A computer-aided diagnosis system of parkinson’s disease based on hilbert spectrum features of speech

Abstract

Early detection and intervention are very important for the diagnosis and treatment of Parkinson’s disease. Firstly, according to the characteristics of dysarthria and abnormal sensitivity to the pronunciation of words such as ‘[pɑ˥’, ‘pjɑʊ˥’, ‘ pən˥’, and ‘pjɛn˥’ in patients with early Parkinson’s disease, a Chinese tongue twister “[pɑ˥ paɪ⇃ pjɑʊ˥ pɪŋ˥ pən˥ peɪ⇃ pʰo˥, pʰɑʊ˥ ˩pɪŋ˥ pɪŋ˥ pʰaɪ↿ peɪ⇃ pjɛn˥ pʰɑʊ⇃” was designed. The original recordings of Parkinson’s suspected patients were collected in a hospital geriatrics clinic to construct a tongue twister speech database. Secondly, Hilbert transform is introduced to process audio data to obtain the instantaneous amplitude signal of speech, the peak value, peak position, peak width and peak area of each syllable corresponding to the instantaneous amplitude and the total speech length, are extracted to construct the machine learning speech feature data set. Finally, SVM is introduced to train and evaluate the speech features of Parkinson’s disease suspected patients, its performance is compared with those of BP and LSTM models. The results show that the recognition accuracy of the three machine learning models is 92.10%, 88.94% and 90.18%, respectively, which indicates that the speech features extracted in this paper are effective. Combined with instantaneous amplitude features of speech and SVM model, a computer-aided diagnosis system for Parkinson’s disease is built. The clinical trial results show that the recognition accuracy of the system for Parkinson’s disease patients can reach 91.43%, and the system has fast response speed and strong anti-interference ability, which can fully assist doctors in making real-time and remote selections of patients with Parkinson’s disease.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-32543-4.

Introduction

Parkinson’s disease(PD) is a common, chronic and progressive neurodegenerative movement disorder, which is related to the aggregation of abnormal α-synuclein in Lewy bodies and the loss of dopaminergic neurons in the substantia nigrostriatum. The neurodegenerative mechanisms in PD are poorly understood and there is currently no cure for PD¹. In 2021, there were over 3 million Parkinson’s disease patients in China. With the rapid development of the aging population trend, it is estimated that the number of Parkinson’s disease patients will rise to 5 million by 2030, accounting for 50% of patients with Parkinson’s disease worldwide. The prevalence rate among people over 65 years old is as high as 1.7%, and the number of new patients each year is nearly 100,000². In recent years, Parkinson’s disease has become “younger”, with middle-aged and young patients accounting for 5% to 10% of the total number of patients³, this situation is not optimistic.

Parkinson’s disease symptoms are divided into motor symptoms and non-motor symptoms. Motor symptoms mainly include dysarthria, muscle rigidity, bradykinesia, resting tremor, postural balance disorders, etc., and non-motor symptoms mainly include sensory disorders, sleep disorders, mental disorders, autonomic nerve dysfunction, etc⁴. As the duration of the disease increases, these symptoms will gradually deepen until the patient completely loses self-control. Parkinson’s disease is characterized by being incurable, having a long course of disease, and a high disability rate, which brings huge psychological pressure and economic burden to patients and their families. Early intervention for Parkinson’s disease can effectively alleviate the malignant development of the disease, so early detection and diagnosis are crucial for patients with Parkinson’s disease. The “2022 Annual Report on the Diagnosis, Treatment and Quality of Life of Parkinson’s Disease Patients in China” shows that due to the uniqueness and complexity of Parkinson’s disease, more than 90% of Parkinson’s disease patients need more than one year to be diagnosed, which indicates that it is urgent to establish an automatic detection system of Parkinson’s disease.

Chinese tongue twisters are a unique linguistic form of the Chinese language. They consist of sentences with repeated, overlapping, and difficult-to-pronounce words that have similar initial consonants, vowels, or tones, and require rapid reading in one breath. Studies have shown that tongue twisters are highly sensitive to articulation disorders. Therefore, this study designed the Chinese tongue twister “[pɑ˥ paɪ⇃ pjɑʊ˥ pɪŋ˥ pən˥ peɪ⇃ pʰo˥, pʰɑʊ˥ ˩pɪŋ˥ pɪŋ˥ pʰaɪ↿ peɪ⇃ pjɛn˥ pʰɑʊ⇃” and collected the voices of suspected Parkinson’s disease patients in the geriatric outpatient department of a certain hospital. The aim was to explore valuable voice features and build a computer-aided diagnosis system for Parkinson’s disease by combining machine learning algorithms. Firstly, the collected tongue twister voices of suspected Parkinson’s disease patients were edited and organized in combination with clinical diagnosis to form a voice database. Secondly, the Hilbert transform was introduced to obtain the instantaneous amplitude signal of the voice, and the peak value, peak position, peak width, peak area corresponding to each syllable, and total voice length were extracted to form a voice feature database. Thirdly, three machine learning models SVM, BP and LSTM were introduced to train and evaluate these speech features, and their performances were compared with that based on amplitude features⁵ and the gray-scale distribution features of the complex value graph of the analytic signal, all of these results will provide a theoretical basis for the construction of the computer-aided diagnosis system. Finally, the SVM model based on the instantaneous amplitude feature was selected to build the computer-aided diagnosis system for Parkinson’s disease, and it has been verified through clinical trials that the the recognition accuracy of the system for Parkinson’s disease patients can reach 91.43%, so it is stable and reliable.

