tests the ADReSSo English dataset and the ADReSS-M Greek dataset to prove the effectiveness of the method proposed in the present invention in a cross-language environment.

[0037] Example

[0038] A scale-based Alzheimer's disease detection method in a cross-language environment, the specific steps are:

[0039] S1. Build a four-layer fully connected neural network structure for data modeling and prediction;

[0040] S2. We first used the text dataset (ADReSS20) from the Interspeech 2020ADReSS challenge, which includes voice recordings and transcribed texts of the Cookie Theft picture descriptions from the Boston Diagnostic Aphasia Examination; the transcribed texts were annotated using the CHAT coding system; then we used the dataset (ADReSSo) designed specifically for detecting Alzheimer's disease through speech features from the 2021 ADReSSo challenge; finally, we used the ADReSS-M Greek dataset from the ICASSP 2023 signal processing challenge to verify cross-language accuracy.

[0041] Since the ADReSSo dataset and the ADReSS-M dataset are not in text form, Whisper has the function of translating multiple voices into English to convert these voice materials into English text form, so that the differences in language expression, grammatical structure, and vocabulary usage between AD and HC can be more intuitively compared and analyzed.

[0042] S3. The specific method is:

[0043] S3-1. Data preprocessing

[0044] The goal of text processing is to automatically and effectively segment and classify continuous dialogue text, splitting the text content into a list of question sentences, a merged answer sentence string, and a separate answer sentence list.

[0045] First, the input dialogue text is segmented at the sentence level using regular expressions to ensure that each independent sentence is correctly recognized without being affected by specific contexts (such as abbreviations). Then, for each segmented text, a series of strategies based on grammatical rules and keyword matching are used to distinguish between question sentences and answer sentences.

[0046] Specifically, when a sentence ends with a question mark, starts with a preset set of question words (including but not limited to "Who", "What", "When", "Where", "Why", "How", and supports case insensitivity), or contains potential inquiry structures such as "want youto" and "like you to", the sentence is judged as a question sentence and added to the question sentence list; otherwise, it is regarded as an answer sentence and added to the answer sentence list, and all answer sentences are connected into a coherent string with spaces as separators.

[0047] S3-2. Feature extraction

[0048] Inspired by the scale rules, lexical feature analysis was performed during the in-depth analysis of the "Cookie Theft" image description task in the ADReSS20 dataset. These analyses also provide clinicians with a set of standardized and systematic evaluation tools. Based on these lexical analyses, they were converted into seven statistically significant Z-score indicators, which are key characteristic parameters for measuring patients' performance in this task.

[0049] S3-2-1. Analysis of vocabulary characteristics

[0050] First, the original text is segmented, a vocabulary list is generated, and the total vocabulary is calculated. This step provides a basic reference for subsequent comprehensive analysis. Then, an exclusion vocabulary library is established to filter out words that distort the original meaning of the text or reduce the efficiency of expression.

[0051] At the level of language fluency, colloquial words such as "uh", "um" and "hmm" are identified and counted, and their use in the text is used as an important parameter for evaluating discourse coherence using regular expression search functions.

[0052] Because of the possible repetition of the content in the "Cookie Theft" images, the original vocabulary list was compared with the deduplicated version, and the repeated content words between the two were quantitatively analyzed. This helps to gain insight into the emphasis, thinking patterns, and possible thinking focus problems of patients when expressing. Simply changing this content table can be applied to other tasks and datasets.

[0053] When analyzing content relevance, we search for vocabulary clusters closely related to the preset topic through the original relevant vocabulary list, and further verify and statistically analyze them with regular expressions, revealing the degree of relevance between the text and the description content.

[0054] In the structured vocabulary analysis stage, the present invention presets a set of key syntactic structural units including conjunctions, prepositions, etc., and uses regular matching rules to sort out text vocabulary and count the frequency of use of structural vocabulary, thereby indirectly reflecting the internal structural organization of the text and the rigor of the logical hierarchy.

[0055] In evaluating the performance of the text in describing details and the richness of the content, the present invention adopts part-of-speech tagging technology, and pays special attention to the use of adjectives and adverbs. By counting the number of modifying words, the present invention can quantitatively judge the participants' delicacy and thoroughness in describing details.

[0056] Finally, after completing the above-mentioned screening and classification, the present invention identifies other words that are not included in the above-mentioned categories as non-key content words. Although these words do not directly carry core information or deepen the theme in the text, their distribution can still reflect the diversity and complexity of the text, and have important supplementary reference value for a comprehensive understanding of the language characteristics of the text.

[0057] S3-2-2. Z-score calculation

[0058] The general calculation formula for a Z-score (also called a standard score) is as follows:

[0059]

[0060] In this formula:

[0061] X represents a specific observation or data point.

[0062] It is the average of the entire population or sample data.

[0063] S is the standard deviation of a population or sample, which measures how dispersed the data points are around the mean.

[0064] For each different language feature index x, such as the total number of words, the number of repeated words, the number of structural words, etc., the present invention regards it as an observation value and subtracts the average value of the corresponding feature from the observation value Then divide it by the standard deviation S of the feature to obtain the corresponding Z-score Z. The present invention uses these Z-scores as features extracted from the scale.

[0065] S3-3. Feature Analysis

[0066] First, SHAP is used to analyze the seven features. This method is based on the Shapley value theory and improves the global and local interpretability of the model by decomposing the model prediction results into the contribution of each feature.

[0067] S4. Experimental design

[0068] In order to adapt to the classification task with less data, the present invention uses traditional machine learning models including but not limited to the K-nearest neighbor algorithm (K-NN) and the newer neural network KAN (Kolmogorov-Arnold Networks) to predict the task of the present invention. In KAN, the input of the present invention is 7-dimensional, the output is 2-dimensional, and there is a hidden layer in the middle containing 16 hidden units. The number of network intervals and the order of the piecewise polynomial of the present invention are both 3. Here, LBFGS is used as an optimizer to minimize the loss function. The training process will execute 20 steps, and the parameters of the model will be updated at each step to minimize the loss function.

