Early Identification of Autism Using Cry Analysis: A Systematic Review and Meta-analysis of Retrospective and Prospective Studies

This systematic review included all levels of research evidence (e.g., randomized controlled trials, observational studies, longitudinal studies, cross-sectional studies, and quantitative research) and aimed to integrate best practice systematic review methodology, the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (Moher et al., 2010).

A comprehensive search strategy was employed to identify relevant studies on cry analysis in relation to early autism identification. The PICO (Participant, Intervention, Comparison, Outcome) framework was used to structure the search related to the research questions (Cooke et al., 2012). We systematically searched multiple electronic databases that cover key biomedical, psychological, and multidisciplinary research areas, including PubMed, Scopus, Web of Science, and PsycINFO, using keywords and subject headings related to cry analysis and autism. Other databases, such as EMBASE, were excluded due to significant overlap with PubMed and Scopus, ensuring comprehensive yet efficient study identification. The search terms included "autism spectrum," "cry," and “screening;” and the search was conducted on 09 September 2024. The keyword combinations used in the literature search are presented in Table 1. Electronic database searches were performed using the key terms related to “Population” (Newborns with high-risk or with a diagnosis of ASD), “Intervention” (early detection of ASD), and “Outcome” (acoustical cry analysis). For the outcome, we included only cry terms to capture all potentially relevant articles. The search strategy did not specify a comparison/control group. These terms were considered in the inclusion and exclusion criteria. After retrieving studies from the searches, duplicates were removed, and the paper titles, abstracts, and associated meta-data were compiled into a single table for further review. We also reviewed reference lists of relevant articles to identify additional studies.

All research within Oxford levels of evidence I–IV (Howick, 2011), including case studies and single-case experimental design studies reporting objective outcome measures were eligible for inclusion, if they met the following criteria: a) Studies focusing on cry analysis; b) Studies conducted on children under 60 months of age (this age range encompass early autism identification and the period of significant neuroplasticity, aligning with the typical timeline of diagnosis and recognition by caregivers and educators); c) The evaluation process follows a uniform diagnostic protocol (including the last three versions of the Diagnostic and Statistical Manual of Mental Disorders (DSM) criteria (DSM-IV, DSM-IV-TR, DSM-V)); or diagnosis made by a specialist in neurology, psychiatry, and pediatrics; or based on an extensive, rigorous and detailed neuropsychological evaluation; d) Reported in full text; and e) English language manuscripts.

The exclusion criteria taken into account were as follows: a) Participants with other medical conditions (included neurological, neurodevelopmental, genetic, perinatal, metabolic, and chronic illnesses that could confound cry analysis); b) Information relative to statistical analysis, ML, or AI was not reported; c) Utterance analysis was based on other types of infant vocalizations such as speech or babbling, because they do not directly assess cry-related biomarkers.

From the systematic review, the final eligible studies included were assessed to evaluate crying as a potential vocal biomarker for autism. For inclusion in the meta-analysis, only studies that reported sufficient data on F0, such as the mean and standard deviation for both AutI/EL and TD/DL groups, were retained. The studies were included if they met the following criteria:

To ensure reliability, prevent duplication, and adhere to standard systematic review protocols, this review was registered in PROSPERO (publication code: CRD42024592154). Selected articles were imported into Rayyan (Ouzzani et al., 2016), used to facilitate the screening process through its collaborative and filtering functionalities. This was done following the revision process by two reviewers (BC and SP), as recommended in PRISMA guidelines (Moher et al., 2010), based on the established selection criteria. The manuscripts selected in the title and abstract review were thoroughly reviewed. The disagreements regarding the selection were resolved through discussion or a third reviewer (screening protocols and final selection are depicted in Fig. 1). The third reviewer (AL) arbitrated disagreements that could not be resolved by discussion. An almost perfect level of agreement was obtained for title, abstract screening, and full-text screening (Cohen's kappa coefficient, k = 0.983, SE-kappa = 0.017, CI:0.949–1.00).

