From Clinic to Home: Validating the Social Communication Scale for Telemedicine-Supported Early Autism Screening in Chinese Toddlers Authors: Li Wang, Yujia Shi, Fei Yuan, Xiujuan Geng, Ziyan Meng, Yipeng Wang, Hon-Cheong So, Patrick C. M. Wong Published: 2026 ABSTRACT Telemedicine is reshaping how early autism detection can be delivered, particularly in resource-limited settings. Yet, most existing tele-assessment tools were developed in Western contexts, limiting their cultural applicability. To address this gap, we evaluated the Social Communication Scale (SCS), a brief video-based tool, for telemedicine-supported screening in Chinese toddlers. A total of 170 Chinese toddlers aged 16 to 30 months were recruited. Parent–child interaction videos were recorded at home and subsequently coded using the SCS scheme. Autism likelihood (Elevated vs. Low) was defined using the Autism Diagnostic Observation Schedule, Second Edition (ADOS-2) classifications. Reliability and validity were evaluated. Screening accuracy was evaluated using total-score cutoff, logistic regression, and Random Forest (RF) models. The SCS demonstrated good inter-rater reliability (intraclass correlation coefficient [ICC] = 0.87) and moderate convergent validity with ADOS-2 and parent-reported measures of socialization and communication. Both logistic regression and RF outperformed the total-score approach, with RF yielding more balanced performance across metrics. Including demographic variables further improved RF classification performance (area under the receiver operating characteristic curve [AUC] = 0.91; sensitivity = 0.91; specificity = 0.91). The SCS demonstrates potential as a telemedicine-supported screening tool for early autism identification. Further validation in independent samples and against clinical diagnoses is needed to establish its generalizability and diagnostic accuracy.Lay AbstractTelemedicine is changing how early autism detection can be delivered, particularly in areas where specialist services are difficult to access. Telemedicine refers to the use of the internet and other digital technologies to provide health care remotely. Although several telemedicine tools for autism screening and diagnosis have shown promising results, most were developed and tested in Western countries and may not work equally well in different cultural settings. In this study, we evaluated a telemedicine tool called the Social Communication Scale (SCS) in 170 Chinese toddlers aged 16 to 30 months. A short video of a parent and child playing naturally together at home or in another familiar setting was recorded. Trained coders then reviewed the videos and assessed the child’s early social communication behaviors associated with autism. The SCS was able to distinguish between toddlers identified by the Autism Diagnostic Observation Schedule, Second Edition (ADOS-2) as having an elevated or low likelihood of autism. Using artificial intelligence (machine learning) alongside the SCS further improved its accuracy. These findings suggest that the SCS could become a useful addition to early autism screening, particularly in places where specialist services are limited. With further testing in larger and more diverse groups of children, this approach could help more families access earlier assessment and support. Introduction The way people seek and receive health care has changed dramatically over the past two decades (Duffens et al., 2025). For families facing barriers related to distance, resources, or time, the ability to access specialist services from home has reshaped service delivery. Defined as the use of telecommunications and virtual technologies to provide care outside traditional clinical settings, telemedicine has become an essential bridge between patients and providers (Field, 1996; Stavropoulos et al., 2022). It has been successfully applied across diverse domains, including infectious diseases, mental health, stroke, and maternal care (Barbosa et al., 2021; Monaghesh & Hajizadeh, 2020). This rapid growth is driven by telemedicine’s capacity to enhance access, reduce costs, and promote health equity (Dorsey & Topol, 2016; Tilhou et al., 2024; Ye et al., 2023). Meta-analytic evidence further indicates that telemedicine often yields outcomes equivalent to or better than in-person care (Shigekawa et al., 2018). These advances reflect a broader shift toward more patient-centered health care. This shift holds particular promise for addressing the challenges of early identification in autism spectrum disorder (ASD). Despite evidence that autism can be diagnosed as early as 18 months (Zeidan et al., 2022), and that timely identification and intervention are crucial for improving both core symptoms and broader developmental outcomes (Anderson et al., 2014; Lord et al., 2018), many autistic children still experience delayed diagnosis, especially in developing countries (Anderson et al., 2014; Lord et al., 2018; van’t Hof et al., 2021). China provides a representative example, with 86.24% of autistic children not receiving a formal diagnosis until after 24 months, with delays particularly pronounced in rural and less-developed regions (Li et al., 2025). These delays are compounded by structural barriers, including the concentration of autism specialists in large urban cities, long travel distances to specialized clinics, and widespread misdiagnosis or under-recognition of early symptoms (Li et al., 2025; Y. Liu et al., 2016). These challenges highlight the urgent need for accessible and scalable approaches that can facilitate earlier identification and intervention for autistic children in China and beyond. In recent years, studies have increasingly explored the feasibility and effectiveness of telemedicine-based approaches for autism screening and diagnosis. Among the technological innovations in this area, video-based methods have emerged as one of the most feasible and satisfactory alternatives to face-to-face assessments (Alfuraydan et al., 2020; Dahiya et al., 2020). A recent meta-analysis of 41 studies identified three major applications of video in autism detection, including video conferencing, recorded videos, and machine learning–based analysis of short clips (Wang, Meng, et al., 2025). The results reported a pooled sensitivity of 0.88 and specificity of 0.76, indicating good sensitivity and moderate specificity overall. Specifically, the machine learning–based analysis showed the highest pooled sensitivity (0.93) and specificity (0.83) among all three categories. Such approaches may offer advantages over conventional statistical methods by capturing complex and nonlinear relationships among behavioral indicators (Stahl, 2024). However, most of these studies were conducted in the United States, with limited representation from other regions. Although machine learning approaches demonstrated the highest diagnostic accuracy, their performance is highly dependent on the quality and representativeness of their training datasets. The predominant reliance on data from a single cultural context raises concerns about the generalizability of these algorithms across populations. The importance of cultural and contextual factors cannot be overemphasized (Mailick et al., 2025). Autism is not experienced in isolation but within specific cultural contexts that shape how symptoms are expressed, recognized, interpreted, and reported (De Leeuw et al., 2020). Building on the awareness of cultural and contextual variations in autism-related social communication, Wang, Qi, et al. (2025) developed the first Social Communication Scale (SCS) specifically for Chinese preschoolers. The SCS uses video-recording method, in which parents record naturalistic play interactions with their