The main contributions of this study are as follows:

• A Chinese tongue twister speech database of suspected Parkinson’s disease patients is established. Unlike the continuous pronunciation of vowel letters, the features contained in this database are unique and effective, with strong anti-interference ability, it can provide a new research idea for automatic detection of Parkinson’s disease.

• We introduced the Hilbert transform to study the instantaneous amplitude signal of speech and extracted effective numerical features to describe the speech characteristics of Parkinson’s patients, and constructed a speech feature database for machine learning.

• We compared the performances of three machine learning models, namely SVM, BP and LSTM, on speech features of amplitude, instantaneous amplitude, and gray distribution law of complex value graph. Finally, the instantaneous amplitude features and SVM model were combined to build a computer-aided diagnosis system for Parkinson’s disease.

• The system has simple operation, high accuracy, and fast response time, and it can automatically activate re-training module in good time to improve recognition accuracy rate. Then the system can fully assist clinical doctors in real-time or remote screening of Parkinson’s disease patients.

Related works

Parkinson’s disease automatic detection methods are mainly divided into regression methods and machine learning methods. In 2015, Bayestehtashk et al.⁶ used openSMILE to extract speech features of subjects and compared the effectiveness of ridge regression, LASSO regression and support vector regression in helping doctors to infer the severity of Parkinson’s disease patients. In 2018, Camnos-Roca et al.⁷ extracted acoustic features from continuous vowel recordings and tested the ability of three regularization methods, LASSO regression, ENET regression and ridge regression, to distinguish healthy people from Parkinson’s disease patients. The recognition accuracy rates were 85.6%, 87.1% and 88.5%, respectively.

With the development of artificial intelligence technology, machine learning and deep learning have been widely used in the automatic detection of Parkinson’s disease. For now, Researchers have developed a large number of Parkinson’s disease automatic detection technologies based on machine learning methods, such as artificial neural network(ANN)⁸, random forest(RF)⁹, support vector machine(SVM)¹⁰, decision tree(DT)¹¹, multilayer perceptron(MLP) and k-nearest neighbor(KNN) network¹². In recent years, various deep learning technologies have begun to be applied in the field of automatic detection of Parkinson’s disease, such as the deep dual-side ensemble model based on deep feature learning and deep sample learning¹³, the end-to-end deep learning model based on convolutional neural network(CNN)¹⁴, and the multi- feature fusion model based on deep convolutional neural network(DCNN)¹⁵, which significantly improve the detection accuracy of Parkinson’s disease.

According to the different features extracted and trained by the model, machine learning methods of Parkinson’s disease are generally divided into the following four categories.

1) Machine learning based on articulatory features. Studies have shown that 89% of patients with early Parkinson’s disease have dysarthsia¹⁶, which is mainly manifested as single tone, low pitch, hoarse voice, low clarity, reduced loudness, reduced pressure, pitch fluctuation, shortness of breath, large pause, and inaccurate pronunciation¹⁷. On the one hand, patients with early Parkinson’s disease have insufficient vowel pronunciation and different vocalization durations^18,19, especially the vowel articulation index (VAI) is significantly lower than that of non-Parkinson patients of the same age²⁰. Therefore, exploring the unique pronunciation characteristics of vowel letters is the main method for intelligent detection of Parkinson’s disease. On the other hand, The characteristics of Parkinsonian speech was not only a stronger acceleration of articulation rate in the course of speaking but also a significant reduction of the total numbers of pauses, indicating an impaired speech rhythm and timing organization²¹, then quantitative acoustic analysis of phonological and prosodic disorders through text reading and monologue is also effective to detect Parkinson’s disease patients²², especially stress modification(including interrogative, imperative and declarative sentences) and prosodic language tasks(such as poetry) have higher accuracy in classifying and identifying speech disorders²³. Little et al.²⁴ performed feature selection and classification based on the repeated vowel recordings of the subjects, and the overall classification accuracy of Parkinson’s disease patients could reach 91.4%. Mekyska et al.²⁵ analyzed 36 new features of vowel pronunciation, and the recognition accuracy of Parkinson’s disease was 67.9% on the feature dataset of vowel letter ‘a/’. Cai et al.²⁶ established an intelligent diagnosis system for Parkinson’s disease based on the continuous pronunciation characteristics of the vowel letter ‘a/’, with an accuracy of 97.24%. Based on the continuous pronunciation characteristics of five vowel letters, Tripathi et al.²⁷ introduced empirical mode decomposition(EMD) and convolutional neural network(CNN) to identify patients with Parkinson’s disease, and the average accuracy was 73.76%. Shibina et al.²⁸ transformed voice inputs into spectrograms by short term Fourier transform(STFT), and then introduced the deep learning architectures, such as convolutional neural networks(CNN) and long short-term memory(LSTM) networks to identify Parkinson’s disease, the accuracy was 95.67%. Montaña et al.¹⁰ extracted the temporal and spectral features of the amplitude of consonant ‘k/’ and two plosive consonants ‘p/’ and ‘t/’ to identify patients with Parkinson’s disease. The Orozco-Arroyave et al.²⁹ evaluated the ability of four speech tasks: isolated words, rapidly repeated syllables ‘/pa/-/ta/-/ka/’, sentences, and spoken text to automatically distinguish patients with Parkinson’s disease in three languages: Spanish, German, Czech. Bocklet et al.³⁰ fused speech, pronunciation, prosody and language features to detect Parkinson’s disease patients, and the accuracy rate was 80%. Vasquez-Correa et al.³¹ developed the m-FDA scale based on multiple language tasks, which refines the scoring criteria of Parkinson’s disease patients into 13 items based on articulation, speech, prosody and comprehension.

