Voice as a Biomarker of Pediatric Health: A Scoping Review

Abstract

The human voice has the potential to serve as a valuable biomarker for the early detection, diagnosis, and monitoring of pediatric conditions. This scoping review synthesizes the current knowledge on the application of artificial intelligence (AI) in analyzing pediatric voice as a biomarker for health. The included studies featured voice recordings from pediatric populations aged 0–17 years, utilized feature extraction methods, and analyzed pathological biomarkers using AI models. Data from 62 studies were extracted, encompassing study and participant characteristics, recording sources, feature extraction methods, and AI models. Data from 39 models across 35 studies were evaluated for accuracy, sensitivity, and specificity. The review showed a global representation of pediatric voice studies, with a focus on developmental, respiratory, speech, and language conditions. The most frequently studied conditions were autism spectrum disorder, intellectual disabilities, asphyxia, and asthma. Mel-Frequency Cepstral Coefficients were the most utilized feature extraction method, while Support Vector Machines were the predominant AI model. The analysis of pediatric voice using AI demonstrates promise as a non-invasive, cost-effective biomarker for a broad spectrum of pediatric conditions. Further research is necessary to standardize the feature extraction methods and AI models utilized for the evaluation of pediatric voice as a biomarker for health. Standardization has significant potential to enhance the accuracy and applicability of these tools in clinical settings across a variety of conditions and voice recording types. Further development of this field has enormous potential for the creation of innovative diagnostic tools and interventions for pediatric populations globally.

1. Introduction

The human voice is often referred to as a unique print for each individual. It contains biomarkers that have been linked in the adult literature to various diseases ranging from Parkinson’s disease [1] to dementia, mood disorders, and cancers, such as laryngeal and glottal [2,3,4]. The voice contains complex acoustic markers that depend on respiration, phonation, articulation, and prosody coordination. Recent advances in acoustic analysis technology, especially when coupled with machine learning, have shed new insights into the detection of diseases. As a biomarker, the voice is cost-effective, easy, and safe to collect in low-resource settings. Moreover, the human voice contains not only speech, but other acoustic biomarkers such as cry, cough, and other respiratory sounds.

The existing body of systematic literature on vocal biomarkers for disease detection spans multiple disciplines, including acoustic analysis, machine learning, and application to clinical medicine. Key studies on this topic include a review by Fagherazzi et al. (2023) [5]. Fagherazzi et al. discuss the integration of vocal biomarkers into clinical practice and highlight the potential of non-invasive voice analysis for diagnosing conditions such as Parkinson’s disease, depression, and cardiovascular conditions. Sara et al. (2023) [6] discuss the feasibility of remote health monitoring using voice analysis and demonstrate significant advancements in telehealth, leveraging technology, and machine learning for the detection of conditions such as COVID-19 and chronic obstructive pulmonary disease (COPD). Lastly, Idrisoglu et al. (2022) [7] present a systematic review on various machine learning models used in voice analysis, finding that Support Vector Machine (SVM) and Neural Network models show high accuracy in diagnosing voice disorders. Despite key findings presented in the current literature, several gaps remain. Most studies focus on adult populations leaving a gap in research on vocal biomarkers for the diagnosis of pediatric-specific conditions. Additionally, most studies are region-specific, limiting the generalizability of findings across more diverse populations. Lastly, there is a lack of comprehensive reviews that synthesize findings across many condition types. Our scoping review aims to synthesize existing knowledge on the application of AI in the analysis of pediatric voice as a biomarker for health to foster a deeper understanding of its potential use as an investigative or diagnostic tool within the pediatric clinical setting.

2. Materials Methods

2.1. Registration and Funding

This scoping review was registered with the Open Science Framework (OSF) to enhance transparency and reproducibility. The review was registered on 24 July 2023 under the OSF registration https://doi.org/10.17605/OSF.IO/SC6MG The full registration details, including the review protocol and objectives, can be accessed at https://osf.io/sc6mg (accessed on 23 May 2024). All phases of this study were supported by the National Institutes of Health grant number: 1OT20D032720-01.