child at home or in other familiar settings, and trained coders later evaluate the child’s social communication abilities from these recordings. The tool has been validated as an outcome measure for detecting changes in social communication behaviors, such as pointing, facial expressions, and requesting, over short intervention periods (Wang, Qi, et al., 2025). Psychometric evaluations indicate that the SCS demonstrates excellent inter-rater reliability (intraclass correlation coefficient [ICC] = 0.91) and strong convergent validity with established instruments, such as the Autism Diagnostic Observation Schedule, Second Edition (ADOS-2). Importantly, the SCS is sensitive to subtle behavioral changes. For example, the SCS captured a significant reduction in SCS total scores (average rate of change = −1.52) across a 20-week intervention (Wang, Qi, et al., 2025). While previous findings established the SCS as a reliable outcome measure developed for use in Chinese preschool-aged children, its potential utility for earlier autism identification remains underexplored. Moreover, it remains unclear whether more advanced analytic approaches, such as machine learning, can meaningfully enhance the screening performance of the SCS beyond conventional scoring methods. Such improvements could have important clinical implications, including reducing unnecessary referrals, improving the prioritization of children who may benefit from further evaluation, and increasing the efficiency of telemedicine-supported screening pathways. To address these gaps, this study evaluated the SCS in a younger cohort of toddlers aged 16 to 30 months. We aimed to (a) further examine its reliability and convergent and divergent validity in this age group, and (b) assess its ability to differentiate toddlers with elevated versus low likelihood of autism using both conventional statistical approaches, including a total-score cutoff and logistic regression, and a machine learning approach based on Random Forest (RF). By integrating psychometric evaluation with rigorously validated classification approaches, this study extends the potential application of the SCS from an outcome measure toward a reliable screening tool for early autism detection in telemedicine contexts. Method Participants Toddlers aged 16 to 30 months were recruited from China via social media platforms (e.g., WeChat) and through word of mouth between August 2023 to July 2025. Children with severe hearing or visual impairments were excluded. A total of 145 toddlers were recruited from the general community without requiring a prior clinical diagnosis. Among them, four (2.76%) had an older sibling diagnosed with autism. Ninety-four families provided information about their reasons for participating in the study. Of these, five children (3.45%) had been identified as elevated likelihood for autism requiring surveillance by clinicians. Twenty (13.8%) had not been diagnosed but were reported by parents to have developmental concerns. The remaining participants had neither a diagnosis nor reported concerns. Due to the limited number of elevated-likelihood cases identified in the community sample (n = 23), baseline data from an existing intervention cohort were included for an additional 25 toddlers aged 24 to 30 months (Wang et al., 2026). Within this subgroup, 11 children had a confirmed autism diagnosis, and 14 were identified as having an elevated likelihood for autism by qualified clinicians. As expected, children from the existing cohort demonstrated greater symptom severity compared with those recruited from the community. While this inclusion may introduce potential sampling bias, it also allowed the sample to better capture the full range of autism symptom severity. To evaluate its impact, sensitivity analyses were conducted excluding these participants. The final sample comprised 170 toddlers. Based on classifications from the ADOS-2 Toddler Module (Lord et al., 2012), children were categorized into low-likelihood (little-to-no concern) and elevated-likelihood (mild-to-moderate or moderate-to-severe concern) groups. Group comparisons on demographic, developmental, and clinical characteristics are presented in Table 1. The elevated-likelihood group showed higher autism symptom severity and lower language and adaptive functioning, consistent with established developmental profiles of autistic toddlers (Ellis Weismer et al., 2010; Georgiades et al., 2022; Saulnier et al., 2022). This group also included a higher proportion of males, shorter gestational age at birth, and had older paternal age, all of which are established risk factors for autism (Crump et al., 2021; Gardener et al., 2009; Modabbernia et al., 2017). Given that these variables are theoretically and empirically associated with autism, they were not treated as confounders and were therefore not adjusted for in the primary analyses. In contrast, significant group differences in child age and household income were also observed, which are not core features of autism and may reflect sampling variability. These variables were therefore treated as potential confounders and adjusted for in sensitivity analyses to ensure the robustness of the findings. Table 1.Demographic, Developmental, and Clinical Characteristics of Toddlers in the Low- and Elevated-Likelihood Groups.CharacteristicsLow likelihood(n = 122)Elevated likelihood(n = 48) p d Child Sex Female 59 (48.36%)12 (25%).005 Male 63 (51.64%)36 (75%)Age (months)23.54 (3.90)25.52 (3.18) <.001 −0.58 Gestational age at birth (weeks)39.31 (1.53)38.55 (1.46).003 0.51 Autism symptom ADOS-2 total score4.06 (2.36)16.13 (5.21) <.001 −2.99 Social affect 3.74 (2.23)14.02 (4.13) <.001 −3.10 Restricted and repetitive behavior 0.34 (0.86)2.13 (1.65) <.001 −1.36 M-CHAT-R/F Positive 1 (0.82%)8 (34.78%) <.001 (n = 122 vs.23) Negative 121 (99.18%)15 (65.22%)SCQ10.65 (4.34)16.85 (6.90) <.001 −1.08 Language ability CDI (% Rank)(n = 115 vs. 32) Vocabulary Production 46.20 (31.08)17.34 (21.96) <.001 1.07 Sentence Complexity 60.50 (26.78)31.25 (26.61) <.001 1.10 Late takers a Yes 23 (20%)21 (65.63%) <.001 No 92 (80%)11 (34.37) Adaptive skill VABS-3 (standard scores)(n = 117 vs. 47) Communication 91.50 (30.37)40.43 (29.92) <.001 1.69 Daily living 40.11 (20.11)20.64 (10.82) <.001 1.21 Socialization 68.33 (23.91)37.92 (17.34) <.001 1.46 Motor 59.14 (11.64)52.92 (9.32) <.001 0.59 Family information Maternal age (years)30.68 (3.23)31.69 (4.07).13−0.27Maternal education High school and below 1 (0.82%)1 (2.08%) College and bachelor’s 59 (48.36%)32 (66.67%).06 Master’s and above 62 (50.82%)15 (31.25%)Paternal age (years)31.95 (4.00)34.02 (5.59).02 −0.43 Paternal education High school and below 4 (3.28%)1 (2.08%) College and bachelor’s 71 (58.20%)33 (68.75%).44 Master’s and above 47 (38.52%)14 (29.17%)Annual household income (RMB)(n = 122 vs. 47) 100,000 and below 9 (7.37%)2 (4.25%) 100,001–200,000 5 (4.10%)13 (27.66%) 200,001–300,000 11 (9.02%)7 (14.89%) <.001 300,001–500,000 33 (27.05%)16 (34.04%) 500,001 and above 64 (52.46%)9 (19.16%)aA toddler was considered a late talker if the vocabulary production was equivalent to the 10th percentile or below corresponding to the age cohort. Values are presented as mean (SD) for continuous variables and n (%) for categorical variables. Continuous variables were compared between groups using Welch’s t-tests, and categorical variables were compared using chi-square tests. ADOS-2 = Autism Diagnostic Observation Schedule, Second Edition; SCQ = Social Communication Questionnaire; CDI = MacArthur-Bates Communicative Development Inventories (Words and Sentences, Long Form); VABS-3 = Vineland Adaptive Behavior Scales, Third Edition. Bold values indicate statistical significance at p < .05. Cohen’s