The included studies' quality and risk of bias were assessed using the National Heart, Lung and Blood Institute (NHLBI) quality assessment tool for observational cohort and cross-sectional studies (NHLBI, 2021). The NHLBI tool is a widely recognized framework designed to systematically evaluate the internal validity and reliability of studies, focusing on factors such as study design, sample size, and measurement methods. Two different reviewers (JAZV and SP) independently applied the protocol as mentioned earlier. Discrepancies were solved through discussion or by a third reviewer (BC).

Data were extracted using a standardized form to ensure consistency and accuracy. The information extracted included study characteristics such as Study Design, Audio Features studied, Statistical and ML Analysis, Findings, Sample size and Study limitations. Two reviewers (AL and SP) performed this extraction process and discrepancies during data extraction were resolved by consensus between the two reviewers. The information is summarized in Table 2 and Table 3 according to the study type (retrospective vs. prospective studies). The Study limitations presented in Tables 2 and 3 reflect those identified by each author in their respective manuscripts. Most of these limitations are common challenges in this type of research, such as heterogeneity, sample size, and issues related to audio recordings.

Data for the meta-analysis were independently extracted by two reviewers (SP and AL) to guarantee accuracy and consistency following the full-text review conducted during the systematic review. Discrepancies during meta-analysis data extraction were resolved by consensus between the two reviewers. The extracted information included (Table 4):

The meta-analysis was conducted using Jamovi software (version 2.4.14) and its meta-analysis module (MAJOR) to assess the variability and reliability of findings across studies. SMD were calculated to compare F0 values between AutI/EL and TD/DL groups, enabling the integration of data from studies with different measurement scales. Positive SMD values indicated higher F0 in AutI/EL infants, while negative values indicated higher F0 in TD/DL infants. The random-effects model, following the DerSimonian and Laird method (DerSimonian & Laird, 2015), was used to account for variability in study designs, populations, and methodologies, providing a weighted average effect size while assuming true effect sizes varied across studies.

Heterogeneity among the included studies was evaluated using Cochran’s Q test to determine statistical significance, the I² index to quantify the proportion of variation due to heterogeneity rather than chance (with values interpreted as low, moderate, or high at 25%, 50%, and 75%, respectively). The tau² statistic was used to estimate the variance in effect sizes. Model fit was assessed using Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), where lower values indicated better fit and model parsimony.

The overall summary effect and 95% confidence intervals (CI) were calculated and visualized using forest plots, which displayed individual study effect sizes alongside the pooled effect size. Publication bias was assessed through funnel plot analysis (Fig. 2) and further evaluated statistically using Egger’s regression test and Begg’s rank correlation test, with significant results indicating potential bias (Table 5). To ensure robustness, sensitivity analyses were performed by excluding studies with small sample sizes, outliers, or methodological differences, and any changes in effect size or heterogeneity were reported.

Risk of bias assessment revealed that 7 out of 11 studies were of fair quality, while the remaining four were rated as good (Esposito et al., 2014, p. 201; Esposito & Venuti, 2010b; Manigault et al., 2023; Unwin et al., 2017). None were classified as poor quality. Key strengths included appropriate sample size justification and power analysis, though limitations were noted in reporting methods for cry measurements and blinding of outcome assessors. Additionally, few studies provided validation for cry measurement protocols, and inconsistencies in the reporting of diagnostic and exposure data were common (Tables 2 and 3).

Table 2 summarizes five retrospective studies comparing AutI and TD infants. These studies analyzed cry features across a wide age range (5 months to 4 years). Statistical methods, including ANOVA, GLM, and correlation analysis, were employed in four studies (Esposito & Venuti, 2009, 2010b, 2010a; Moffitt et al., 2022), while one applied ML using Support Vector Machines (SVM) (Khozaei et al., 2020). Two studies include 20 infants (Esposito & Venuti, 2009, 2010a), one includes 28 children (Esposito & Venuti, 2010b), and the most recent studies report 62 (Khozaei et al., 2020) and 61 participants (Moffitt et al., 2022), respectively. ADOS-2 is the primary diagnostic tool used in 4 out of 5 studies. Notably, one study (Khozaei et al., 2020) has employed ML to distinguish AutI cries from TD ones. Specifically, SVM demonstrated classification accuracies exceeding 85% when distinguishing AutI from TD infants based on cry features like F0, Mel Frequency Cepstral Coefficients (MFCC), and quality voice features.