2.2. Search Strategy

Precise searches were conducted to identify relevant keywords and controlled vocabulary for the following concepts: artificial intelligence, voice, pediatrics, and disorders. Controlled vocabulary terms were combined logically by a medical librarian using Boolean logic, with keywords searched in the title and abstract to form a sensitive search strategy. The final search strategy utilized 217 keywords, including 91 related to “artificial intelligence”, 45 related to “voice”, 20 related to “pediatric”, and 61 related to “disorder”, as shown in Appendix A. The original PubMed search was translated into the following databases: Embase, Web of Science Core Collection, and the Cochrane database. Google Scholar and ClinicalTrials.gov were searched in order to pull in the grey literature. All searches were run in May 2023 and de-duplicated in EndNote using the validated deduplication method put forth by Bramer et al. [8]. Results were imported into Covidence, a systematic review software. Titles and abstracts were independently reviewed by two reviewers against pre-defined inclusion criteria. Relevant texts were moved to the full-text review, whereby the same process evaluated PDFs of eligible citations. Conflicting votes were resolved via discussion until the two original reviewers reached a consensus. The PRISMA flow chart of article inclusion is shown in Figure 1.

2.3. Inclusion Criteria

Each study was required to include voice recordings in pediatric populations aged 0–17 years. Studies involving both pediatric and adult cohorts were considered on the basis that pediatric data were collected and analyzed separately from adult data. A minimum of 10 pediatric participants were required in each study. All pediatric health conditions were considered except for newborn or infant cry to indicate hunger, discomfort, pain, or sleepiness. Studies were limited to peer-reviewed prospective or retrospective research studies written originally in English and excluded scoping reviews, literature reviews, and meta-analyses. Studies were required to utilize one or more feature extraction methods to produce a vocal dataset and required an analysis of pathological biomarkers contained in voice, cry, or respiratory sounds using one or more machine learning or artificial intelligence models.

2.4. Data Extraction

At the final stage, 62 studies met the inclusion criteria (Figure 1). A study was eligible for data extraction after two independent reviewers reached a consensus on its inclusion in the title, abstract, and full-text review phases. Utilizing the data extraction template in Covidence, we customized a tool to collect general study information, study characteristics, participant characteristics, recording sources and data, feature extraction methods, and machine learning or artificial intelligence model types. When available, accuracy, sensitivity, and specificity data were collected for each diagnostic model. These data were synthesized to determine the combination of feature extraction method(s) and artificial intelligence model that results in the highest diagnostic accuracy. Only models that used at least one feature extraction method and presented data for accuracy, sensitivity, and specificity were considered for highest diagnostic accuracy. When more than one model was presented by a study, only the model with the highest diagnostic accuracy was summarized. Four condition groups were determined: developmental conditions, respiratory conditions, speech language conditions, and other non-respiratory conditions. The best models were determined for each condition type within each condition group, with the exception of other non-respiratory conditions.

3. Results

3.1. Global Representation

Across 62 studies, 25 countries were represented (Appendix B). The global distribution and frequency of publication are shown in Figure 2. Pediatric populations from the United States, India, and China were the most frequently studied. Data primarily represented pediatric populations from North America, Asia, Europe, and Oceania and were less representative of Central and South America, Africa, and the Middle East.

3.2. Studies by Year

This review identified studies published between 2015 and 2023, and data were extracted on 25 May 2023. The number of studies per year is shown in Figure 3, with an average of 7 pediatric voice studies per year between 2015 and 2023 and a peak of 15 publications in 2019.

3.3. Funding Sources

Research funding supported 29 studies (46.7%), and 56 different funding sources were represented (Appendix C). Organizations that provided funding to two or more studies included the National Natural Science Foundation of China, Manipal University Jaipur (India), SMART Innovation Centre (USA), Austrian National Bank, National Institute on Deafness and Other Communication Disorders (USA), Austrian Science Fund, Natural Sciences and Engineering Research Council of Canada, and the Bill & Melinda Gates Foundation (USA). Most funding came from public and private organizations from the United States, China, India, and Austria.

3.4. Participant Age

Each study had, on average, 202 participants [range: 12–2268], with a median of 76 participants. A total of 27% of participants (n = 3347) were distinguished by sex, of which 61% were male. School-aged children (ages 5–12 years) were the most commonly studied (25 studies). Newborn (ages 0–2 months), infant (ages 3–11 months), toddler (ages 1–2 years), preschool (ages 3–4 years), school-aged (ages 5–12 years), and teenage (ages 13–17) groups were also represented in at least five studies each as shown in Figure 4. The specific pediatric age group being studied was not defined for 12 studies.

3.5. Recording Characteristics

As shown in Figure 5, studies included three types of vocal recordings: voice (38 studies), cry (13 studies), and respiratory sounds (12 studies). The majority of studies (45 studies) collected unique vocal data, while 17 studies utilized 13 different existing datasets to conduct their studies, of which recordings from the Baby Chillanto Infant Cry Database (Mexico) and the LANNA Research Group Child Speech Database (Czech Republic) were the most commonly studied.