d effect size interpretation: 0.2 = small, 0.5 = medium, 0.8 = large. Demographic, Developmental, and Clinical Characteristics of Toddlers in the Low- and Elevated-Likelihood Groups. A toddler was considered a late talker if the vocabulary production was equivalent to the 10th percentile or below corresponding to the age cohort. Values are presented as mean (SD) for continuous variables and n (%) for categorical variables. Continuous variables were compared between groups using Welch’s t-tests, and categorical variables were compared using chi-square tests. ADOS-2 = Autism Diagnostic Observation Schedule, Second Edition; SCQ = Social Communication Questionnaire; CDI = MacArthur-Bates Communicative Development Inventories (Words and Sentences, Long Form); VABS-3 = Vineland Adaptive Behavior Scales, Third Edition. Bold values indicate statistical significance at p < .05. Cohen’s d effect size interpretation: 0.2 = small, 0.5 = medium, 0.8 = large. Measures Social Communication Scale The SCS is a play-based observational tool developed to assess autism-related social communication behaviors in children through parent–child interactions. It comprises four semi-structured yet naturalistic activities: 4 min of free play, 4 min of pretend play, 2 min of snack time, and 2 min of bubble play. Across these activities, seven key domains of social communication are evaluated: (a) Spontaneous Vocalization, (b) Pointing, (c) Eye Contact, (d) Requesting, (e) Social Overtures, (f) Initiation of Joint Attention, and (g) Enjoyment in Interaction. Each domain is rated on a 7-point scale with total scores ranging from 0 to 49. Higher scores indicate greater difficulties in the domains. One of the key advantages of the SCS is its remote administration format, which allows assessments to be conducted in the home setting with minimal preparation from parents, thus increasing ecological validity and feasibility. The tool has demonstrated strong reliability and validity among Chinese-speaking autistic children (Wang, Qi, et al., 2025). In this study, we followed the protocol and coder training procedures described in the work by Wang, Qi, et al. (2025). Parents were provided with detailed instructions on how to set up the environment and conduct the SCS activities. During the tele-assessment, a trained facilitator guided the session via Tencent Meeting, ensuring appropriate camera positioning, toy selection, and face-to-face seating arrangement. The facilitator provided only procedural guidance (e.g., timing of activity transitions and adjustments to ensure the parent and child remained within the camera frame) during interaction and did not offer prompts, or interaction feedback that could influence parent–child behavior. Parents were encouraged to interact with their child naturally, and all sessions were video-recorded for subsequent coding. A total of 168 video recordings were independently coded using the SCS coding scheme by trained coders who were masked to group classification and all other assessment procedures and results. Two videos were excluded due to poor video quality that prevented reliable coding. Details of coder training and reliability procedures are reported in the work by Wang, Qi, et al. (2025). Autism Diagnostic Observation Schedule, Second Edition The ADOS-2 Toddler Module was administered by certified assessors, who were masked to the SCS administration and coding. This module includes a series of semi-structured, play-based tasks designed to elicit symptoms and behaviors related to autism in toddlers aged between 12 and 30 months. Scores are generated across two domains: Social Affect (SA) and Restricted and Repetitive Behaviors (RRB), which are summed to produce a total algorithm score. Higher scores indicate greater symptom severity. The ADOS-2 is widely recognized as the gold standard in autism diagnostic assessment, and its scoring classifications were therefore used as the primary reference for grouping toddlers rather than as a confirmed diagnostic outcome in this study. Specifically, based on the total algorithm score, toddlers were classified into one of three levels of concern: little-to-no concern, mild-to-moderate concern, or moderate-to-severe concern. For the purposes of analysis, toddlers in the little-to-no concern category were grouped as low likelihood, whereas those in the mild-to-moderate or moderate-to-severe concern categories were grouped as elevated likelihood. The ADOS-2 toddler module demonstrated high sensitivity (0.83–0.95) and specificity (1.00) in Chinese toddlers (Chen et al., 2023). Parent-Reported Autism Symptoms Social Communication Questionnaire (SCQ) is a 40-item caregiver-report screening tool designed to assess communication skills and social functioning in individuals potentially experiencing autism (Rutter et al., 2003). The current Chinese form was administered in this study. The Chinese version has been validated in children aged 2 to 12 years and demonstrates good internal consistency, moderate test–retest reliability, and high sensitivity (0.85–0.96) and specificity (0.88–0.98) (S. Liu et al., 2022). Although the Chinese version of the SCQ is recommended for children aged 2 years and above, it was included to provide complementary caregiver-reported information alongside clinician-administered ADOS-2 assessments. Given the limited availability of standardized caregiver-report measures targeting social communication in children younger than 30 months, the SCQ served as a pragmatic proxy for capturing parent-perceived functioning in this age group. SCQ scores were used for convergent validity analyses only and not for classification or screening purposes. Modified Checklist for Autism in Toddlers, Revised with Follow-Up (M-CHAT-R/F) is a two-stage parent-report screening tool used to identify toddlers at risk for autism (Robins et al., 2014). The first stage consists of 20 yes-or-no items assessing early social communication and behavior. Children scoring between 0 and 2 are classified as low-risk, and those scoring between 8 and 20 as elevated-risk. For scores in the medium-risk range (3–7), a structured follow-up interview is conducted to clarify initial responses. If the follow-up score remains ⩾2, the child is considered screen positive; if it is between 0 and 1, the child is classified as screen negative. The Chinese version of the M-CHAT-R/F has shown adequate inter-rater and test–retest reliability, good sensitivity (0.96) and specificity (0.87) (Guo et al., 2019). Only the 145-toddler sample completed the M-CHAT-R/F. Language Development MacArthur-Bates Communicative Development Inventories (CDI) is a parent-report measure designed to assess early language development in children aged 8 to 48 months (Hutchins, 2013). It comprises two forms: Words and Gestures (8–16 months), which evaluates emerging receptive and expressive vocabulary and the use of communicative or symbolic gestures, and Words and Sentences (16–48 months), which assesses vocabulary production and early grammatical development. The Chinese version of the CDI has been adapted and validated for use in Chinese-speaking populations (Tardif et al., 2008). In this study, the Words and Sentences form was administered. Two indices, including vocabulary production and sentence complexity, were derived. Adaptive Functioning Vineland Adaptive Behavior Scales, Third Edition (VABS-3) is a standardized instrument used to assess adaptive functioning across the lifespan (Sparrow et al., 2014). It evaluates an individual’s performance in four core domains: Communication, Daily Living Skills, Socialization, and Motor Skills. It can be completed through a semi-structured interview or caregiver/teacher rating forms. In this study, the caregiver-report