2) Machine learning based on gait feature. Gait abnormality is one of the core motor symptoms in patients with Parkinson’s disease, which is mainly manifested as reduced stride length, drag walking, accelerated step frequency, frozen gait and postural instability, then gait features can be extracted to recognize the patients with Parkinson’s disease. Balaji et al.³² developed a motor gait classification system for early detection and stage classification of Parkinson’s disease using four machine learning models, including decision tree(DT), support vector machine(SVM), ensemble classification and Bayesian classification. Vidya et al.³³ introduced empirical mode decomposition(EMD), convolutional neural network (CNN) and long short-term memory (LSTM) network into gait signal learning for the diagnosis and classification prediction of Parkinson’s disease patients. Ferreira et al.³⁴ introduced the naive Bayes algorithm into spatio-temporal gait parameter learning to detect Parkinson’s disease patients, with an accuracy of 84.6%. Aydn et al.³⁵ developed a classification system of gait pattern for Parkinson’s disease patients by use of advanced ensemble learning algorithm and Hilbert-Huang transform, with an accuracy of 98.79%.

3) Machine learning based on brain image features. The study found that brain images of patients with Parkinson’s disease showed the loss of the normal “dovetail” on magnetic susceptibility sequences, which was related to the absence of nigrosome-1 in the substantia nigra. Recently, Hazhirkarzar et al.³⁶ developed a machine learning model based on baseline dopamine transporter single photon emission computed tomography(DAT-SPECT) imaging and multimodal magnetic resonance imaging(MRI) for predicting the progression of Parkinson’s disease, with an accuracy of 90%, a sensitivity of 1.00, and a specificity of 0.62. Bhakar et al.³⁷ established an automatic detection model of Parkinson’s disease based on brain SPECT image features by introducing artificial neural network, with an accuracy of 93%, a sensitivity of 1.00, and a specificity of 0.86.

4) Machine learning based on biochemical features. The researches found that the biochemical indicators of Parkinson’s patients may change, especially have a decrease in dopamine levels in the brain. In 2010, the Michael J. Fox Foundation and a core team of scientists and industry partners launched the Parkinson’s Progression Marker Initiative (PPMI) to create a robust open-access database; Based on the PPMI database, the research team combined multiple advanced machine learning techniques and multi-modal fusion strategies to develop an adaptive ensemble stacking model to improve the early recognition and diagnosis ability of Parkinson’s disease³⁸. In 2020, Wang et al.³⁹ developed an innovative deep learning technology to detect patients with early Parkinson’s disease by taking into account multiple indicators, such as rapid eye movement period, anosmia, cerebrospinal fluid data and dopaminergic imaging substances, with an accuracy of 96.45%. In 2025, the European Molecular Biology Laboratory (EMBL) and Quadram Institute in the UK integrated 22 independent research teams from 11 countries, and used a variety of mainstream machine learning algorithms to systematically evaluate the value and generalization ability of gut microbiota characteristics in the diagnosis of Parkinson’s disease⁴⁰.

The construction of feature database is also a hot topic in Parkinson’s disease computer-aided diagnosis system. In 2007, Little et al.²⁴ established the Oxford Parkinson’s disease dataset(OPDD). In 2010, the Parkinsons telemonitoring data set (PTDS) collecting audio signals through telephone interviews was formally formed⁴¹. In 2013, Sakar et al.⁴² established a Parkinson’s disease audio training database consisting of 26 speech samples, each of which contained linear and time-frequency features of vowels, numbers, words and short sentences. In 2024, Ma et al.⁵ constructed the Parkinson’s phonetic features on tongue twister(PPFTT) data set by extracting features from the speech amplitude signal of Chinese tongue twistel. In 2025, Matsumono et al.⁴³ extracted kinematic statistical features such as rhythm, amplitude, frequency, frequency standard deviation, slope, and established a feature databases for identifying Parkinson’s disease patients.