3.6. Clinical Conditions

Vocal recordings were analyzed, using AI, as a biomarker for 31 clinical conditions, represented in Appendix D. Among these conditions, developmental conditions (21 studies), respiratory conditions (21 studies), speech and language conditions (13 studies), and non-respiratory conditions (7 studies) were represented. The most frequently studied conditions included autism spectrum disorder (ASD) (12 studies), intellectual disabilities (7 studies), asphyxia (7 studies), and asthma (5 studies).

3.7. Feature Extraction Methods

Among 62 studies, 33 feature extraction methods were utilized (Appendix E). Mel-Frequency Cepstral Coefficients (MFCCs) were the most utilized feature extraction method (43 studies), followed by Spectral Components (10 studies), Cepstral Coefficients (10 studies), Pitch and Fundamental Frequency (9 studies), and Linear Predictive Coefficients (9 studies).

3.8. Artificial Intelligence and Machine Learning Models

Across the studies, 33 artificial intelligence or machine learning models were utilized (Appendix F). The most common AI/ML models were Support Vector Machine (SVM) (34 studies), Neural Network (31 studies), Random Forest (9 studies), Linear Discriminant Analysis (LDA) (7 studies), and K-Nearest Neighbor (KNN) (5 studies).

3.9. Model Accuracy

Among 54 studies (excluding those on non-respiratory conditions), 85 models were summarized based on feature extraction methods, AI model types, and their diagnostic accuracy, sensitivity, and specificity. Out of these, 39 models were evaluated and compared for diagnostic accuracy (Appendix H). Fifteen models achieved high diagnostic accuracy: three for developmental conditions (

Table 1. Developmental Conditions: Models with Highest Diagnostic Accuracy.

Developmental Conditions: Models with Highest Diagnostic Accuracy
Study	Condition Type	Recording Type	Feature Extraction Methods	Artificial Intelligence Model	Accuracy	Sensitivity	Specificity
Jayasree 2021	Autism Spectrum Disorder	Voice	Mel Frequency Cepstral Coefficients	Neural Network	100	100	100
Sharma 2020	Intellectual Disability	Voice	Local Binary Patterns	Support Vector Machine	98.7	99	99.2
Ribeiro 2020	Dyslexia	Voice	Local Binary Patterns	Support Vector Machine	94.4	100	80

), eight for respiratory conditions (

Table 2. Respiratory Conditions: Models with Highest Diagnostic Accuracy.

Respiratory Conditions: Models with Highest Diagnostic Accuracy
Study	Condition Type	Recording Type	Feature Extraction Methods	Artificial Intelligence Model	Accuracy	Sensitivity	Specificity
Hariharan 2018	Asphyxia	Cry	Linear Predictive Coefficients Nonlinear Entropies Wavelet Packet Transform	Improved Binary Dragonfly Optimization	100	100	100
Gouda 2019	Wheezing	Respiratory	Discrete Wavelet Transform Mel Frequency Cepstral Coefficients Short Time Fourier Transform;	Neural Network	100	100	100
Amrulloh 2015a	Cough Segments	Cough	Mel Frequency Cepstral Coefficients Non-Gaussianity Score Shannon Entropy Zero Crossing Rate	Linear Discriminant Analysis	97.4	97.5	92.8
Amrulloh 2018	Wet/Dry Cough	Cough	Formant Frequency Mel Frequency Cepstral Coefficients Non-Gaussianity Score Shannon Entropy	Neural Network	96.4	96.5	96.6
Khalilzad 2022a	Respiratory Distress Syndrome	Cry	Cepstral Coefficients Harmonic-to-Noise Ratio	Support Vector Machine	95.3	95	95
Amrulloh 2015b	Asthma	Cough	Formant Frequency Mel Frequency Cepstral Coefficients Non-Gaussianity Score Shannon Entropy	Neural Network	94.4	100	88.9
Sharan 2017	Croup	Cough	Mel Frequency Cepstral Coefficients Cochleagram Image Feature	Support Vector Machine	91.2	88.4	91.6
Sharan 2021	Whooping Cough	Respiratory	Mel Frequency Cepstral Coefficients Cochleagram Image Feature	Neural Network	90.5	95.2	85.7

), and four for speech–language conditions (

Table 3. Speech Language Conditions: Models with Highest Diagnostic Accuracy.

