Parkinson’s disease (PD) is a neurological degenerative disease characterized by a group of motor symptoms, including bradykinesia, rigidity, postural instability, and resting tremors, as well as a group of non-motor symptoms, including hyposmia, cognitive impairment, and autonomic dysfunctions (Solla et al., 2020; Ercoli et al., 2022), which occur due to progressive loss of dopaminergic neurons in the substantia nigra of the midbrain (Hornykiewicz, 1998). Both motor and non-motor symptoms have significant impact on the quality of life of patients with PD (Solla et al., 2020; Ercoli et al., 2022). Although PD remains incurable, accurate diagnosis and subsequent treatment of mild conditions can improve the patient’s quality of life (Lang et al., 2013). However, accurate diagnosis of early-stage PD is difficult because mild motor symptoms are often misconstrued as typical signs of aging. A survey reported a clinical diagnostic accuracy of only 80% by experienced movement disorder specialists, whereas the precision is even lower when performed by general neurologists (Adler et al., 2014). In recent years, several tools have been used for the early diagnosis of PD, including dopamine transporter (DaT) scans, 7-Tesla magnetic resonance imaging, and radioactive iodine metaiodobenzylguanidine scans, with satisfactory accuracy (Brooks, 1993; Gaenslen et al., 2008; Cosottini et al., 2014; Kawazoe et al., 2019; Vaswani et al., 2022). However, these examinations are time-consuming, costly, and rely on specialized equipment and reagents that are only feasible in a limited number of top-tier tertiary medical centers, thus restricting their widespread application. Hence, a reliable, inexpensive, and convenient tool that can accurately identify early-stage PD is required.

Dysarthria, a prevalent motor manifestation of PD, is observed in approximately 90% of patients. Notably, it may manifest up to 5 years before the occurrence of motor symptoms, highlighting its potential as an early indicator of PD (Postuma et al., 2012). Dysarthria in PD is characterized by hypokinetic symptoms, including reduced voice volume and pitch variation, breathy voice, tremors, hoarse voice quality, inconsistent voice rates, and imprecise articulation. Hypokinetic dysarthria reflects the involvement of all dimensions of voice production, including respiration, phonation, resonance, articulation, and prosody (Moro et al., 2021). Objective acoustic analysis of voice in individuals with PD can reveal abnormalities in specific features, including decreased signal-to-noise ratios and increased jitter and shimmer (Suphinnapong et al., 2021). Moreover, the advancement of data processing technology has led to the utilization of machine learning algorithms for dynamic analysis of voice feature parameters, aiming to achieve accurate classification between patients with PD and healthy controls (HCs). Many studies utilized linear voice feature parameters such as jitter, shimmer, or noise, in conjunction with various machine learning algorithms that typically result in an accuracy exceeding 0.75 when distinguishing individuals with PD from HCs (Khojasteh et al., 2018; Rusz et al., 2018; Moro-Velazquez et al., 2019). Little et al., utilized nonlinear voice feature parameters, including Recurrence Period Density Entropy (RPDE), Detrended Fluctuation Analysis (DFA), and Pitch Period Entropy (PPE), in combination with the SVM algorithm to achieve an accuracy of 0.91 in differentiating individuals with PD from HCs (Little et al., 2007; Orozco-Arroyave et al., 2016). Additionally, other types of feature parameters such as Mel-Frequency Cepstral Coefficients (MFCC), and Band Bark Energies (BBE) have been employed alongside several machine learning algorithms to provide accuracies exceeding 0.85 (Sakar et al., 2019; Vasquez-Correa et al., 2020). Recently, several studies have utilized representation learning feature parameters in combination with machine learning algorithms to discriminate individuals with PD from HCs (Zhang, 2017; Cummins et al., 2018). These feature parameters were extracted using deep learning methods and contained additional hidden voice information that is difficult to interpret compared to conventional voice features. Moreover, these features could help to improve the accuracy of different models to classify pathological voice (Cummins et al., 2018). The aforementioned studies have successfully used machine learning algorithms to analyze conventional voice feature parameters and representation learning feature parameters to accurately distinguish individuals with PD from HCs. However, most studies have focused on discriminating between individuals with PD and HCs without considering the differentiation between those with mild/early PD and HCs. Even studies that have specifically focused on diagnosing mild/early PD typically use modified Hoehn–Yahr staging (mH&Y) criteria of ≤3 or mH&Y ≤ 2 to define mild/early PD. These studies also had limited datasets, usually encompassing <50 patients with mild/early PD (Rusz et al., 2011; Defazio et al., 2016; Lim et al., 2022; Wang et al., 2022). Indeed, in clinical practice, patients with mH&Y ≥ 2 are easier to identify, whereas those presenting with mH&Y ≤ 1.5 often exhibit mild motor symptoms that frequently lead to misdiagnosis. Therefore, it is imperative to use voice to differentiate individuals with “real” mild/early PD (mH&Y ≤ 1.5) from HCs. Otherwise, since current research on using voice to distinguish between PD and HCs mainly focuses on English-speaking populations, there is a relative lack of studies on native Chinese speakers. Given this gap, the present study has employed a Chinese database for both model construction and validation, aiming to strengthen the evidence base in this field.