Methods

Data source

Four Chinese words such as ‘[pɑ˥’, ’pjɑʊ˥’, ‘pən˥’, and ‘pjɛn˥’ were selected from the “Mandarin Articulation Disorder Assessment Word List⁴⁴” jointly produced by Jinan University in China, the University of Central Arkansas in the United States, and Ohio State University to form the tongue twister “[pɑ˥ paɪ⇃ pjɑʊ˥ pɪŋ˥ pən˥ peɪ⇃ pʰo˥, pʰɑʊ⇃ pɪŋ˥ pɪŋ˥ pʰaɪ↿ peɪ⇃ pjɛn˥ pʰɑʊ⇃”. From April 3 to November 30, 2023, suspected Parkinson’s disease patients who could speak standard mandarin and were willing to cooperate with the experiment were selected in the geriatric outpatient department of a certain hospital. They recorded the tongue twister in a relatively quiet environment. After editing and sorting, a total of 97 audio recordings were collected to form a voice database⁵, among which 67 belonged to Parkinson’s disease patients (including 42 males and 25 females, with an average age of 72.5 years) and 30 belonged to non-Parkinson’s disease patients (including 18 males and 12 females, with an average age of 71.2 years). All voice data met the following requirements: (1) Completeness. The voice was basically complete, with more than 10 syllables clearly distinguishable, and the recording length did not exceed 15 s; (2) Clarity. The voice was clear and distinguishable, and environmental noise did not cause significant interference; (3) Validity. Other articulation disorder diseases of the patients, such as senile dementia, stuttering, and Alzheimer’s disease, were completely excluded.

This experiment was conducted in accordance with the relevant operation guidelines and regulations. Therefore, the participants were fully aware of the purpose of the experiment, and we confirm that the informed consent has been obtained from all subjects and/or their legal guardians.

Speech preprocessing and feature selection

Speech preprocessing

Signals are usually divided into stationary signals and non-stationary signals. The distribution law or distribution parameters of non-stationary signals will change with time, that is, the statistical characteristics of samples such as expectation and variance are different in different time periods. In contrast, the statistical properties of stationary signals do not change over time⁴⁵. The vast majority of signals in nature are non-stationary signals, and speech signals are also generally non-stationary signals.

The Fourier transform(FT) converts time-domain signals into frequency-domain signals, but it cannot capture the local frequency characteristics that change over time and is only suitable for processing stationary signals. To handle non-stationary signals, time-frequency analysis of the signal is required⁴⁶. The short-time Fourier transform(STFT) is a two-dimensional function concerning time and frequency. Through STFT, narrowband spectrograms with high frequency resolution and wideband spectrograms with strong time resolution can be obtained. However, the STFT has a fixed frame length globally and can only provide the same frequency and time resolution. To display high frequency resolution in the low-frequency region and time resolution in the high-frequency region, the wavelet transform(WT)⁴⁷ is needed. The WT processes the time-varying positive and negative frequency attenuation into finite-length wavelets, showing local frequency and time resolution characteristics. However, like STFT, it is limited by the Heisenberg uncertainty principle and cannot precisely distinguish between frequency and time. Moreover, it requires the manual setting of basic functions and can’t adaptively adjust. The Hilbert transform (HT) is an extension of the FT and can perform time-frequency analysis on signals⁴⁸. Signals are classified into single-frequency signals and multi-frequency signals based on their frequency characteristics. Single-frequency signals only emit and receive one band of frequencies. HT mainly assesses the instantaneous frequency of single-frequency signals and has no requirements for the positive or negative nature of the original signal, allowing it to handle signals with all non-negative frequencies.

The HT of the function x(t) is defined as

1

Here, t,τ denotes time, and PV denotes the principal value of the singular integral, which transforms the time real-valued function x(t) into the time complex conjugate function y(t). The original signal x(t) and its complex conjugate function y(t) form the analytic signal z(t) as follows:

2

The instantaneous amplitude signal of the analytical signal z(t) is obtained from Eq. (2) as follows:

3

Here, x(t) represents the real part of z(t), and y(t) represents the imaginary part of z(t). It should be noted that the instantaneous amplitude of the analytic signal is different from the amplitude of the original signal. The amplitude of the original signal can be positive or negative, while the instantaneous amplitude of the analytic signal is always positive. Based on the analytic signal, other features such as the instantaneous phase, instantaneous frequency, Hilbert spectrum and Hilbert marginal spectrum of the signal can also be obtained. Due to space limitation, this paper only considers the instantaneous amplitude feature of speech and compares it with the amplitude feature of the original signal⁵ and the gray-scale distribution feature of the complex value graph of the analytic signal. These features are trained and evaluated through three machine learning models: SVN, BP and LSTM. The research on other numerical features will be discussed in another paper.