Speech Language Conditions: Models with Highest Diagnostic Accuracy
Study	Condition Type	Recording Type	Feature Extraction Methods	Artificial Intelligence Model	Accuracy	Sensitivity	Specificity
Hariharan 2018	Deafness	Cry	Linear Predictive Coefficients Non-Linear Entropy Wavelet Packet Transform	Improved Binary Dragonfly Optimization	100	100	100
Barua 2023	Speech Language Impairment	Voice	Wavelet Packet Decomposition	Support Vector Machine	99.9	99.9	99.9
Miodonska 2016	Sigmatism	Voice	Delta Coefficients Mel Frequency Cepstral Coefficients	Support Vector Machine	94.5	96	93
Suthar 2022	Speech Disorder	Voice	Landmark Analysis	Linear Discriminant Analysis	93	94	92

).

For diagnosing developmental conditions, the best models often utilized voice recordings with Mel-Frequency Cepstral Coefficients (MFCCs) and Support Vector Machines (SVM) or Neural Networks. Jayasree (2021) [9] achieved 100% accuracy, sensitivity, and specificity for diagnosing autism spectrum disorder using MFCCs and a Feed-Forward Neural Network.

For respiratory diagnosis, the top models frequently used recordings of coughs, respiratory sounds, and cries. MFCCs were the most common feature extraction method, often combined with the Non-Gaussianity Score. Neural Networks and SVM were the most utilized AI models. Notably, Hariharan (2018) [10] and Gouda (2019) [11] achieved 100% accuracy for diagnosing asphyxia and wheezing, using Improved Binary Dragonfly Optimization and Artificial Neural Networks, respectively.

In the category of speech–language conditions, voice recordings were commonly used, but there was no dominant feature extraction method among the four models with high accuracy. Hariharan (2018) [10] and Barua (2023) [12] developed models with 100% and 99.9% accuracy for detecting deafness and speech–language impairment, using Improved Binary Dragonfly Optimization and SVM.

Overall, MFCCs combined with SVM resulted in the highest diagnostic accuracy across all condition groups and types.

4. Discussion

The human voice contains unique, complex acoustic markers that vary depending on one’s coordination between respiration, phonation, articulation, and prosody. As technology progresses, especially in artificial intelligence and acoustic analysis, voice is emerging as a cost-effective, non-invasive, and accessible biomarker for the detection of pathologies. Our primary objective was to determine what is currently known about using pediatric voice paired with AI models for the early detection, diagnosis, and monitoring of pediatric conditions. This review identified 62 studies that met the inclusion criteria, utilizing pediatric voice, cry, or respiratory sounds for the detection of 31 pediatric conditions among four condition groups, representing pediatric populations from 25 countries.

4.1. Developmental Conditions

Twenty-one of the included studies trained and evaluated machine learning algorithms using voice data to classify children with developmental disorders. Speech was the predominantly utilized feature, with studies considering various aspects of speech, including vocal, acoustic, phonetic, and language features. Acoustic features [13,14,15] and phonetic features [16] were extracted to train machine learning algorithms in classifying children with intellectual disabilities. A majority of the included studies centered on training machine learning algorithms to classify children with autism spectrum disorder (and Down Syndrome [9]) using acoustic features [9,17,18,19,20,21], vocal features [22,23], voice prosody features [24], pre-linguistic vocal features [25], and speech features [26,27]. In particular, Wu et al. (2019) [21] focused on acoustic features of crying sounds in children of 2 to 3 years of age, while Pokorny et al. (2017) [28] concentrated on pre-linguistic vocal features in 10-month-old babies. Speech features were also utilized in training machine learning algorithms to classify children with developmental language disorders [29], specific language impairments [30,31], and dyslexia [32,33]. Acoustic and phonetic features were commonly extracted using Mel-Frequency Cepstral Coefficients to train Neural Network or Support Vector Machine algorithms.

4.2. Respiratory Conditions

Twenty-one of the included studies focused on the unintentional air movement across vocal cords by cry, cough, or breath. Machine learning techniques characterized infant cries in the setting of asphyxia [34,35,36,37,38]. Spontaneous pediatric coughs are rigorously described through AI methodology [39,40,41,42,43] and analyzed to detect specific clinical entities such as croup [42,44,45], pertussis [45], asthma [46], and pneumonia [40]. Asthma, a common childhood illness, has also been studied through AI analysis of pediatric breath sounds [47,48]. Nearly all respiratory studies utilized Mel-Frequency Cepstral Coefficients to extract features from cough, respiratory sound, and cry recordings which were applied to Neural Network and Support Vector Machine algorithms.