form was used for all the toddlers, with parents rating their child’s everyday functioning across the four domains. The Chinese version of the VABS-3 has demonstrated moderate to excellent reliability and validity (Zhang et al., 2020). As standard scores are not available for toddlers younger than 24 months in the Chinese version, raw scores were used for analysis. Procedures Interested parents accessed the study registration form via a QR code embedded in recruitment posters circulated on social media platforms. The form included brief screening questions regarding the toddler’s age and the presence of any visual or hearing impairments to determine initial eligibility. Trained research assistants subsequently contacted all registrants by phone to confirm eligibility, explain study procedures, and address any questions. Parents who provided verbal consent during the phone call and written consent during the in-person visit participated in a series of assessments. The assessment protocol began with the online administration of the SCS, followed by in-person evaluations, including the ADOS-2 and the M-CHAT-R/F. In addition, parents completed online questionnaires, including the SCQ, the CDI, the VABS-3, and a family demographics survey. Data Analysis Statistical analyses were conducted using Python (Python Software Foundation, 2026), R (R Core Team, 2024), and JASP (JASP Team, 2023). Reliability To evaluate inter-rater reliability of the SCS, 40 videos (approximately 24% of the total sample) were randomly selected and independently double-coded by two trained coders. This proportion is consistent with recommended practice in observational coding studies, where double-coding of approximately 10% to 25% of the dataset is considered typical to obtain stable and representative estimates of reliability while maintaining feasibility (O’Connor & Joffe, 2020). Item-level and total score reliability were assessed using the ICC, implemented via the irr package (Gamer et al., 2019). Specifically, we employed a two-way random-effects model, which assumes that both raters and items are randomly sampled and assesses absolute agreement. Interpretation of ICC values followed Koo and Li (2016) guidelines: values above 0.90 were considered excellent, 0.75 to 0.90 as good, 0.50 to 0.75 as moderate, and below 0.50 as poor reliability. Validity Convergent validity, defined as the degree to which a measure correlates with other instruments assessing related constructs (Duckworth & Kern, 2012), was calculated using Pearson correlation coefficients between the SCS total score and a range of established measures. These included the ADOS-2 total score, and its SA and RRB subscores, to assess alignment with gold-standard observational measures of autism symptoms. Parent-reported SCQ and VABS-3 socialization scores were also included to evaluate consistency with parental perceived social traits. In addition, we examined correlations between the SCS and indicators of communicative competence, including the CDI vocabulary production and VABS-3 Communication scores. Because nearly all M-CHAT-R/F results were scored as 0 or 1 (indicating low variability, see Table 1), correlations involving the M-CHAT-R/F were not computed. Divergent validity, defined as the degree to which a measure shows low or non-significant associations with theoretically unrelated constructs (Campbell & Fiske, 1959), was examined between the SCS total score and maternal and paternal ages as markers of unrelated demographic factors. Given the number of correlations examined, p-values were adjusted using the Benjamini–Hochberg false discovery rate (FDR) procedure to control for multiple comparisons. All reported p-values for validity analyses are FDR-adjusted. Classification performance of the SCS To evaluate the SCS’s ability to distinguish toddlers with elevated versus low likelihood of autism, we first applied conventional statistical approaches including the total-score and logistic regression cutoffs. The total-score method summed the seven SCS items, whereas logistic regression allowed differential weighting of each item. To determine optimal cutoffs while limiting overfitting, we used the cutpointr package (Thiele & Hirschfeld, 2021), which estimates robust thresholds through bootstrapping. Following the recommendations of Thiele and Hirschfeld (2021), we adopted the Youden index as our primary performance metric and selected the cutoff that maximized the in-sample Youden index. Using the Youden index facilitates a balanced trade-off between sensitivity and specificity in cutoff selection. In practice, we generated 1,000 outer bootstrap resamples, each with 100 inner replicates, to obtain stable estimates of the cutoff and associated classification metrics, including sensitivity, specificity, accuracy, and the area under the receiver operating characteristic curve (AUC). Following interpretive guidelines (Çorbacıoğlu & Aksel, 2023), AUC values ⩾ 0.90 were classified as excellent, 0.80 to 0.89 as considerable, 0.70 to 0.79 as fair, 0.60 to 0.69 as poor, and <0.60 as fail. For clinical interpretability, the final cutoff was rounded to the nearest integer. Subsequently, RF models were employed for classification (Breiman, 2001; Pedregosa et al., 2011). Two RF models were developed: one using the seven SCS items only and another incorporating child and family demographic variables (age, sex, gestational age, parental ages/education, and household income). To assess potential multicollinearity, we examined pairwise correlations among SCS items and demographic variables and computed the condition number of the feature matrix. Pearson correlations were generally below the commonly used threshold for high intercorrelation (r > .70; Boslaugh, 2012), with only the correlation between maternal and paternal age approaching this level (r = .71; see Supplemental Figure S1). The condition index of the feature matrix was 4.46, below the threshold of 10 (Kim, 2019), indicating negligible multicollinearity. Nonetheless, demographic variables were incorporated only into the RF models and not into the logistic regression model. This decision was made because logistic regression is more sensitive to correlated predictors, which may lead to unstable coefficient estimates, whereas RF models are relatively robust to multicollinearity due to their ensemble-based, nonparametric structure. Accordingly, the inclusion of demographic variables was not expected to adversely affect RF model performance. Models were implemented in scikit-learn (Pedregosa et al., 2011) with nested fivefold cross-validation, a procedure recommended for its high statistical confidence and unbiased estimation of model performance (Ghasemzadeh et al., 2024). Given the class imbalance between the two groups, the Synthetic Minority Oversampling Technique (SMOTE) was applied within the inner training folds to balance the class distribution, ensuring that synthetic samples did not influence the held-out test data (Chawla et al., 2002). Hyperparameter tuning was conducted using a grid search within the inner cross-validation loop, varying the number of trees (100, 300, and 500) and the maximum number of features considered at each split (2–7). The optimized parameters obtained from this process were then applied in the outer loop to fit the testing data. Model robustness was further tested using 1,000 permutation iterations, comparing observed accuracy against chance performance. To understand how individual features contributed to autism-likelihood classification and enhance interpretability, SHAP (SHapley Additive exPlanations) values were computed using the TreeSHAP algorithm (Lundberg et al., 2020), which provides exact feature-attribution