Regarding statistical analysis, in general, the studies highlight key audio features, specifically the cry duration and F0. These retrospective studies consistently found that AutI exhibited more prolonged hyperphonation (F0 > 1000 Hz) (Esposito & Venuti, 2009) periods, fewer silence or pauses and shorter cry vocalizations (Esposito & Venuti, 2009), more dysphonation or irregular patterns in cry amplitude (Esposito & Venuti, 2009), and higher F0 (Esposito & Venuti, 2010a, 2010b), which are statistically significant markers. Additionally, F0 was also predicted by developmental assessments and diagnostic tools. Specifically, (Moffitt et al., 2022) suggest that both acoustic and phonological features of child vocalizations correlate with expert clinician evaluations of autism-related traits.

Six prospective studies examining EL and DL infants are detailed in Table 3. These six studies encompass a diverse age range, spanning from neonates to 18 months infants. Two of them used ML (SVM, Random Forest), and four applied Statistical Analysis (ANCOVA, ANOVA, t-test, Pearson). One study included the smallest sample size of 24 infants (Orlandi et al., 2012), while four studies enrolled between 40 and 50 infants (Esposito et al., 2014; Santos et al., 2013; Sheinkopf et al., 2012; Unwin et al., 2017). The largest sample size is reported by Manigault et al., 2023, who included 363 children. The ADOS was utilized in most studies, with the exception of Orlandi et al., 2012, which did not report any diagnostic assessment, and Manigault et al., 2023, which employed a broader developmental assessment approach combined with parental questionnaires (Bayley Scales of Infant and Toddler Development: Bayley-III; M-CHAT; and Child Behavior Checklist: CBCL, respectively).

These studies generally supported the findings of the retrospective studies, with EL infants showing distinct cry features, particularly increased F0 and shorter cry vocalizations. Prospective studies also identified more hyperphonation (F0 > 1000 Hz) on pain cries (Sheinkopf et al., 2012), higher resonance frequencies (F1 and F2) (Orlandi et al., 2012), smaller amplitude or energy (Sheinkopf et al., 2012), with no difference in phonation (Sheinkopf et al., 2012) and melodic patterns (Orlandi et al., 2012). Some studies (Esposito et al., 2014; Sheinkopf et al., 2012) show discrepancies in F0 variation. The same studies show F0 higher in EL infants while F0 is reported lower in (Orlandi et al., 2012) and (Unwin et al., 2017) although the studies included a small sample size.

One noteworthy AI model (Santos et al., 2013) achieved a 97.796% accuracy in effectively differentiating cries between EL and DL infants. Additionally, the study conducted by (Manigault et al., 2023), introduces a Random Forest algorithm capable of discerning EL cries in infants as young as 1 month. Although the accuracy of this model is not explicitly reported (Manigault et al., 2023), the model trained using acoustic cry characteristics was associated with clinical and developmental assessments at 2 years, being amplitude, F0, resonance frequencies, and signal quality as the most relevant variables.

Several consistent trends were observed across both retrospective (see Table 2) and prospective studies (see Table 3) for specific audio features that may help differentiate cries of AutI/EL from TD/DL infants:

ML models, particularly SVM and Random Forest classifiers, consistently outperformed traditional statistical methods, achieving classification accuracies upwards of 90% across various studies (Khozaei et al., 2020; Manigault et al., 2023; Santos et al., 2013), validating their use in cry analysis supporting early autism identification.