4.3. Speech and Language Conditions

The detection and evaluation of voice and speech disorders in children is uniquely challenging due to the intricate nature of speech production and the variability inherent in children’s speech patterns. To address these challenges, researchers have explored a variety of computational approaches leveraging machine learning, neural networks, and signal processing techniques aimed toward the early identification of speech delay [49,50]. Several studies highlight promising methodologies to identify stuttering and specific language impairment (SLI) using acoustic and linguistic features [51,52]. Feature extraction techniques and convolutional neural networks can help to detect hypernasality in children with cleft palates [53,54,55]. Voice acoustic parameters have been developed to identify dysphonia and vocal nodules in children [56,57]. Automatic acoustic analysis can also be used to differentiate typically developing children from those who are hard of hearing, language-delayed, and autistic [58]. Other notable research has utilized deep learning models and computer-aided systems to identify SLI and sigmatism, also known as lisping [59,60,61]. Similar to developmental and respiratory condition diagnosis, Mel-Frequency Cepstral Coefficients were the preferred feature extraction method. However, there was variation in the AI model used. It is notable that across the studies included, each model presented for the diagnosis of speech and language conditions achieved diagnostic accuracy ≥ 86 percent, most with accuracy ≥ 90 percent.

4.4. Other Non-Respiratory Conditions

Researchers have explored using voice recordings and AI to identify other non-respiratory genetic or medical conditions, usually based on known characteristics affecting cry, voice, or speech that can lead to a clinical suspicion that a diagnosis is present. A Voice Biometric System was developed using recordings of 15 two-syllable words to identify whether a child has cerebral palsy and the severity of the condition, with potential usefulness to evaluate therapeutic benefit [62]. A hierarchical machine learning model using voice recordings of the standardized PATA speech test was able to identify and grade the level of severity of dysarthria associated with ataxia [56]. Early detection of anxiety and depression using a 3 min Speech Task in 3-to-8-year-olds showed reasonable accuracy when recordings were high quality [63], and multimodal text and audio data were able to discriminate adolescents with depression based on recorded interviews [64]. Recordings of cry sounds have also been evaluated using machine learning and have shown reasonable accuracy in detecting life-threatening sepsis in neonates [65,66] and neonatal opioid withdrawal syndrome [67]. Mel-Frequency Cepstral Coefficients, Cepstral Coefficients, and Harmonic-to-Noise Ratio were most often used for feature extraction and applied to Support Vector Machine algorithms for the detection of non-respiratory conditions. The data are best viewed by condition type, due to the variety of conditions included in this group.

4.5. Limitations

This review was restricted to studies published in English, which may not capture the full scope of research in non-English-speaking regions. Additionally, the inclusion criteria required a sample size of at least 10 pediatric participants in each study. Studies with fewer than 10 participants may offer valuable insights, but they did not meet inclusion criteria within this review.

Another limitation of this study was the inherent lack of standardization in feature extraction methods and AI model architectures. This variability makes it challenging to compare results across studies. To identify models with the highest diagnostic accuracy, we designed a strict inclusion criteria for comparison. However, it should be noted that there is considerable nuance in the development of models within the current state of voice diagnostics. Standardizing these methods in this developing field could facilitate more accurate comparisons and lead to more robust meta-analyses in the future.

5. Conclusions

This scoping review highlights the current and potential applications of AI in analyzing pediatric voice as a biomarker for health. AI models have been used with pediatric voice data for the early detection, diagnosis, and monitoring of 32 pediatric conditions, including autism spectrum disorder (ASD), intellectual disabilities, asphyxia, and asthma. While most applications focus on developmental, respiratory, and speech and language conditions, this review also explores using pediatric voice analysis for detecting non-respiratory conditions such as anxiety and depression, sepsis, and jaundice.

Research thus far demonstrates the enormous potential of using voice recordings to detect and monitor diseases and conditions in children. The data indicate that using Mel-Frequency Cepstral Coefficients (MFCCs) as a feature extraction method, combined with Neural Networks or Support Vector Machines, results in high diagnostic accuracy across various conditions and vocal recording types. However, the lack of standardization in feature extraction methods and AI model architectures presents a challenge in comparing results across studies. Future research should investigate the use of standardized methods to facilitate more accurate comparisons and robust meta-analyses. Additionally, future research should explore the use of AI diagnosis for pediatric conditions not yet covered in this review.

While most studies have recorded voices in clinical settings, there is potential for using voice recordings as biomarkers in non-clinical settings where children are more comfortable, such as at home or school. Further development in this field could lead to innovative diagnostic tools and interventions for pediatric populations globally.