values for tree-based models, including RF. SHAP provides a theoretically grounded framework for feature attribution, quantifying the marginal contribution of each feature to model predictions. The SHAP beeswarm plot was used to visualize global feature importance and the direction and magnitude of each feature’s influence on classification outcomes. Sensitivity Analysis Two sensitivity analyses were conducted to evaluate the robustness of the findings. First, given the inclusion of 25 toddlers from an existing intervention cohort to increase the number of elevated-likelihood cases, a sensitivity analysis was performed by excluding these participants. All validity and classification analyses were repeated in the community-only sample to assess whether the main results were influenced by recruitment source or symptom severity differences. Second, because the elevated- and low-likelihood groups differed significantly in child age and household income (see Table 1), a propensity score matching (PSM) analysis was conducted to examine the potential impact of these group differences. Propensity scores were estimated using a logistic regression model with group (elevated vs. low likelihood) as the outcome and child age (continuous) and household income (categorical) as predictors. One-to-one nearest-neighbor matching without replacement was performed using the MatchIt package (Ho et al., 2007). After matching, all validity and classification analyses were repeated using the matched dataset. Results Descriptive Statistics and Differences for SCS Items Between Groups Descriptive statistics for individual SCS items and the total score are presented in Table 2. Independent-samples t-tests were conducted with p-values adjusted using the FDR procedure. The results showed that the low-likelihood group scored significantly lower than the elevated-likelihood group on most SCS items and the total score, with large effect sizes, except for the Requesting item, which showed no significant group difference (p = .81, d = 0.04). Table 2.Descriptive Statistics and Group Differences for SCS Items.IndexItemLow likelihoodM (SD)Elevated likelihoodM (SD) p d 1Spontaneous vocalization2.66 (1.71)4.19 (1.64) <.001 −0.912Pointing2.33 (1.82)3.91 (2.11) <.001 −0.833Eye contact2.13 (1.79)3.57 (1.61) <.001 −0.834Requesting3.08 (1.87)3.00 (2.02)0.810.045Social overtures1.37 (1.84)3.79 (2.28) <.001 −1.236Initiation of joint attention4.18 (1.35)5.32 (1.02) <.001 −0.907Enjoyment in interaction1.88 (1.72)3.66 (1.91) <.001 −1.00Total score17.34 (7.81)26.81 (8.36) <.001 −1.19Note. p-values are based on independent-samples t-tests and are Benjamini–Hochberg false discovery rate (FDR)-adjusted. Bold values indicate statistical significance at p < .05. Cohen’s d effect size interpretation: 0.2 = small, 0.5 = medium, 0.8 = large. Descriptive Statistics and Group Differences for SCS Items. Note. p-values are based on independent-samples t-tests and are Benjamini–Hochberg false discovery rate (FDR)-adjusted. Bold values indicate statistical significance at p < .05. Cohen’s d effect size interpretation: 0.2 = small, 0.5 = medium, 0.8 = large. Reliability As shown in Table 3, the overall ICC for the SCS total score was 0.87 (95% confidence interval [CI] [0.77, 0.93]), indicating good inter-rater consistency. Item-level ICCs ranged from 0.55 to 0.89, reflecting moderate to good agreement across items. The highest reliability was observed for Spontaneous Vocalization and Initiation of Joint Attention, whereas the lowest, though moderate, was found for Eye Contact. Table 3.Inter-Rater Reliability of the SCS Items.IndexItemICC [95% CI]1Spontaneous Vocalization0.89 [0.80, 0.94]2Pointing0.73 [0.54, 0.85]3Eye Contact0.55 [0.29, 0.74]4Requesting0.78 [0.62, 0.88]5Social Overtures0.70 [0.50, 0.83]6Initiation of Joint Attention0.84 [0.72, 0.91]7Enjoyment in Interaction0.71 [0.52, 0.84]Total score0.87 [0.77, 0.93] Inter-Rater Reliability of the SCS Items. Validity Convergent Validity As shown in the Figure 1, significant correlations were observed between the SCS and ADOS-2 total score, r(166) = .51, p < .001; ADOS-2 SA subscore, r(166) = .52, p < .001; and ADOS-2 RRB subscore, r(166) = .30, p < .001. The SCS also correlated positively with parent-reported autism symptoms on the SCQ, r(166) = .26, p < .001, and negatively with VABS-3 Socialization scores, r(160) = −.37, p < .001. In addition, higher SCS scores were associated with reduced communicative competence, as shown by negative correlations with CDI vocabulary production, r(144) = −.22, p = .01, and VABS-3 Communication scores, r(166) = −.48, p < .001. Correlations coefficients between SCS total score and other measures. Divergent Validity No significant associations were found between the SCS and maternal age, r(166) = .14, p = .07, or paternal age, r(166) = .15, p = .07. These results suggested that the SCS total score was not significantly influenced by these variables, indicating good divergent validity. Classification Performance of the SCS Total Score Classification Performance As shown in Figure 2 and Table 4, receiver operating characteristic (ROC) analysis of the SCS total score yielded an AUC of 0.79 (95% CI [0.72, 0.87]), indicating fair discriminatory power. Using Youden’s index, the optimal cutoff score was identified as 23. At this threshold, the sensitivity was 0.72 (95% CI [0.59, 0.86]), specificity was 0.72 (95% CI [0.64, 0.79]), and overall classification accuracy was 0.72 (95% CI [0.65, 0.79]), leaving approximately one quarter of toddlers classified as elevated likelihood by the ADOS-2 would not be identified using this cutoff alone. Receiver operating characteristic (ROC) curves comparing classification performance of the SCS total score and logistic regression-weighted score. Table 4.Model Classification Performance With 95% Confidence Intervals.ModelAUCSensitivitySpecificityAccuracy Conventional statistical model Total score0.79 [0.72, 0.87]0.72 [0.59, 0.86]0.72 [0.64, 0.79]0.72 [0.65, 0.79]Logistic regression score0.90 [0.85, 0.95]0.87 [0.77, 0.96]0.78 [0.70, 0.85]0.80 [0.74, 0.86] Random Forest model Baseline Permutation (null)0.54 [0.43, 0.66]0.57 [0.41, 0.73]0.51 [0.33, 0.69]0.54 [0.43, 0.66] Nested cross-validation SCS only0.87 [0.78, 0.94]0.88 [0.74, 1.00]0.87 [0.73, 0.96]0.87 [0.78, 0.94] SCS + Demographics0.91 [0.86, 0.96]0.91 [0.83, 0.98]0.91 [0.85, 0.96]0.91 [0.86, 0.98]Note. Results for the Random Forest models are reported using nested fivefold cross-validation, with performance metrics derived from the outer test folds. Permutation (null) results represent model performance expected by chance. Model Classification Performance With 95% Confidence Intervals. Note. Results for the Random Forest models are reported using nested fivefold cross-validation, with performance metrics derived from the outer test folds. Permutation (null) results represent model performance expected by chance. Logistic Regression Classification Performance Incorporating item-specific weights derived from a logistic regression model improved classification performance relative to the total-score approach (Figure 2 and Table 4), yielding an AUC of 0.90 (95% CI [0.85, 0.95]). The optimal cutoff for the weighted score was −1.3, at which sensitivity was 0.87 (95% CI [0.77, 0.96]), specificity was 0.78 (95% CI [0.70, 0.85]), and overall accuracy was 0.80 (95% CI [0.74, 0.86]). The final model was specified as follows: Score = −6.56 + 0.48 × Spontaneous Vocalization + 0.24 × Pointing + 0.37 × Eye Contact − 0.65 × Requesting + 0.07 × Social Overtures + 0.64 × Initiation of Joint Attention + 0.31 × Enjoyment in Interaction. RF Classification Performance RF modeling using the seven SCS items alone yielded a slightly lower AUC compared with the logistic