Feature selection of speech signal

Syllable is the most basic speech unit, and each syllable corresponds to a Chinese character. The Chinese tongue twister “[pɑ˥ paɪ⇃ pjɑʊ˥ pɪŋ˥ pən˥ peɪ⇃ pʰo˥, pʰɑʊ˥ ˩pɪŋ˥ pɪŋ˥ pʰaɪ↿ peɪ⇃ pjɛn˥ pʰɑʊ⇃” consists of 14 Chinese characters, and the corresponding speech signal has 14 peaks. The peak value, peak position, peak width and peak area of the 14 peaks corresponding to the instantaneous amplitude signal of speech are selected in this paper. Together with the total length of the speech, a total of 57 features form the feature vector.

The amplitude signal of speech has obvious physical meaning, and the speech length describes the length of time for the collected tongue twisters to be recorded. The pronunciation of Parkinson’s disease patients is accompanied by hesitation and stuttering, which prolongs the speech length. The peak value of speech reflects whether the voice is loud and whether the voice is strong. The increased muscle tension of patients with Parkinson’s disease leads to unstable breath and low loudness when speaking, which is significantly different from that of non-Parkinson’s disease(non-PD) patients. The peak width of speech reflects the coherence and fluency of speech. The breath of Parkinson’s disease patients is not stable, and the characteristics of single tone, elongation, long interval between syllables, and vagueness of words are not uniform, and the peak width is generally higher than that of non-PD patients. The peak area describes the energy of syllables. Because the energy size of each syllable in PD patients is generally less balanced than that in non-PD patients, because the speech of PD patients is accompanied by plosive sounds and the breath is strong and weak. From the speech amplitude and instantaneous amplitude of PD patients and non-PD patients(Fig. 1), it can be found that in the amplitude and instantaneous amplitude of PD patients, the waveform of each syllable is thin, the line coherence is poor, the volume suddenly increases, plosive sounds or syllables do not appear, and the pause gap between syllables is long, and the fluency is weak. In the amplitude and instantaneous amplitude of non-PD patients, the syllables were full, the wave area was large, the sample contained syllables at a glance, the connections between syllables were compact, and the overall length was generally smaller than that of PD patients. In general, there are obvious differences in the balance of peak value, peak energy and peak width, and the distribution uniformity of peak positions, so it is meaningful to use them as numerical features to identify PD patients and non-PD patients.

Fig. 1.

Amplitude and instantaneous amplitude of speech signal on tongue twister.

In this study, first, we refer to reference⁵ to extract 6-dimensional features of speech amplitude as the first set of features, and then extract gray distribution features of complex- valued map as the second set of features, and extract 57-dimensional features of instantaneous amplitude as the third set of features, and then introduced SVM, BP and LSTM models to train and evaluate the effectiveness of the features. Finally, the instantaneous amplitude features and SVM machine learning model were used to construct the computer-aided diagnosis system of Parkinson’s disease.

In order to accurately extract the amplitude features and instantaneous amplitude features of speech signals, speech signals need to be compressed, denoised, reconstructed and segmented⁵. After the signals were segmented, the 14 largest peaks were selected as the speech signals corresponding to the 14 syllables. Figure 1 shows the instantaneous amplitude signal and amplitude signal of the Chinese tongue twisters for both PD patients and non-PD patients. Figure 2 illustrates the amplitude signal and instantaneous amplitude signal of speech for PD patient and non-PD patient in Fig. 1 after 10 times compression, as well as noise reduction and reconstruction, segmentation processing. In Figs. 1 and 2 and x(t) represents amplitude signal of speech, while a(t) represents instantaneous amplitude signal. From Fig. 2, it can be found that the 14 syllables of the speech signal are clear and recognizable, which provides great convenience for extracting the numerical features. Moreover, the peak width, peak value and peak energy of 14 syllables in PD patients were significantly smaller than those in non-PD patients, and the distribution of peak position was more uneven. In addition, the speech length of non-PD patients was also significantly lower than that of PD patients.

Fig. 2.

Preprocessing results for amplitude and instantaneous amplitude.

Feature extraction

HT is a very practical tool for signal analysis. Although it largely loses physical meaning of signal, but demonstrates powerful mathematical meaning, which is beneficial to describe many essential properties of signals.

Instantaneous amplitude features

After the instantaneous amplitude signal is normalized, the numerical features such as peak value, peak position, peak width and peak energy of each syllable, as well as the total speech length can be extracted.

Peak value refers to the maximum amplitude of a syllable, which describes the intensity of the pronunciation. The peak value of the ith syllable is represented by V_i, and the formula is

4

Here, L_i and U_i are the starting and ending positions of the ith syllable, and a(t) represents the instantaneous amplitude.

Peak position refers to the position of the peak of the syllable wave, and its distribution characterizes the fluency of the speech. P_i represents the position of the ith syllable wave peak, and the formula is

5

Peak energy refers to the wave area of the syllable, which describes the energy of the pronunciation. E_i represents the peak energy of ith syllable, and the formula is

6

Peak width refers to the duration of pronunciation of a syllable, which describes the fluency and stability of speech, and its formula is

7

The peak value V_i, peak position P_i, peak energy E_i, and peak width W_i of the instantaneous amplitude, as well as the total speech length L can form the feature vector T, that is.

8

Where N denotes the number of phonetic syllables. Because the tongue twister contains a total of 14 syllables, the instantaneous amplitude feature T is a 57-dimensional vector.

Amplitude features

The extraction of amplitude features can be referred to reference⁵. The 6-dimensional feature vector is composed of the length S_D of each speech, the average peak width M_PW, the average peak value M_PV, the average peak area M_PA, the average peak distance M_PD and the average peak span M_PS of 14 syllables, and is expressed as.

9

Grayscale distribution features of complex- value graphs

Figure 3 shows the gray-scale distribution law of analytical signals on PD patients and non-PD patient. It can be seen from Fig. 3 that the body of the complex-valued map of patients with PD is flat and loose, with a large range of spillover and diffusion, while the body of the complex-valued map of patients without PD is round and compact, with no obvious spillover and a small diffusion area. In this study, the gray-level distribution law of complex-valued maps were used to recognize patients with PD and non-PD. Firstly, the complex-valued image of the analytic signal is transformed into a binary image, and then it is evenly divided into M×M regions, and the ratio of black pixels in each region is counted, so that the gray distribution law of the complex-valued image can be described by an M ² dimensional feature vector, that is.

10

Fig. 3.

Grayscale distribution of complex-valued maps of speech signals.

Where s_i(1 ≤ i ≤ M ²) is the ratio of black pixels of ith region.

In order to find a region partition with good classification performance for every machine learning model, M = 3,4,5,6,7 are selected for testing and evaluation in this paper.

Model training and evaluation

In this study, there are 97 speech in the tongue twister speech database, of which 67 belong to Parkinson’s disease patients and 30 belong to non-Parkinson’s disease patients. Therefore, the input of amplitude feature is 97 × 6 matrix, the input of instantaneous amplitude feature is 97 × 57 matrix, and the input of gray distribution feature is 97×M ² matrix (M = 3,4, 5,6,7). The 97 × 1 vector is output as the class label, where the PD patient label is 0 and the non-PD patient label is 1. By five-fold cross validation, 80% of the data is taken as the training set, and the remaining 20% of the data is taken as the validation set. The parameter settings of BP, LSTM and SVM models are shown in Table 1, the average accuracy of 20 independent trials is taken to evaluate the effectiveness of features and the classification performance of the models. The classification accuracy of the three machine learning models based on the three feature is shown in Table 2. In Tables 1 and 2, AM represents amplitude, IAM represents instantaneous amplitude, CVI represents complex-valued image.

Table 1.

Parameters of the three machine learning models.

Models	Parameters	Values
BP	Number of network layers	4(input layer, hidden layer 1, hidden layer 2, output layer)
	Number of neurons in the hidden layer	20, 7(AM) 135, 57(IAM) 20, 7(CVI, M = 3) 35, 16(CVI, M = 4) 50, 25(CVI, M = 5) 80, 36(CVI, M = 6) 120, 49(CVI, M = 7)
	Initial learning rate	0.001
	Minimum learning batch	10
	Maximum iterations	350
LSTM	Number of network layers	5(input layer, 3 hidden layers, output layer)
	Number of neurons in the hidden layer	400, 400, 100
	Initial learning rate	0.001
	Mminimum learning batch	10
	Maximum iterations	350
SVM	Penalty parameter	1
	Kernel function	Radial basic functions
	Maximum iterations	350

Table 2.

Performance of machine learning models.

Models	Features	min/%	max/%	mean/%	sd/%
SVM	IAM	89.47	100	92.10	5.11
	AM	84.21	100	91.56	5.64
	CVI(M = 3)	63.16	68.42	67.37	2.10
	CVI(M = 4)	63.16	78.95	70.53	4.21
	CVI(M = 5)	68.42	78.95	74.21	4.38
	CVI(M = 6)	68.42	78.95	75.79	4.44
	CVI(M = 7)	68.42	84.21	75.26	6.68
BP	IAM	82.76	100	90.18	5.03
	AM	84.21	100	88.94	8.02
	CVI(M = 3)	68.42	94.74	80.00	7.37
	CVI(M = 4)	68.42	100	79.95	8.81
	CVI(M = 5)	68.42	94.74	79.95	8.80
	CVI(M = 6)	75.86	89.66	80.69	5.05
	CVI(M = 7)	68.42	94.74	79.47	6.84
LSTM	IAM	78.95	100	88.94	7.21
	AM	78.95	100	87.37	8.30
	CVI(M = 3)	57.89	73.68	65.26	6.17
	CVI(M = 4)	63.16	89.47	75.79	9.34
	CVI(M = 5)	68.42	73.68	70.00	7.00
	CVI(M = 6)	63.16	89.47	75.44	9.11
	CVI(M = 7)	68.42	78.95	73.16	4.60

Matlab 2018b was used for programming with an Intel (R) Core (TM)i7-4702MQ CPU@2.20 GHz CPU configuration.

It can be seen from Table 2 that the average recognition accuracy for Parkinson’s disease patients of SVM, BP and LSTM models based on instantaneous amplitude features is 92.1%, 90.18% and 88.94%, respectively, and SVM has the best performance. The average recognition accuracy for Parkinson’s disease patients of SVM, BP and LSTM based on amplitude features is 91.56%, 88.94% and 87.37%, respectively, and SVM had obvious advantages. The average recognition accuracy of the SVM model based on the gray distribution of the complex value image is between 67.37% and 75.79%, the BP model is between 79.95% and 80.69%, and the LSTM model is between 65.26% and 75.79%, BP has the best performance. The recognition performance of three machine learning models based on gray distribution is generally the best when M = 6, the accuracy of SVM, BP, and LSTM is 75.79%, 80.69%, and 75.79%. More than 75%, but the performance is still not satisfactory. Comprehensive consideration, the SVM model based on instantaneous amplitude features has the best recognition performance, the highest recognition accuracy is 100%, the lowest is 89.47%, and the average is 92.1%. The variance of 20 independent experiments is only 5.11%, which indicates that the SVM model has stable performance.

Figure 4 visually shows the accuracy of three machine learning models SVM, BP and LSTM corresponding to three sets of numerical features, where the gray and white dotted line represents the baseline with an accuracy of 80%, and the blue, purple and brown solid lines show the accuracy of BP, LSTM and SVM models respectively. The green, yellow, and red solid circles indicate that the corresponding model accuracy ranks first, second, and third. The green, yellow, and pink areas correspond to amplitude features, instantaneous amplitude features, and five groups of gray distribution features on complex value image for different dimensions, respectively. It can be seen from Fig. 4 that the SVM, BP and LSTM models based on instantaneous amplitude features have high recognition accuracy and strong stability, with the average accuracy of 92.10%, 90.18%, and 88.94% respectively, and the variances of 5.11%, 7.21% and 5.03% respectively. It shows that the SVM model based on instantaneous amplitude features can effectively identify patients with PD and non-PD.

Fig. 4.

Classification accuracy of machine learning models.

Computer aided diagnosis system

System architecture

With the rapid development of science and technology, the application of artificial intelligence is increasing day by day. More and more machine learning models are used in the medical field to assist doctors to make more accurate and objective diagnosis results based on clinical practice and patient symptoms. By comparing the recognition performance of SVM, BP and LSTM models based on amplitude features, instantaneous amplitude features and gray distribution features of complex-valued image, the SVM model based on instantaneous amplitude features is selected to construct a computer-aided diagnosis system for Parkinson’s disease. The system consists of four main modules:

1) Data preprocessing module. Firstly, the instantaneous amplitude signal of every original speech of suspected PD patients is obtained by Hilbert transform. Secondly, we perform normalization on the instantaneous amplitude. Finally, after further compression, noise reduction, reconstruction, and cutting operations, the 14 largest syllable signals of the normalized instantaneous amplitude signal are given.

2) Feature extraction module. Calculated the peak value, peak position, peak energy, and peak width of each syllable signal, together with the total length of the speech, it constitutes the 57-dimensional feature vector of the speech. The speech feature database is constructed by combining the patient’s labels given by the doctor’s clinical diagnosis.

3) Machine learning module. SVM is used to train the extracted instantaneous amplitude features, optimize the model structure and parameters, and obtain a satisfactory SVM classification model.

4) System application module. The tongue twister speech signal of the tester recorded in real-time was input into the diagnosis system. The system automatically extracts the instantaneous amplitude features and outputs the diagnosis results. The speech signal features of the tester and the computer diagnosis results were evaluated and confirmed by the clinicians and then imported into the speech feature database. When the scale of new speech data reaches the set threshold, the system automatically starts SVM re-training to improve the universality and accuracy of the model.

Figure 5 shows the architecture of the computer-aided diagnosis system for PD in this study.

Fig. 5.

Architecture of the computer-aided diagnosis system.

System configuration

In this study, data preprocessing, feature extraction, model training and system application are all carried out in the Matlab environment. The system is basically configured with Intel i7 processor or AMDx86-64 dual-core processor with 16GB memory, and the graphics card supports Open-GL3.3 and has 1GBGPU memory. Commonly used desktop computer or notebook computer can meet the system operation requirements.

System application

Open the system application interface in the Matlab environment, set the sampling frequency of 15 kHz and the time of 15s. In preparation time, testers first familiarize themselves with the tongue twister information and precautions, and then the recording was completed in a quiet environment, the system automatically completes the preprocessing and feature extraction of the recording data, and then pouring into the diagnosis system and output the diagnosis results. The specific process is: the command window output rules prompt “Please read the tongue twister within 15 s, [pɑ˥ paɪ⇃ pjɑʊ˥ pɪŋ˥ pən˥ peɪ⇃ pʰo˥, pʰɑʊ˥ ˩pɪŋ˥ pɪŋ˥ pʰaɪ↿ peɪ⇃ pjɛn˥ pʰɑʊ⇃”. Followed by the command, “Are you ready to start recording?” If you type “no”, the rule prompt is output again. If you type “yes”, the system will display “start recording” after 3 s, and voice prompt “end recording, thank you for your cooperation” after 15s. The recording is saved as a wav document by the system program ‘audiorecorder’, which is automatically imported into the diagnosis system after preprocessing and feature extraction to obtain the classification label. If the label is 0, the command window output prompts “You may be a patient with Parkinson’s disease, please go to the hospital for diagnosis as soon as possible!” If the label is 1, the output says, “You don’t have Parkinson’s disease!” After the diagnosis is completed, the system automatically imports the instantaneous amplitude features and diagnostic labels of tester’s speech into the speech feature database, and the system will automatically start the retraining process for each additional 20 data samples in the database in order to increase the accuracy and reliability of the diagnosis system.

The system was tested on suspected Parkinson’s disease patients in the outpatient clinic of a hospital from June 3 to September 30, 2024, and 32 Parkinson’s disease patients were identified from 35 suspected patients (21 males and 14 females) with an accuracy of 91.43%, and the average response time of the system was 3.5s. It shows that the computer-aided diagnosis system of Parkinson’s disease has rapid response and high accuracy, which can effectively assist doctors in real-time or remote screening of Parkinson’s disease patients, and clinicians can give targeted diagnosis and treatment opinions after comprehensive evaluation of the patient’s condition.

During the deployment of the computer-aided diagnosis system for Parkinson’s disease, we conducted follow-up visits with 20 patients and physicians who had used the system for auxiliary diagnosis purposes. We also collected and analyzed 10 user reviews regarding the application. All participants rated the system above 90 points across three key dimensions: safety, stability, and usability.

Conclusion

In this study, according to the characteristics of dysarthsia and abnormal sensitivity to tongue twisters in the early stage of Parkinson’s disease, the instantaneous amplitude features of speech signal based on Hilbert transform were used for SVM model, and a computer-aided diagnosis system of Parkinson’s disease was constructed. The trial test results show that the system can assist doctors in real-time, remote and accurate screening of patients with early Parkinson’s disease.

At present, the system still has the following shortcomings:

1) The data is unbalanced, the test results are somewhat dependent on language tasks, and the dysarthrias caused by other diseases cannot be completely excluded, such as mild stutter and mild Alzheimer’s disease.

2) The validity and accuracy of the model still lack of long-term test, the reasons for system diagnosis errors need further analysis, and the system needs to be optimized and improved in the later stage.

3) The sample size is small, and the information collection is susceptible to environmental noise, which affects the performance of the system.

4) A significant investment of time is needed to augment the model’s training data in order to enhance its generalization capability and mitigate the adverse effects of local dialects and other variables on diagnostic accuracy.

Future research can consider increasing the sample size and extracting more numerical features with physical meaning in the speech for classification learning. When the sample reaches a certain level, the severity of Parkinson’s disease patients can also be graded to improve the pertinency and effectiveness of clinicians’ diagnosis and treatment. In the later stage, it can be considered to make the system into a simple mobile phone APP, and the doctor or the relatives of patient can conduct the screening of Parkinson’s disease by themselves, so as to achieve early detection, early diagnosis and early treatment, all of this will be helpful to delay the process of the disease, reduce the patient’s physical pain, psychological pressure and economic burden.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Author Contributions Statement This paper was independently completed by our research team in terms of research design, data collection and data analysis. Ma Yuhong was responsible for writing the paper, Zheng Fangyuan and Wang Juntao were responsible for data collection and proofreading of the paper, Lu Jinye was responsible for data analysis and program design, and Ma Yuhong was responsible for reviewing the objectivity and authenticity of the paper and completing the submission. The authorship order was determined based on the actual contributions of each author and was approved by all authors. Finally, all authors contributed to the writing and revision of the paper, provided constructive suggestions and improvement suggestions for the writing of the paper, and ensured the authenticity, accuracy and objectivity of the research results. In addition, all members of the research team will receive consistent information on research participants and result reports. All authors of this paper jointly emphasize that the corresponding author is only responsible for the research results of this study. No single author bears full responsibility. The research institutions and cooperative institutions mentioned in this paper only provide research materials and research platforms, and do not include any other purposes and cannot be used as a basis for diagnosis and treatment.

Funding

This research was not funded by any institution or organization. The research funds were entirely self-funded by the team members.

Data availability

The audio dataset (speech recordings of the tongue twister) used in this study contains potentially sensitive information and is therefore not publicly available to protect participant privacy. However, the dataset is available for non-commercial academic research purposes upon reasonable request to the corresponding author, subject to approval from the institutional ethics committee and the execution of a data sharing agreement.

Declarations

Competing interests

The authors declare no competing interests.

Ethical statement

The study was approved by Changfeng Hospital Medical Ethics Committee, fully meeting the review requirements of the independent ethics committee.