The gold standard for ASD diagnosis combines DSM-5 criteria with the ADOS, developmental assessments, and parental questionnaires (Mcmorris et al., 2013; Wiggins et al., 2019). The reviewed studies consistently relied on established tools, including the DSM-5 criteria, ADOS (assessment for social interaction and communication), developmental assessments such as the Mullen Scales of Early Learning (Akshoomoff, 2006), Griffiths Mental Development Scales (Pino et al., 2024), MacArthur Communicative Development Inventory (Charman et al., 2003), Child Behavior Checklist (CBCL) (Rescorla et al., 2019) or Bayley-III (evaluation for cognitive, motor, and developmental milestones) (Torras-Mañá et al., 2016). These tools were often used in combination with parental input and medical history to provide a comprehensive diagnostic framework. These methods align with international practices (Lordan et al., 2021), though some regions also utilize tools like Childhood Autism Rating Scale (CARS) (Rellini et al., 2004), Gilliam Autism Rating Scale (GARS) (Samadi & McConkey, 2014) or rely more on clinical judgment due to limited access to specialized instruments.

Most studies diagnosed autism between 18 and 54 months, with a peak around 36 months. While early signs were detected as young as 12 months, delays in formal diagnosis—often beyond age 4—highlight the need for earlier developmental screenings to address this gap.

A total of six studies were included in the analysis of F0 comparing AutI/EL and TD/DL infants (Table 4). The estimated overall SMD for F0 differences was 0.685 (standard error = 0.358), yielding a Z-value of 1.91. However, the result was not statistically significant at the 5% level (P = 0.114), as the CI: – 0.24 to 1.60 included zero, indicating a lack of strong evidence for differences in F0 between EL and DL groups (see Table 5 for more details on statistics).

The studies demonstrated substantial heterogeneity, with a Tau² value of 0.6196 (SE = 0.4943) and an I² of 80.2%, reflecting high variability in effect sizes across studies. This variability likely stems from differences in study designs, sample sizes, methodologies, and the age ranges of infants studied (0 to 18 months). Model fit statistics, including an AIC of 18.477 and aBIC of 22.477, indicated a reasonable fit but highlighted the need for greater consistency in study designs (see Table 5).

Individual study findings varied, as illustrated in the forest plot (Fig. 2A). Studies by (Esposito & Venuti, 2010a, 2010b; Esposito et al., 2014) reported positive effect sizes with a CI that did not cross zero, indicating significant differences in F0 between EL and DL groups. In contrast, studies by (Orlandi et al., 2012) and (Unwin et al., 2017) found non-significant differences.

Despite the trend towards higher F0 in EL infants (combined SMD = 0.68), the results for this meta-analysis were not statistically significant, underscoring the need for further research. The symmetrical funnel plot (Fig. 2B) suggested no significant publication bias, and the distribution of study results appeared consistent, supporting the reliability of the findings despite high heterogeneity.

The role of cry analysis in early autism identification is particularly important given the current challenges in diagnosing autism, which often depends on behavioral assessments that are not feasible, accessible and often challenging in very young infants. Early identification during infancy aligns with the critical developmental window when neuroplasticity is most pronounced, optimizing the effectiveness of early interventions (Dawson, 2008; Jacobson et al., 1998; Nadel & Poss, 2007; Towle et al., 2020). Evidence strongly supports the benefits of these interventions, ranging from cognitive and language improvements to better social outcomes, underscoring their transformative impact on children, families, and society (Boyd et al., 2010; Okoye et al., 2023; Pickles et al., 2016).

In brief, cry analysis, especially when combined with ML algorithms, represents a significant advancement in the field of early autism screening, potentially enabling clinicians to detect early signs or identify infants with an increased likelihood during routine pediatric check-ups, thereby facilitating timely early intervention programs.