regression model but produced more balanced performance across evaluation metrics (all ⩾0.87; Figure 3 and Table 4). When child and family demographic variables were incorporated, performance improved further across all metrics (all ⩾0.91). Classification performance of Random Forest models. Distributions of (A) area under the ROC curve (AUC), (B) accuracy, (C) sensitivity, and (D) specificity across the Random Forest models trained with null (red), SCS items only (blue), and SCS combined with demographic variables (purple). SHAP-Based Feature Importance To further examine the contribution of individual features to classification and enhance interpretability, SHAP analysis was applied to the two RF models. Figure 4 presents the beeswarm plots for the SCS-only model and the SCS + demographics model, illustrating both the relative importance and direction of each feature’s effect on classification. Across both models, Enjoyment in Interaction, Spontaneous Vocalization, and Initiation of Joint Attention consistently emerged as the most influential predictors, as indicated by their larger mean absolute SHAP values and greater dispersion. In contrast, the importance of Requesting varied across models. Notably, Requesting showed only weak correlations with demographic variables (r = −.15 to .13; see Supplemental Figure S1), suggesting that its contribution was largely independent of demographic context. In the SCS + demographics model, child age and household income also emerged as influential features. However, these variables differed significantly between groups (Table 1), indicating that their contributions may reflect underlying group imbalances rather than independent predictive effects. For most SCS items, higher scores, indicating greater social communication difficulty, were associated with an increased likelihood of autism classification. An exception was Requesting, where lower scores (i.e., more frequent and high-quality requesting) were associated with elevated autism likelihood. SHAP beeswarm plot showing feature contributions to the RF models. Sensitivity Analysis Two sensitivity analyses were conducted to examine the robustness of the classification and validity findings. In the first, the 25 toddlers from an intervention cohort were excluded to account for potential influences of recruitment source and differences in symptom severity. Validity and classification performance remained broadly consistent with the full-sample results across all approaches. RF model performance increased in this subsample, reaching near-ceiling levels. This pattern likely reflects reduced clinical heterogeneity following exclusion of the intervention cohort participants, who exhibited more severe and diverse symptom profiles. Their removal resulted in a more homogeneous classification problem that is easier for machine learning models to learn. These findings underscore the importance of sample diversity when interpreting machine learning classification performance. In the second sensitivity analysis, PSM was used to address group differences in child age and household income. Despite balancing these demographic factors, validity and classification performance closely mirrored those of the full sample. While child age and household income were identified as influential in the SHAP analysis of the full sample, they may reflect differences between groups rather than contributing independent predictive value. Detailed results for both sensitivity analyses are provided in the Supplemental Material. Discussion Early and effective identification of autism is a global challenge, particularly in regions with limited access to diagnostic specialists. This study sought to address this gap by evaluating the SCS, a telemedicine-based observational measure, to support early autism identification in Chinese toddlers. Results indicate that the SCS demonstrates sound psychometric properties and shows promise for distinguishing toddlers with elevated versus low likelihood of autism. These findings support the potential extension of the SCS from an outcome measure to a telemedicine-supported screening approach for early autism identification. By leveraging short parent–child interaction videos, the SCS offers a family-centered assessment approach that aligns with global efforts to expand equitable access to early detection assessment (Lord et al., 2022). The reliability analysis revealed that the SCS demonstrated good inter-rater reliability, with an overall ICC of 0.87, confirming consistency in scoring across independent raters. This reliability is comparable to the inter-rater consistency reported for the ADOS-2 total score, ranging from 0.79 to 0.98 (McCrimmon & Rostad, 2014), despite the SCS being a much briefer and asynchronous measure. Consistent with the initial validation study, the coders in this study included doctoral students and research assistants with limited prior autism experience, further highlighting the SCS’s potential to reduce dependence on shorted specialists (Wang, Qi, et al., 2025). At the item level, reliability varied across behaviors, suggesting that some items may not yet be suitable for independent use. “Spontaneous Vocalization” and “Initiation of Joint Attention” achieved the highest ICCs. In contrast, “Eye Contact” yielded the lowest ICC, consistent with previous research showing that gaze-related behaviors are inherently difficult to code from videos due to subtle timing differences, occlusions, and camera angle variability (Venker et al., 2020; Wass et al., 2013). The validity analysis further established the construct coherence of the SCS. The SCS total score was significantly correlated with the ADOS-2 total score and its SA subscore, indicating alignment with gold-standard clinician-administered observations of core autism symptoms. Interestingly, significant correlations were also observed with the ADOS-2 RRB subscore. Although the current SCS does not include items that explicitly target RRBs, this correlation suggests that children exhibiting more pronounced social communication challenges may also display greater RRBs (Chaxiong et al., 2022). Future refinements of the SCS could consider incorporating items that capture RRB-related behaviors to further enhance its diagnostic comprehensiveness. Moreover, significant correlations between the SCS and parent-reported measures, including the SCQ, CDI vocabulary production, and the VABS-3 Socialization and Communication domains, further support the SCS’s ecological validity. These convergent patterns mirror findings from the initial SCS validation in an older cohort of autistic children (i.e., 24–60 months; Wang, Qi, et al., 2025). Across different age ranges of young children, the findings suggest that the SCS captures both observed social communication behaviors and parent-reported variations in everyday functioning. Importantly, these validity findings remained broadly unchanged across both sensitivity analyses, suggesting that they were not driven by recruitment source, symptom severity differences, or demographic imbalances between groups. This study also evaluated the potential of the SCS as a screening measure for differentiating toddlers with elevated versus low likelihood of autism, as defined by ADOS-2 classifications. When trained using the seven SCS behavioral items, both logistic regression and RF model achieved higher evaluation metrics, compared with the total-score method. The superior performance of the two models is likely attributable to their ability to assign differential weights to individual behavioral features. Unlike conventional total scoring methods that treat all items as equally weighted, these models estimate the relative contribution of each feature