This study aimed to establish machine learning models for analyzing conventional and representation learning feature parameters extracted from sustained vowels, which could be used to distinguish patients with PD and mild PD (mH&Y ≤ 1.5) from age- and sex-matched HCs. We also compared the diagnostic performance of the models with that of general neurologists, who are not experts in movement disorders.

In this study, we used a machine-learning algorithm to analyze conventional and representation learning feature parameters extracted from sustained vowel and developed diagnostic models to differentiate between HCs, patients with PD, and patients with mild PD. The models were independently validated using separate datasets. Subsequently, the diagnostic performance of the model was compared with that of general neurologists. Our results demonstrated that compared with general neurologists, the diagnostic models showed satisfactory performance in distinguishing between patients with PD and HCs, as well as patients with mild PD and HCs.

Our study revealed that within the best feature parameter subset for developing the mild PD model, the vowel [o] exhibited a significantly higher number of feature parameters ([a]: 25 parameters, [o]: 39 parameters, [i]: 17 parameters). This finding suggests that the vowel [o] contained more information to distinguish mild PD and HCs. In the research on using sustained vowel tasks to differentiate patients with PD from HCs, the vowels [a], [o], and [i] are commonly employed (Moro et al., 2021). However, there is still controversy regarding which vowel provides more information for distinguishing PD patients from HCs. Hireš et al. found that, when compared to other vowels, the vowel [a] demonstrated the highest diagnostic performance in distinguishing between patients with PD and HCs, achieving an accuracy of 0.99, a sensitivity of 0.86, a specificity of 0.93, and an AUROC of 0.89 (Hireš et al., 2022). However, Orozco-Arroyavede et al. found that when distinguishing between patients with PD and HCs, the vowel [i] achieved the highest accuracy of 0.76 among the 5 vowels ([a], [e], [i], [o], [u]) (Orozco-Arroyave et al., 2013). While, Song et al. revealed that when distinguishing between patients with mild PD (mH&Y ≤ 3) and HCs, the vowel [o] achieved the highest accuracy of 0.85, compared to vowels [a] and [i] (Song et al., 2020). The possible causes for why the vowel [o] contains more information to distinguish mild PD from HCs, we speculated as follows: The position of the tongue within the oral cavity varies when phonating different vowels (Zhenni et al., 2020). When producing the vowel [a], the tongue is positioned in the middle horizontally and at its lowest point vertically within the oral cavity (Zhenni et al., 2020). This indicates that the tongue is in a relaxed state during vowel [a] production. While producing vowels [o] and [i], the tongue must remain in specific positions within oral cavity. When pronouncing the vowel [i], the tongue is positioned close to the palate, where there is relatively limited space in the oral cavity, making it easier to maintain a stable tongue position (Zhenni et al., 2020). However, when pronouncing the vowel [o], the position of the tongue is close to the back of the oral cavity, and it is centered in the vertical direction (Zhenni et al., 2020). This makes it more difficult for the tongue to maintain a stable position. In patients with mild PD, the abnormal movement of the tongue due to bradykinesia and rigidity makes it challenging to maintain a stable position (Mefferd and Dietrich, 2019). Therefore, it is more difficult for patients with mild PD to produce sustained and stable vowel [o] compared to vowels [a] and [i].