The meta-analysis examined differences in the F0 of crying between AutI/EL and TD/DL infants, synthesizing data from six studies. The pooled effect size indicated a trend toward higher F0 in AutI/EL infants, but the result was not statistically significant. This finding aligns with hypotheses suggesting that atypical vocal characteristics could serve as early markers for autism (Sheinkopf et al., 2012). However, the lack of statistical significance and substantial heterogeneity among studies underscore the need for caution in drawing definitive conclusions. Variability in methodologies, including differences in F0 measurement techniques and the inclusion of infants at varying developmental stages, contributed to the inconsistency in findings. Notably, studies with significant results (Esposito & Venuti, 2010a, 2010b) highlight the importance of further research into these early acoustic markers. While no significant publication bias was detected, the substantial variability suggests a need for standardized methodologies in future research to improve comparability and robustness.

While the existing research provides a strong foundation for the use of cry analysis in early autism screening, there is a growing need for longitudinal large-scale studies that follow-up infants from the general population from birth through early childhood. Most prospective studies reviewed here focused on EL of autism, such as infants with siblings already diagnosed. While these studies provide valuable insights, they may not capture the full spectrum of autism traits or generalize to the broader population. Additionally, it is important to note that, despite the accuracy demonstrated by cry acoustic ML models in this review, they are not yet widely used in this field. This is likely due to the difficulty of gathering a sufficiently large, heterogeneous, and diverse sample size, as well as the low prevalence of relevant cases, which makes it challenging to train and validate robust ML models.

Therefore, future multicentric large-scale longitudinal studies following infants from various demographic, education, genders, races, ethnicities and socioeconomic backgrounds populations are essential to validate cry analysis as a universal screening automatic tool for early autism identification. Such studies would allow researchers to track the developmental trajectory of acoustic cry features over time, providing a more comprehensive understanding of how cry patterns evolve in both AutI and TD infants. Furthermore, identifying critical developmental windows where cry analysis is most predictive of later autism diagnoses could enhance early intervention strategies. The potential of cry analysis in early autism identification is supported by the success of similar technologies in clinical settings. For example, automated auditory screening tools like otoacoustic emission testing for newborn hearing loss (Akinpelu et al., 2014) have been seamlessly integrated into standard pediatric care due to their reliability and non-invasive nature. Furthermore, advanced diagnostic tools employing data from eye-tracking systems (Jones et al., 2023) and video-based behavioral analyses (Megerian et al., 2022), combined with ML algorithms, have demonstrated high accuracy and clinical utility in identifying early autism traits.

These approaches highlight how technology-driven, multimodal data-rich methodologies can be effectively incorporated into routine screenings, laying the foundation for cry analysis to become a practical tool in pediatric healthcare.

The studies analyzed in this systematic review and meta-analysis had several limitations that should be considered when interpreting the findings. Small sample sizes were a common issue, reducing statistical power and generalizability, particularly in studies involving ML and acoustic cry analysis. Additionally, there was significant heterogeneity in participants’ age, autism characteristics, and cry elicitation methods, complicating comparisons across studies. The variability in recording quality, often influenced by non-standardized environments, further compromised data reliability. The lack of standardized methods for eliciting and recording cries also posed challenges, with some studies not differentiating between cries triggered by pain, or other needs or emotions. Moreover, missing or incomplete developmental follow-up information in several studies hindered the ability to draw long-term conclusions about the relationship between early acoustic markers and later autism diagnosis. Furthermore, limited demographic reporting restricted insights into the influence of socioeconomic, cultural, and biological factors. In fact, details related to gender, ethnicity, race, educational, socio-economic background, and co-occurring conditions that could influence cry patterns are usually not reported. Missing developmental follow-up data in many studies precluded definitive conclusions about the long-term predictive value of early acoustic markers. Finally, the meta-analysis itself was constrained by the substantial heterogeneity in study design, population characteristics, and cry measurement methods, emphasizing the need for standardization in future research. Specifically, the meta-analysis of F0 exhibited high heterogeneity, driven by differences in sample demographics, cry recording methods, and analytical approaches. The limited number of studies further constrained subgroup analyses and precision of effect size estimates. The non-significant pooled result highlights the potential for underpowered conclusions. These challenges emphasize the need for more robust, large-scale, and standardized studies to validate cry analysis as a screening tool for autism.