based on its predictive value (Y. Liu & Zhao, 2017). Consistent with this, the regression coefficients and SHAP analysis indicated substantial variability in the classification relevance of the seven SCS items. These findings suggest that early social communication behaviors are not uniformly informative for autism identification. Instead, specific domains, such as Enjoyment in Interaction, Spontaneous Vocalization, and Initiation of Joint Attention, appear to carry greater discriminative value than others. From a developmental perspective, this pattern is consistent with the central role these behaviors play during the second year of life, a period characterized by rapid growth in social communication abilities (Carpenter et al., 1998). Enjoyment in Interaction reflects a child’s motivation to engage socially with others and is considered an important foundation for later social development (Gampe et al., 2025). Spontaneous Vocalization, particularly when directed toward another person, reflects emerging communicative intentionality and is closely associated with early language development (Watt et al., 2006). Initiation of Joint Attention requires children to actively coordinate attention between people and objects for the purpose of sharing experiences, a developmental achievement that has been repeatedly identified as an early indicator of autism-related social communication differences (Charman et al., 1997; Mundy et al., 2007; Mundy & Newell, 2007). These behaviors together reflect the development of social motivation, communicative intentionality, and shared attention, capacities that undergo substantial change during toddlerhood and may therefore be particularly informative indicators of elevated autism likelihood in this age range. An interesting exception was the Requesting item. Although Requesting did not significantly differentiate the elevated- and low-likelihood groups in univariate analyses, it contributed to classification in both the logistic regression and RF models. This pattern may reflect the distinct developmental function of requesting behaviors. Unlike joint attention or shared enjoyment, requesting is primarily directed toward obtaining desired objects, activities, or assistance rather than sharing experiences with others (Bates et al., 1975). As a result, requesting behaviors may be less sensitive to autism-related social communication differences during toddlerhood and may remain relatively intact in some children with elevated autism likelihood, particularly when interacting with their parents. At the same time, the way requesting occurs in relation to other social behaviors may still carry meaningful information. For example, a toddler who requests frequently but rarely initiates joint attention, shares enjoyment, or engages in reciprocal interaction may present a different developmental profile from a toddler whose requesting occurs within richer social exchanges. Logistic regression and RF models are able to evaluate these behaviors jointly rather than in isolation, allowing them to capture patterns of co-occurrence that may not emerge in univariate group comparisons. Thus, while Requesting may have limited value as a stand-alone indicator, it may nevertheless contribute useful contextual information when considered alongside other social communication behaviors. Compared with logistic regression, the SCS-only RF model produced more balanced performance across evaluation metrics (all ⩾0.87), particularly in terms of specificity. The higher specificity achieved by RF (0.87 vs. 0.78 for logistic regression) indicates fewer false positives, which is especially important in screening contexts where over-identification may lead to unnecessary use of resources and increased burden and stress for families. This more balanced performance may reflect the ability of RF models to capture complex, nonlinear relationships and interactions among features, which are not explicitly modeled in standard logistic regression (Lin et al., 2023). From a clinical perspective, the findings suggest that different analytic approaches may offer distinct advantages within autism screening pathways. Logistic regression retains the benefits of transparency, interpretability, and ease of implementation, making it potentially well suited for routine screening applications where straightforward risk estimation is desirable. In contrast, the RF findings suggest that more complex modeling approaches may provide additional information that could assist decision-making in cases where risk classification is less clear. However, these results should be interpreted cautiously. The observed performance reflects internal validation within the present dataset and does not establish superiority in real-world screening settings. Further external validation in independent and more diverse samples is required before conclusions can be drawn regarding the clinical utility or implementation of RF-based screening approaches. At present, the findings support the potential value of machine learning approaches as complementary tools for enhancing autism screening rather than as replacements for more transparent and interpretable methods. Incorporating child and family demographic characteristics into the RF model further improved classification performance, yielding higher AUC, sensitivity, specificity, and overall accuracy. This suggests that contextual factors may provide additional information that complements SCS items in distinguishing autism likelihood. However, this improvement should be interpreted with caution. As indicated by the SHAP analysis and sensitivity analyses, some demographic variables (e.g., child age and household income) differed between groups and may therefore have contributed to classification through their association with group membership rather than independent predictive value. Consistent with this interpretation, when demographic imbalances were addressed through PSM, classification performance remained stable, and the SCS items alone retained strong discriminative ability. These findings suggest that SCS items are the primary drivers of classification performance, whereas demographic variables may provide supplementary contextual information. Future studies should evaluate the added value of demographic features in larger and more diverse independent samples to determine whether these effects generalize beyond the present dataset. The classification performance observed in the current sample aligns with the highest benchmarks reported for video-assisted telemedicine technologies, where machine learning models have reached pooled sensitivities of 0.93 and specificities of 0.83 (Wang, Meng, et al., 2025), as well as with established toddler screening tools (Sánchez-García et al., 2019). A key strength of this study lies in its rigorous methodological and analytical approach to classification. To strengthen the robustness of the RF modeling results, we first implemented nested fivefold cross-validation to separate hyperparameter tuning (inner loop) from evaluation on held-out data (outer loop), providing a more reliable estimate of generalization performance and reducing the risk of overfitting. In addition, SHAP analyses were applied to enhance interpretability by quantifying the contribution of each feature to model predictions. These approaches enhance the transparency and reliability of the findings, addressing common concerns regarding interpretability and overfitting in machine learning models (Heinrichs & Eickhoff, 2020). A second strength lies in the development and validation of the SCS within a culturally and geographically specific context outside of Western populations. While existing international telemedicine instruments for autism screening and diagnosis like the TAP (formerly Telehealth Autism Spectrum Disorder-Pediatrics [TELE-ASD-PEDS]; Corona et al., 2020), the Naturalistic Observation Diagnostic Assessment (NODA; Smith et al., 2017), and the Screening Tool for Autism in Toddlers and Young Children (STAT; Juárez et al., 2018) have demonstrated efficacy, their overwhelming reliance on Western samples highlights the need for the development of psychometrically validated instruments for non-Western cultural contexts (Wang, Meng, et al., 2025). The video-based observational format of the SCS may be particularly useful in this context. Unlike parent-report questionnaires, the SCS enables trained coders to directly evaluate toddlers’ social communication behaviors during brief parent-child interactions. This may complement caregiver-reported screening, especially in settings where parents’ understanding of autism-related behaviors may be limited or influenced by stigma. Previous studies have reported limited autism-related knowledge among Chinese parents (Li et al., 2025; Su et al., 2021), as well as stigma that may affect caregiver reporting (Yu et al., 2020). Consistent with these concerns, among the 145 parents in the current sample who completed the M-CHAT-R/F, 23 toddlers were classified as having elevated autism likelihood based on the ADOS-2, but only eight of these children (34.78%) screened positive on the M-CHAT-R/F. This finding suggests that reliance on caregiver-report screening alone may miss some toddlers with elevated autism likelihood. Accordingly, the SCS may be best positioned as a complementary component within a broader stepped screening pathway rather than as a stand-alone diagnostic tool. Used alongside parent-report questionnaires, the SCS could provide an additional behavioral observation layer to inform decisions regarding developmental surveillance, comprehensive assessment, and specialist referral. In community child health settings, brief parent–child interaction tasks could potentially be incorporated into routine developmental checkups, for example, during waiting periods for physical examination results or other health reports. Such an approach may allow initial behavioral information to be collected with relatively low burden on families and frontline providers. Nevertheless, further implementation research is needed to examine the feasibility, acceptability, workflow integration, and referral outcomes of incorporating the SCS into routine community screening services. Despite these strengths, several limitations should also be considered when interpreting the present findings. First, the inclusion of additional participants from an existing intervention cohort may introduce sampling bias. Children in this subgroup were older and demonstrated greater symptom severity compared with those recruited from the community. While this inclusion allowed the sample to better capture the full spectrum of autism symptom severity, it may also have inflated classification performance. To address this concern, sensitivity analyses were conducted, including exclusion of this subgroup and matching on key demographic variables, with results showing a consistent pattern of findings. Nevertheless, future studies should prioritize recruiting more balanced and population-representative samples to minimize potential confounding at the design stage. Second, although the present sample represents one of the larger datasets in telemedicine-based autism screening research, particularly within Chinese populations, it remains modest for machine learning applications. Models such as RF typically require larger and more heterogeneous datasets to ensure model stability, reduce overfitting, and enhance generalizability. In addition, model evaluation relied on nested cross-validation, which provides a robust internal estimate of performance but does not constitute true external validation. Future work should therefore prioritize multisite data collection and independent validation cohorts to further evaluate the generalizability of the SCS and associated classification models. Third, the relatively broad developmental range (16–30 months) introduces potential heterogeneity, given the rapid changes in social communication and language development during this period. Screening performance may vary across younger and older toddlers, which could not be fully examined in the current sample due to the limited and uneven distribution of participants in the elevated-likelihood group. Future studies with larger samples should investigate age-specific performance or stratified models to better account for developmental effects. Fourth, the SCS and ADOS-2 both assess observable social communication behaviors associated with autism, raising the possibility of shared method variance. This overlap may partially contribute to the observed classification performance, as both measures rely on similar behavioral constructs. Future research should incorporate additional independent criteria, such as clinical diagnoses to further establish the discriminant validity of the SCS. Fifth, although the SCS protocol provides general guidance on the types of toys used during assessment, specific materials were not standardized across families to preserve ecological validity and feasibility in home settings. While the coding scheme targets core social communication behaviors expected to generalize across contexts, variability in materials may introduce unquantified heterogeneity. Future research could systematically examine the trade-off between standardization and ecological validity to optimize measurement consistency. Finally, this study focused on concurrent classification based on ADOS-2-defined likelihood categories rather than longitudinal diagnostic outcomes. As such, the findings demonstrate screening performance at a single time point but do not establish predictive validity for later clinical diagnoses. Longitudinal follow-up studies are needed to determine whether early SCS-based classifications reliably predict diagnostic outcomes and developmental trajectories over time. Conclusion This study provides evidence supporting the reliability and validity of the SCS as a telemedicine-supported screening approach for identifying toddlers with elevated versus low likelihood of autism in a Chinese sample. Integrating brief parent–child interaction videos with rigorously validated classification approaches, the SCS demonstrated good psychometric properties and promising classification performance within the present dataset. These findings suggest that the SCS may serve as a practical and family-centered complement to existing early autism screening methods, particularly in settings where access to specialist services is limited. Future work should prioritize external validation in larger and more diverse populations, evaluate against independent clinical diagnoses, and explore the integration of automated video coding and AI-assisted analysis to further enhance precision and efficiency. The SCS therefore represents an important step toward more accessible and culturally sensitive early autism detection in China and other resource-limited contexts. Supplemental Material sj-docx-1-aut-10.1177_13623613261480732 – Supplemental material for From Clinic to Home: Validating the Social Communication Scale for Telemedicine-Supported Early Autism Screening in Chinese Toddlers Supplemental material, sj-docx-1-aut-10.1177_13623613261480732 for From Clinic to Home: Validating the Social Communication Scale for Telemedicine-Supported Early Autism Screening in Chinese Toddlers by Li Wang, Yujia Shi, Fei Yuan, Xiujuan Geng, Ziyan Meng, Yipeng Wang, Hon-Cheong So and Patrick C. M. Wong in Autism