Our study also found that the representation learning feature parameters accounted for the majority of the best feature subsets used to construct the two diagnostic models, indicating that these parameters provided more information for distinguishing patients with PD from HCs than conventional voice feature parameters. The phenomenon of voice is inherently complex, characterized by high-dimensional data, and the effects of PD on voice are also multidimensional. Conventional voice feature parameters may not adequately capture enough information to characterize the voice signals associated with PD (Vasquez-Correa et al., 2020). The application of deep learning methods to automatically extract abstract and unexplained hidden features from voice and distinguish PD patients from HCs has been attracting increasing attention. Correa et al., utilized RAE and CAE to extract representation learning feature parameters from voice data, which were then classified using the SVM algorithm. The accuracy achieved for discriminating patients with PD and HCs was 0.84 (Vasquez-Correa et al., 2020). Zhang et al., using stacked autoencoders (SAE) to extract representation learning feature parameters from voice data, which were then classified using the KNN algorithm. The accuracy achieved for discriminating patients with PD and HCs was 0.97 (Zhang, 2017). Subsequently, they compared the diagnostic performance of representation learning feature parameters with that of conventional feature parameters and found that the representation learning feature parameters had a higher classification accuracy than conventional feature parameters.

Currently, four primary tasks are utilized for detecting dysarthria in patients with PD: sustained vowel, syllable repetition, passage reading, and monologue tasks (Rusz et al., 2021). In comparison to other tasks, the sustained vowel task remains unaffected by cognition, language, and dialect (Gerratt et al., 2016). Additionally, the sustained vowel task can be effortlessly executed and adheres to a consistently standardized methodology. However, several studies have pointed out that compared to sustained vowel tasks, syllable repetition, passage reading, and monologue tasks can provide more comprehensive voice information and be more accurate in distinguishing between patients with PD and HCs (Rusz et al., 2013; Godino-Llorente et al., 2017; Wang et al., 2022). While, our study revealed that even sustained vowel task could distinguish mild PD from HC with accuracy ≥0.85. The possible reasons for our high accuracy are as follows: In addition to conventional feature parameters, our study utilized representation learning feature parameters that encompass abstract and unexplained hidden features (not present in conventional features), which are helpful in distinguishing patients with PD from HCs (Vasquez-Correa et al., 2020).

Recent report has indicated that PD affects 3.62 million patients in China, accounting for half of the global number of patients with PD (Qi et al., 2021). With a rapidly aging population, this number is predicted to increase to approximately 5 million by 2030 (Li et al., 2019). However, a significant proportion (an estimated 20–40%) of individuals with PD remain undiagnosed, (MRC CFAS et al., 2006; Bajaj et al., 2010) with even lower figures in rural areas (Zhang et al., 2003). In the early stages of PD, mild motor symptoms may be perceived as an age-related decline in motor function and are easily overlooked by patients and general neurologists, leading to misdiagnosis. Even movement disorder specialists have difficulty distinguishing patients with PD from HCs until the average mH&Y stage reaches 1.8 (Hughes et al., 2002). Furthermore, diagnoses made by general neurologists are likely to have lower diagnostic accuracy and later mH&Y staging (Joutsa et al., 2014). In the present study, the sensitivity of general neurologists to discriminate between patients with PD and HCs was 0.85 and 0.60 for mild PD and HCs, respectively. These results implied that 15% of patients with PD and 40% of those with mild PD are misdiagnosed by general neurologists, which is consistent with the misdiagnosis rates reported previously. Compared with the diagnostic performance of general neurologists, the machine learning-based diagnostic model analyzing voice feature parameters in the present study demonstrated outstanding performance. When using voice feature parameters to differentiate between patients with PD and HCs, most studies exhibited impressive AUROC and accuracies (>0.9) (Little et al., 2007; Orozco-Arroyave et al., 2016; Khojasteh et al., 2018; Rusz et al., 2018; Moro-Velazquez et al., 2019). Studies that focused on discriminating between patients with mild early PD and HCs exhibited satisfactory diagnostic performance (AUROC and accuracy >0.8) (Zhang, 2017; Cummins et al., 2018; Sakar et al., 2019; Vasquez-Correa et al., 2020). However, these studies defined mild/early PD as mH&Y ≤ 3 or mH&Y ≤ 2, and most of the participants with mild/early PD enrolled in these studies had a mH&Y stage of ≥1.5. To the best of our knowledge, the current study is the first to attempt to distinguish PD with mH&Y ≤ 1.5 from HCs using voice feature parameters. The results demonstrated the remarkable diagnostic performance of the model in identifying patients with mild PD and HCs, with an AUROC of 0.93 and an accuracy of 0.85. The sensitivity of the model in identifying patients with mild PD and HCs reached 0.95. Thus, the model can distinguish most patients with mild PD from the general population, making it suitable for screening.

