ML-Based Quantitative Analysis of Linguistic and Speech Features Relevant in Predicting Alzheimer’s Disease

The rest of the paper is organized as follows: Section 2 presents a summary of prior investigations concerning the detection of Alzheimer’s disease and any relevant work on speech-based Alzheimer’s detection or classification. It also highlights the gaps in the literature that this study aims to address. Section 3 describes the data sources, image preprocessing steps, and machine learning techniques, provides a detailed discussion of the classification techniques, and Section 4 provides a clear interpretation of the results and their implications for Alzheimer’s disease detection. It may also highlight any limitations of the study, suggest future research directions, and summarize the study’s main findings and their significance. It may also offer potential clinical applications or future research directions.

2. Literature Review

Zargarbashi and Babaali (2019) proposed an approach to diagnosis systems for the early detection of Alzheimer’s disease using noninvasive data. The authors utilized a classification system based on spoken language and applied three techniques (statistical and neural) to categorize audio signals into two classes: dementia and control. To design a multi-modal feature embedding on the spoken language audio signal, they used N-gram, i-vector, and x-vector techniques. The system’s evaluation was performed on the cookie picture description task from DementiaBank’s Pitt Corpus, and it was found that the system’s accuracy was 83.6%. Therefore, this study implies that machine learning-based classification systems can help detect Alzheimer’s disease early by analyzing spoken language.

Kundaram and Pathak (2021) proposed a deep learning method that uses a DCNN to classify Alzheimer’s disease using MRI samples. The early diagnosis of Alzheimer’s disease is crucial for patient care and treatment. Deep learning techniques allow for learning high-level features from the dataset, resulting in improved performance compared to hand-crafted feature learning methods such as traditional machine learning techniques. The DCNN was trained to classify the disease as Alzheimer’s disease (AD), mild cognitive impairment (MCI), and standard control (NC) using the ADNI dataset. The results of the experiments showed an accuracy of 98.57%, which is higher than other studies. It should be noted that while this study presented promising results, further research is needed to validate the findings and to determine the generalizability of this model to real-world clinical settings. Additionally, the study may have limitations in the dataset and experimental setup, so further research is needed to evaluate the performance of this model in larger, more diverse datasets.

Tóth et al. (2018) discussed that the acoustic parameters of speech were analyzed in individuals with and without dementia using statistical analysis and machine learning techniques. The results indicated noteworthy differences in acoustic parameters such as speech tempo, articulation rate, silent pause, hesitation ratio, length of utterance, and the pause-per-utterance ratio between the two groups. The disparities were most prominent in the speech tempo during the delayed recall task and the number of pauses in the question-answering study.

Kumar et al. (2022) focused on identifying a concise set of speech features that could aid in recognizing dementia and implementing machine learning (ML) and deep learning (DL) models. The study utilized speech samples from the Pitt Corpus in Dementia Bank and proposed a combination of prosodic, voice quality, and cepstral features for the task. The experimental results demonstrate that the ML models outperform the DL models (87.6% vs. 85%) using the proposed speech feature combination, with lower time and memory consumption. These findings are promising and improve upon those of earlier studies on dementia recognition using speech.

Cummins et al. (2020) review Alzheimer’s dementia recognition through the spontaneous speech (ADReSS) challenge, which offers a platform for participants to create speech and language-based systems for recognizing Alzheimer’s dementia (AD) using speech recordings and transcripts. The study aims to evaluate various modern approaches to these modalities. The results indicate that linguistic techniques are more effective than acoustic systems. Combining BoAW, End-to-End CNN, and hierarchical-attention networks yielded the best test-set outcome, surpassing the challenge’s standards.

Pappagari et al. (2021) explore using speech and speaker recognition technologies and natural language processing to detect Alzheimer’s disease (AD) and estimate mini-mental status evaluation (MMSE) scores. The Interspeech 2021 ADReSSo challenge dataset was used to accomplish this. The research concentrates on adapting state-of-the-art speaker recognition and language models individually and in combination to evaluate their complementary performance for the tasks. The study’s most outstanding models yielded an 84.51% accuracy in the automatic detection of AD and 3.85 RMSE in MMSE prediction.

Pan et al. (2021) highlight the efficacy of utilizing acoustic and linguistic information embedded in spontaneous speech recordings for automatic Alzheimer’s disease detection. The paper delves into two cutting-edge ASR paradigms, Wav2vec2.0 and time delay neural networks (TDNN), on the ADReSSo dataset, which includes recordings of people describing the Cookie Theft (CT) picture. The study proposed five models to solely investigate the extraction of acoustic and linguistic information from audio recordings.

Liu et al. (2021) introduced a neural network-based technique for identifying Alzheimer’s disease (AD) in the spontaneous speech of subjects during a picture description task without depending on manual transcriptions and annotations of the lesson. The approach extracts bottleneck features from audio via an ASR model. It utilizes convolutional neural network (CNN) layers for local context modeling, bidirectional long short-term memory (BiLSTM) layers for global context modeling, and an attention-pooling layer for classification.

Mittal et al. (2020) suggested that detecting the early stages of Alzheimer’s disease (AD) is challenging because of the absence of a definitive in vivo diagnosis. The authors propose a multi-modal deep learning approach using speech and corresponding transcripts to identify AD. A CNN-based model predicts the diagnosis for various speech segments combined for the final prediction. When trained and evaluated on the DementiaBank’s Pitt Corpus, the proposed method achieves 85.3% accuracy in 10-fold cross-validation.