[0069] The 4-Layer FCNN model constructed by the present invention includes four interconnected fully connected layers, each of which is followed by a ReLU (Rectified Linear Unit) nonlinear activation function, which is intended to capture complex nonlinear relationships in the data. The input layer of the network accepts 7 feature variables, and after layer-by-layer nonlinear transformation, the final output layer produces a 2-dimensional prediction result. The present invention uses a stochastic gradient descent optimizer with a learning rate of 0.001 to enhance the stability of the optimization process and accelerate the convergence rate. The present invention introduces a momentum term in the stochastic gradient descent optimizer, and its value is set to 0.9. The model of the present invention sets a batch size of 16 and trains a total of 300 epochs.

[0070] The method proposed in this paper outperforms some traditional machine learning and neural network models. At the same time, the method also shows good performance on the cross-language dataset ADReSS-M, proving its cross-language applicability.

[0071] In summary, the present invention innovatively proposes a method for detecting Alzheimer's disease based on scale rules in a cross-language environment. The method aims to reveal and make full use of the common characteristics of Alzheimer's disease detection in different languages ​​and cultural backgrounds. Compared with previous research results, the method of the present invention has also achieved good results on cross-language datasets and ensured a high degree of interpretability and rationality. The shortcoming of the paper is that the language differences caused by cultural regions have not been fully understood and grasped. In future research, it is necessary to further expand and improve it.

Claims

1. A scale-based Alzheimer's disease detection method in a cross-language environment, characterized in that: The specific steps are: S1. Build a four-layer fully connected neural network structure for data modeling and prediction; S2. We first used the text dataset (ADReSS20) from the Interspeech 2020 ADReSS challenge, which includes voice recordings and transcripts of the Cookie Theft picture descriptions from the Boston Diagnostic Aphasia Examination; the transcripts were annotated using the CHAT coding system; then we used the dataset (ADReSSo) designed for detecting Alzheimer's disease through speech features from the ADReSSo challenge; and finally, we used the ADReSS-M Greek dataset from the ICASSP 2023 signal processing challenge to verify cross-language accuracy; S3. The specific method is: S3-1. Data preprocessing The goal of processing the text is to automatically and effectively segment and classify the continuous dialogue text, splitting the text content into a list of question sentences, a combined string of answer sentences, and a separate list of answer sentences; First, the input dialogue text is segmented at the sentence level using regular expressions to ensure that each independent sentence is correctly recognized without being affected by specific contexts (such as abbreviations). Then, for each segmented text, a series of strategies based on grammatical rules and keyword matching are used to distinguish question sentences from answer sentences. S3-2. Feature extraction Inspired by the scale rules, we conducted a lexical feature analysis during an in-depth analysis of the "Cookie Theft" image description task in the ADReSS20 dataset. Based on these lexical analyses, we converted them into seven statistically significant Z-score indicators, which are key characteristic parameters for measuring the patient's performance in this task. S3-2-1. Analysis of vocabulary characteristics First, the original text is segmented, a vocabulary list is generated, and the total vocabulary is calculated; then, an exclusion vocabulary library is established to filter out words that distort the original meaning of the text or reduce the efficiency of expression; When analyzing the relevance of content, we explored the vocabulary clusters closely related to the preset topic through the original relevant vocabulary list, and combined with regular expressions for further verification and statistical analysis, revealing the relevance of the text and the description content; In the structural vocabulary analysis phase, a set of key syntactic structural units such as conjunctions and prepositions is preset, and regular matching rules are used to sort out text vocabulary and count the frequency of use of structural vocabulary; In assessing the performance of the text in describing details and the richness of the content, part-of-speech tagging technology was adopted, with particular emphasis on the use of adjectives and adverbs; the number of modifying words was counted to quantitatively judge the participants' level of delicacy and thoroughness in describing details; Finally, after completing the above-mentioned screening and classification, other words that fail to fall into the above-mentioned categories are identified as non-key content words; S3-2-2. Z-score calculation The general calculation formula for a Z-score (also called a standard score) is as follows: In this formula: X represents a specific observation or data point; It is the average of the entire group or sample data; S is the standard deviation of the population or sample data, which measures the dispersion of the data points around the mean; For each different language feature indicator x, treat it as an observation and subtract the mean value of the corresponding feature from this observation Then divide it by the standard deviation S of the feature to get the corresponding Z-score Z; these Z-scores are used as features extracted from the scale; S3-3. Feature Analysis First, SHAP is used to analyze the seven features, based on the Shapley value theory, by decomposing the model prediction results into the contribution of each feature; S4. Experimental design Traditional machine learning models including but not limited to K-nearest neighbor algorithm (K-NN) and the newer neural network KAN are used to predict the task; in KAN, the input is 7-dimensional, the output is 2-dimensional, and there is a hidden layer in the middle containing 16 hidden units; the number of network intervals and the order of the piecewise polynomial are both 3; LBFGS is used as the optimizer to minimize the loss function; the training process will execute 20 steps, and the parameters of the model will be updated at each step to minimize the loss function; the constructed 4-Layer FCNN model includes four interconnected fully connected layers, each layer is followed by a ReLU nonlinear activation function; the input layer of the network accepts 7 feature variables, and after layer-by-layer nonlinear transformation, the final output layer produces a 2-dimensional prediction result; a stochastic gradient descent optimizer with a learning rate of 0.001 is used, and a momentum term is introduced in the stochastic gradient descent optimizer.

Images