Although PD remains incurable, accurate diagnosis and subsequent treatment can improve the patient’s quality of life However, in real-world settings, the diagnosis and treatment of PD in its early stages are challenging. Owing to the lack of early screening tools and accurate diagnostic support, it is difficult to diagnose most patients earlier than mH&Y stage 2 (Joutsa et al., 2014), thereby missing the best time window for disease-modifying treatment. Until now, the precise diagnosis of early-stage PD relied on a limited number of movement disorder specialists and rare, expensive equipment. Additionally, few patients are aware of their symptoms and consult doctors before mH&Y staging 1.5. Consequently, the development of satisfactorily sensitive and convenient non-invasive screening tools stage is urgently needed to detect prodromal or mild PD. Our study results showed that analyzing voice feature parameters extracted from a simple sustained vowel task, which can be performed using only a smartphone with a microphone for recording and a computer for analyzing, is a convenient and relatively affordable tool for identifying PD before mH&Y stage 1.5.

Our study has some limitations. First, we did not analyze the correlation between voice parameters and disease severity (mH&Y, and MDS-UPDRS III). Dysarthria is just one manifestation of the motor symptoms of PD, and although it may appear earlier than other motor symptoms, relying solely on voice parameters to assess PD severity may not fully reflect the actual severity. Additionally, we only included patients with PD and HCs and excluded patients with other types of parkinsonism (e.g., multiple system atrophy and progressive supranuclear palsy), which would make the model more widely applicable. Currently, diagnostic criteria for prodromal PD have been proposed, and idiopathic rapid eye movement sleep behavior disorder is a prodromal symptom of PD that has garnered significant attention. The use of voice and other multimodal somatosensory parameters to screen for these diseases are future research directions.

In this study, we used a machine learning algorithm to analyze voice feature parameters and developed diagnostic models for differentiating between HCs, patients with PD, and those with mild PD (mH&Y ≤ 1.5). The models were independently validated using separate datasets. Our results demonstrate a remarkable diagnostic performance of the model in identifying patients with mild PD (mH&Y ≤ 1.5) and HCs. Furthermore, we proposed a paradigm for the automatic identification of patients with PD by voice. The results of our study are helpful for screening PD in the early stages in the community and primary medical institutions where movement disorder specialists and special equipment are lacking.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by Ethics committee of Chinese PLA General Hospital. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

MW: Data curation, Formal analysis, Investigation, Methodology, Resources, Validation, Visualization, Writing – original draft, Writing – review & editing. XlZ: Writing – original draft, Writing – review & editing. FL: Writing – original draft, Writing – review & editing. LW: Formal analysis, Methodology, Software, Validation, Writing – original draft, Writing – review & editing. YiL: Data curation, Resources, Writing – review & editing. RT: Investigation, Writing – review & editing. JY: Investigation, Writing – review & editing. SL: Data curation, Formal analysis, Methodology, Software, Writing – review & editing. YZ: Data curation, Formal analysis, Methodology, Software, Writing – review & editing. YuL: Data curation, Formal analysis, Methodology, Software, Writing – review & editing. KR: Data curation, Formal analysis, Methodology, Software, Writing – review & editing. ZC: Data curation, Formal analysis, Methodology, Software, Writing – review & editing. XY: Funding acquisition, Investigation, Resources, Validation, Visualization, Writing – review & editing. ZW: Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Validation, Visualization, Writing – review & editing. ZG: Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – review & editing. XZ: Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Validation, Visualization, Writing – review & editing.