Meghanani et al. (2021) explored the efficacy of Log-Mel spectrograms and MFCC features in Alzheimer’s disease (AD) recognition and mini-mental state examination (MMSE) score prediction using the ADReSS challenge dataset. Three deep neural networks (DNNs) - CNN-LSTM, ResNet-LSTM, and pBLSTM-CNN - were used. CNN-LSTM showed an accuracy of 64.58% using MFCC features, while ResNet-LSTM achieved an accuracy of 62.5% with Log-Mel spectrograms. The proposed approach achieved an accuracy of 62.5% on the ADReSS dataset.

Below Table 1 shows the comparative study of classification approaches using machine learning.

3. Materials and Methods

This paper uses a set of standardized neuropsychological tests to confirm each participant’s cognitive impairment diagnosis. A speech analysis cross-platform software tool capable of extracting various features from audio recordings, such as pitch, intensity, and duration, is used in this paper; we use the CLAN platform. Once the extracted features were preprocessed to remove noise and normalize the data. Feature selection was then performed to identify the most relevant features for the classification task. Here is a proposed framework for Alzheimer’s disease detection and classification. Figure 1 shows the complete flowchart for dementia detection using machine learning (ML) techniques.

3.1. Data Acquisition

In this paper, we collected data from DementiaBank’s Pitt Corpus dataset, which consists of conversations between individuals with various forms of dementia and their interlocutors. Much work has already been done with this dataset, which is discussed in (Chen et al., 2019). The dataset includes codes for transcribing speech and annotating various aspects of language use, such as turn-taking, discourse structure, and speech errors. Overall, the DementiaBank’s Pitt Corpus dataset is a valuable resource for research on neurodegenerative diseases, particularly Alzheimer’s disease, and has been utilized extensively in the field for developing and evaluating machine learning models for diagnosing and predicting these conditions. Initially, our dataset contained 1200 entries and over 100 features, saved in a CSV file. To ensure that the data was fit for training models, we cleaned it extensively, including removing redundant columns and handling missing values and errors. Additionally, you standardized the data to prevent any feature from disproportionately influencing the model. As a result of this preprocessing, we identified 50 relevant features used to train our models.

3.2. Preprocessing

The acquired data needs to be preprocessed to remove any noise, it must be correct for processing, and normalized across patients. This step is essential to ensure that machine learning algorithms can accurately detect and classify Alzheimer’s disease. After collecting the data, we processed it to ensure it was in the appropriate format, making it easier to proceed with our tasks. To achieve high accuracy in speech recognition, high-quality acoustic segments, which are portions of the speech free from background noise or interference, must be used. This can be done using advanced signal processing techniques such as noise reduction, filtering, and speech enhancement (Pan et al., 2020). Figure 2 shows an example of automatic transcript selection.

3.3. Feature selection

Feature extraction is the process of selecting and transforming relevant information from raw data to create a set of features that can be used for analysis, modeling, or machine learning tasks. In this paper, we collected data from the Dementia Bank’s Pitt Corpus. Initially, the dataset contained 100+ features, but after preprocessing, we extracted 50 features that were further processed for execution. Figure 3 shows the correlation between the attributes.

3.4. Classification

The study employed five commonly used machine learning algorithms: k-nearest neighbours, support vector machine, decision tree, random forest, and XGBoost.

3.4.1. Support Vector Machine

In the context of Alzheimer’s disease detection, SVM can classify brain MRI images as either Alzheimer’s disease or healthy controls. The SVM algorithm takes as input a set of labeled brain MRI images, where each image is labeled as either Alzheimer’s disease or healthy control. The algorithm then creates a hyperplane that separates the two classes, maximizing the margin between them.

We must find the optimal values of w and b that maximize the margin between the two classes to train the SVM model. The margin is the distance between the hyperplane and the closest data points from each category. The optimal hyperplane is the one that maximizes the margin. For a given dataset with N data points, the input data is represented as a set of feature vectors xi, each with a corresponding binary class label yi, where yi ∈ {−1, +1}. SVM aims to find a hyperplane that can separate the two classes.

The hyperplane can be represented in Eq (1)

$$\mathrm{w}·\mathrm{x}+\mathrm{b}=0\mathrm{ (1)}$$

Where w is a weight vector perpendicular to the hyperplane, and b is a bias term. This hyperplane defines the decision boundary.

3.4.2. Decision Tree

In the context of Alzheimer’s disease detection, the decision tree can classify brain MRI images as either Alzheimer’s disease or healthy controls. The decision tree algorithm takes as input a set of labeled brain MRI images, where each image is labeled as either Alzheimer’s disease or healthy control. The algorithm then creates a tree-like model that separates the two classes. The decision tree algorithm starts with the root node and recursively splits the data based on the values of the attributes until the leaves are pure or the stopping criteria are met. The goal is to create a tree that generalizes well to new data.

To use a decision tree for Alzheimer’s disease detection, brain MRI images can be preprocessed to extract relevant features, such as shape, texture, and intensity features. These features can then be used as input to the decision tree algorithm, which can be trained to classify the images as either Alzheimer’s disease or healthy controls.

3.4.3. Random Forest

Random forest is an ensemble learning algorithm widely used to improve the accuracy and stability of predictions. It is a supervised learning algorithm that can be applied to solve classification and regression problems. The key concept behind random forest is to create several decision trees using random subsets of training data and features. Each tree in the forest is trained independently, and the final prediction is obtained by combining the outputs of all the trees. This approach helps reduce the variance and improve the model’s generalization capability. Here is an example formula for random forest alzheimer classification:

For each tree in the forest: