{"paper_id":"445e5685-b5ec-48cb-8bfd-2db2e8824e25","body_text":"Large-Scale Assessment of Language, Speech, and Movement in Autism and ADHD with AI | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return\"[object Function]\"==o.call(a)}function e(a){return\"string\"==typeof a}function f(){}function g(a){return!a||\"loaded\"==a||\"complete\"==a||\"uninitialized\"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){(\"c\"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){\"img\"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),\"object\"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height=\"0\",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),\"img\"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||\"j\",e(a)?i(\"c\"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName(\"script\")[0],o={}.toString,p=[],q=0,r=\"MozAppearance\"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&\"[object Opera]\"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?\"object\":l?\"script\":\"img\",v=l?\"script\":u,w=Array.isArray||function(a){return\"[object Array]\"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split(\"!\"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split(\"=\"),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(\".\").pop().split(\"?\").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split(\"/\").pop().split(\"?\")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&\"css\"==i.url.split(\".\").pop().split(\"?\").shift()?\"c\":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Large-Scale Assessment of Language, Speech, and Movement in Autism and ADHD with AI View ORCID Profile Aimar Silvan , View ORCID Profile Lucas C. Parra , Adriana Di Martino , Michael P. Milham , View ORCID Profile Jens Madsen doi: https://doi.org/10.1101/2025.10.20.682864 Aimar Silvan 1 Department of Biomedical Engineering, City College of New York , New York, NY, 10031, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Aimar Silvan Lucas C. Parra 1 Department of Biomedical Engineering, City College of New York , New York, NY, 10031, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Lucas C. Parra For correspondence: parra{at}ccny.cuny.edu Adriana Di Martino 2 Child Mind Institute , New York, NY, 10022, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Michael P. Milham 2 Child Mind Institute , New York, NY, 10022, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jens Madsen 1 Department of Biomedical Engineering, City College of New York , New York, NY, 10031, USA Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Jens Madsen Abstract Full Text Info/History Metrics Preview PDF Abstract Attention-Deficit/Hyperactivity Disorder (ADHD) and Autism Spectrum Disorder (ASD) frequently co-occur with overlapping symptoms. We investigated whether automated behavioral analysis during a clinician-child interview can identify distinct, objective features of the two disorders. Analyzing audio-video recordings of 2,341 youths (ages 5–22) in a broad community sample, multivariate models revealed that language difficulties often attributed to ADHD are primarily explained by age, cognitive ability, or co-occurring ASD. Increased motor activity specifically marked hyperactive-impulsive ADHD, but not ASD or inattentive ADHD. ASD was uniquely characterized by divergent narrative production and perspective-taking, alongside a distinct vocal profile of higher pitch and intensity, despite structurally intact language. While these digital behavioral measures correlate with most diagnostic categories and age, the joint analysis effectively separates the effects of ASD from ADHD. These findings show that scalable digital assessment from recorded clinical interviews can disentangle overlapping ASD and ADHD diagnoses into domain-specific behavioral signatures. Introduction Attention-Deficit/Hyperactivity Disorder (ADHD) and Autism Spectrum Disorder (ASD) are prevalent conditions with long-term functional impacts 1 – 4 . ADHD involves persistent inattention and/or hyperactivity, while ASD is characterized by challenges with social interaction and restricted, repetitive behaviors and interests 5 . These heterogeneous conditions frequently co-occur 6 – 12 , sharing genetic factors 13 – 16 and converging variations in large-scale brain networks 17 – 22 . Current diagnostic assessments for ASD and ADHD rely on clinical observation and reports. The clinical visit is an opportunity to observe and assess patients’ behaviors in a social context. Here, we test whether large-scale, automated analysis of naturalistic clinical interactions can disentangle overlapping ADHD and ASD diagnosis into separable behavioral dimensions. We leveraged the Healthy Brain Network (HBN) cohort 23 , a large community sample of neurotypical and neurodivergent youth. The data included a video-recorded, semi-structured interview where participants were asked about an emotive animated film they just watched. This easily scalable interview was designed to elicit language, emotional expression, and social-cognitive functions like perspective-taking. State-of-the-art AI enables automated assessment of these audio-video recordings across three domains: gross body movements, language use, and manner of speech. In the following, we review relevant literature in these domains in the context of ADHD and ASD. Effective communication depends on both structural language (vocabulary, syntax) and pragmatic language skills (social application, topic coherence) 24 . Social communication deficits are a core diagnostic feature of ASD but not ADHD 5 . However, studies using informant-based assessments suggest a significant degree of overlap, with both groups demonstrating difficulties in both pragmatic and structural language 25 – 27 . Beyond fundamental language skills, perspective-taking is also essential for discourse and narrative comprehension and construction 28 . This is often described as Theory of Mind (ToM), i.e., the ability to infer the mental and emotional states of others 29 . Difficulties in perspective-taking have long been associated with ASD 30 . Studies using automated language analysis have identified repetitive speech and atypical narratives in ASD 28 , 31 , and narrative disorganization in ADHD 32 . However, these methods typically evaluate fewer than 50 participants 34 and analyze conditions separately, limiting generalizability and obscuring shared versus distinct traits. We address these gaps by jointly modeling structural, pragmatic, and higher-order socio-cognitive abilities across both diagnoses to develop a detailed model of traits in both ADHD and ASD on an unprecedented scale. Alongside communication deficits, atypical prosody is a clinical hallmark of ASD 31 . Recent systematic reviews link ASD to higher mean pitch and greater pitch variability 33 , though findings regarding intensity, speech rate, and voice quality remain inconsistent due to methodological heterogeneity and small sample sizes 33 , 34 . Furthermore, the influence of co-occurring ADHD, which may also present distinct prosodic features 35 , 36 , remains poorly understood. While automated vocal analysis shows diagnostic promise 37 – 39 , constrained samples currently prevent the isolation of robust, condition-specific vocal signatures for either group. Motor differences are central to both ADHD and ASD, yet their behavioral signatures are rarely quantified under comparable conditions. While ASD has been associated with diverse motor patterns 40 , 41 , it remains unclear which of these reflect distinct motor phenotypes or are shared manifestations of ADHD. Further complicating matters, language, speech, and motor coordination develop rapidly with age, and symptom presentations can shift 43 . For instance, hyperactivity in ADHD often decreases while inattention persists into adolescence 44 . Without rigorously modeling key covariates like age, sex 45 , 46 , and cognitive ability (IQ) 47 , 48 , observed markers may simply reflect typical developmental or demographic variance rather than true psychopathology. This first analysis of the HBN semi-structured interview presents an opportunity to simultaneously observe these diverse behavioral domains at an unprecedented scale. Using gold-standard clinical diagnoses, we apply a multivariate approach to identify independent drivers of behavioral variance. Deviating from prior work that treats co-occurring ADHD and ASD as a distinct categorical group 49 – 51 , we model ADHD Inattention, ADHD Hyperactivity, and ASD as independent clinical dimensions. By mapping digital phenotypes onto partially overlapping diagnostic dimensions rather than mutually exclusive categories, our findings help reconcile prior inconsistencies in the literature, and isolate condition-specific behavioral signatures. Results A semi-structured interview of a community sample for automatic behavioral analysis To measure naturalistic behaviors, 2,341 participants (aged 5-22) of the HBN cohort were video-recorded during a semi-structured interview after watching \"The Present\", a brief animated film in which a boy’s interaction with a three-legged puppy culminates in the revelation that both share a missing leg. During the interview, participants were asked identical questions one at a time ( Table S1 ) probing narrative recall, factual memory, emotional interpretation, and perspective-taking. We used state-of-the-art computational tools to derive objective, quantitative metrics from the video-recorded interviews, covering expressive language, vocal prosody, as well as facial and body movements ( Fig. 1A ). Download figure Open in new tab Figure 1. Overview of the study cohort. A) Using AI for objective behavioral characterization. Video recordings were processed to quantify body and facial movements, while audio was transcribed and analyzed to extract metrics of language structure, semantic content, and vocal prosody. B) Venn diagram showing the distribution and overlap of the primary diagnostic groups. C) The 12 most common clinician-assigned DSM-5 diagnoses in the full cohort (N=2,341), see Fig. S1 for specific co-occurring diagnoses for the ADHD and ASD children. D) Age and sex distribution of the full cohort, see Fig. S1 for age-sex distributions per diagnostic group in panel B. E) Sex distribution within the children with ASD and/or ADHD diagnostic, ADHD-Combined are here considered positive in both hyperactive and inattentive diagnostic status. The black line indicates chance occurrence based on the sex ratio in the cohort. F) Adjusted odds ratios showing the likelihood of a co-occurring positive diagnosis in individuals with positive specific predictors (or one-year change in the case of age). Females are less likely than males to receive an ADHD or ASD diagnosis (OR < 1). G) Spearman correlation matrix, displaying only significant associations (p<0.01, Bonferroni corrected, white otherwise). The matrix illustrates that clinical instruments for specific neurodevelopmental traits, like ASSQ for autism and SWAN for ADHD, are highly correlated with other measurements like depression (MFQ), anxiety (SCARED), and internalizing/externalizing behaviors (CBCL). They also correlate with age, IQ, and sex, which complicates the attribution of observed behavior to any single instrument scale independently. H) Two-dimensional histogram showing the negative correlation between age and hyperactivity traits (SWAN scale) (ρ(2194) = -0.28, p<0.001). Full names for clinical instruments (SWAN, ASSQ, MFQ, SCARED, CBCL) are provided in the ‘Clinical Instruments’ section in the Methods. Detailed model statistics for panel F are available in Supplementary Section S1.1 . The cohort was diagnostically rich and clinically complex ( Fig. 1B, C & Table 1 ). The sample exhibited an overall male-to-female ratio of approximately 2:1 ( Fig. 1D ), with lower than chance prevalence in females (17% of the ASD group and 23-27% of the ADHD group; Fig. 1E ). Consistent with the clinical literature 52 , a clinician-confirmed diagnostic status of either ADHD-Inattentive or ADHD-Hyperactive significantly increased the odds of co-occurring ASD (Adjusted Odds Ratios (OR) = 2.3 and 1.6, respectively, Fig. 1F ). Clinical overlap extended to instrument ratings, which were highly correlated and strongly influenced by age and sex ( Fig. 1G ). For instance, parent-rated hyperactivity traits decreased significantly with age (ρ(2194) = -0.28, p<0.001, Bonferroni corrected, Fig. 1H ), consistent with typical developmental trajectories 44 . The same was also reflected in the effect of age on the prevalence of ADHD-Hyperactive status ( Fig. 1F , OR = 0.83). Such a change with development is to be expected for language abilities 53 , which are a primary focus of this study. View this table: View inline View popup Download powerpoint Table 1. Demographic, clinical, and interview characteristics Note that in Fig. 1E-F and in all our analyses, ADHD-Combined presentation is coded as both Inattentive and Hyperactive. To distinguish this coding with two binary variables (Inattentive, Hyperactive) from the three DSM-5 ADHD presentations (Inattentive, Hyperactive, Combined), we will refer to these as diagnostic status and presentation , respectively. Developmental factors dominate over diagnostic effects in structural language To characterize naturalistic communication at scale, we processed the audio recordings of the interview using a state-of-the-art AI pipeline that integrates automatic transcription with Large Language Model (LLM)-based diarization. From these transcripts, we extracted objective metrics of structural and pragmatic language abilities ( Fig. 2A ). Download figure Open in new tab Figure 2. Language abilities are primarily driven by development, not by ADHD or ASD. A) The computational pipeline used to extract objective structural and pragmatic language metrics from transcribed audio. B) Hyperactive traits (SWAN) and lexical diversity (Moving-Average Type-Token-Ratio with a window of 10) are correlated (ρ(2160)=-0.16, p<0.001). However, the color gradient, representing age, reveals that this association is confounded by development. C) Conceptual diagram of the multivariate regression model used to isolate the unique effects of each predictor. A “Univariate Model” does not account for covariation, while the “Multivariate Model” controls for shared variance, with Age and Sex as a common source of variance. Arrows are color-coded to the effect size magnitude in the multivariate regression model (panel D). Univariate associations between DSM-5 diagnoses and demographics with lexical diversity (grayed arrows) are instead better explained by age and IQ in the multivariate model. Black arrows indicate effects tested in Fig. 1F . (See Supplementary Section S1.1 ). D) Effect sizes from the multivariate model predicting lexical diversity. Effect sizes represent the unique contribution of each factor (Sex coded as Female = 1, Male = 0). For binary variables, the reported effects are Cohen’s d, while for numerical variables, we report Cohen’s f (see Methods). E) A summary matrix of effect sizes from the multivariate model applied to a range of language metrics. The results show the consistent and strong effects of age and IQ, and the general lack of diagnostic effects across all measured aspects of language abilities. Significance cut-off at p < 0.01 (white otherwise; Bonferroni-corrected across the 9 models tested). Detailed multivariate model reports are available in Supplementary Section S1.2 . To isolate the unique behavioral contributions of each diagnosis while accounting for high co-occurrence and developmental maturation, we employed a multivariate regression model. This approach coded diagnostic status as independent binary variables (ASD, ADHD-Inattentive, ADHD-Hyperactive) while simultaneously controlling for Age, Sex, and IQ ( Fig. 2C ). This multivariate approach demonstrated that structural language abilities are overwhelmingly driven by development, not by ADHD or ASD. For example, a simple univariate analysis suggested a significant negative association between lexical diversity and hyperactivity (ρ(2160)=-0.16, p<0.001, Bonferroni Corrected, Fig. 2B ). However, the multivariate model revealed this variation was actually driven by the effect of Age (Effect Size=0.489; effect sizes (ES) reported in this study represent Cohen’s d for binary and Cohen’s f for continuous variables; see Methods) and IQ (ES=0.273) ( Fig. 2C-D ). This pattern held across almost all structural language metrics ( Fig. 2E ). Older participants spoke more (ES=0.389) and faster (ES=0.563), while females spoke faster (ES=0.165) and answered more coherently (ES=0.189) than males. The only significant diagnostic marker identified in basic language use was an increase in self-referential speech (first-person pronouns) specific to participants with ASD (ES=0.278). Our planned analysis was to model ADHD and ASD as independent additive effects. The alternative approach of modeling the co-occurrence as a separate diagnostic category (ADHD+ASD) 49 – 51 is treated in the Methods under “Alternative analysis approaches” along with Age interactions and other post hoc analyses. ASD is uniquely associated with atypical narrative, story comprehension, and perspective taking While structural language remained largely intact across diagnoses, impairments in social communication often stem from higher-order difficulties in narrative construction and perspective-taking (Theory of Mind). To objectively quantify these functions, we computationally compared the semantic content of each participant’s interview responses to those of their age-matched, Typically Developing peers, generating an objective measure of \"answer typicality\" ( Fig. 3A-C ). Download figure Open in new tab Figure 3. Semantic analysis reveals ASD-specific effects on narrative and social-cognitive abilities. A) The computational pipeline for semantic analysis. Participants’ answers to 23 predefined interview questions were parsed and converted into numerical semantic representations (embeddings). B) Multidimensional Scaling (MDS) representation of the answer embeddings, where similar embeddings are represented closer together. A \"typical answer\" is obtained from the semantic (cosine) similarity to the median embedding of the corresponding age-matched TD group answers. C) Semantic similarity to the age-appropriate prototype is used as a measure of ‘typicality’. For the narrative recall question in panel A, higher similarity reflects a more detailed and developmentally expected recount of the narrative relative to age-matched peers. Subject 1 provides a detailed answer (high similarity), whereas Subject 3 does not recall the story (low similarity). D) Joint distribution illustrating the positive association between age and the production of more typical narratives, but the delayed development in the ASD group compared to the non-ASD group. E) Matrix of effect sizes from the multivariate model, assessing the impact of diagnosis and demographic factors on various social-cognitive domains probed by the interview questions (e.g., factual memory, emotional description, and thematic understanding). Only effect sizes with p<0.01, after Bonferroni correction, are shown in color. See Table S1 for a full list of questions and the typical answers for each age bin, and Supplementary Section S1.3 for detailed multivariate model reports. F) Scatterplot of terms distinguishing ASD (top, purple) from non-ASD (bottom, pink) participants in responses to the question related to thematic understanding, using scattertext 57 (See Methods section ‘Visualization of Distinctive Terms’ for details, and Fig. S2 for visualization of other questions). After controlling for Age, our models revealed a significant, divergent socio-cognitive profile specific to ASD ( Fig. 3E ). An ASD diagnosis was uniquely associated with producing less typical narratives (ES=-0.299, Fig. 3D ), showing different thematic understanding -- such as failing to convey that the boy and the puppy were similar because they were both missing a leg ( Fig. 3F ) -- (ES=-0.266), providing idiosyncratic descriptions of their own emotions (ES=-0.272), and expressing atypical preferences (ES=-0.313). In contrast, neither the ADHD-Inattentive nor ADHD-Hyperactive statuses were significantly associated with any of these social-cognitive domains. Despite measuring similarity of responses relative to age-matched controls, age remained a powerful predictor across some semantic language use measures ( Fig. 3E ). This pattern is consistent with increasing convergence towards age-normative responses, rather than a drift in the semantics of the responses. Indeed, the exemplary ‘typical answers’ for each age bin were remarkably similar ( Table S1 ), suggesting no age-related difficulties in understanding or answering the questions within the Typically Developing population. Consistent with social-cognitive maturation 43 , older participants provided more consistent, detailed narratives (ES=0.394) and recalled the movie’s main message more typically (ES=0.345). Similarly, higher IQ was associated with more typical narrative abilities (ES=0.185), thematic understanding (ES=0.212), and factual memory (ES=0.153). We also observed a secondary finding related to sex, with females tending to produce more typical narratives (ES=0.223), and positive and negative preferences (ES=0.264 & 0.180) compared to male participants ( Fig. 3E ). Autism Spectrum Disorder has a distinct vocal profile The acoustic patterns of pitch, intensity, and rhythm are critical for conveying the intent of an utterance. To this end, we employed an automated pipeline to extract a comprehensive set of acoustic features from the segmented speech of each child, including measures of pitch, loudness, and voice quality ( Fig. 4A ). Download figure Open in new tab Figure 4. Automatic vocal prosody metrics reveal an ASD-specific acoustic profile. A) Automatic extraction of vocal prosody features involved transcribing and diarizing speech, thus isolating the children’s speech from that of the adult interviewer, and extracting audio-based features using automatic acoustic analysis. 58 B) ASD participants showed increased dysphonia compared to non-ASD participants, which may be related to an “unusual” voice in ASD. C) Effect Size matrix illustrating the effects of development and ASD on vocal prosody. Effect Sizes in this matrix are estimated again using a multivariate regression model, with p<0.01, after Bonferroni correction. See Supplementary Section S1.4 for detailed multivariate model reports. Our automated acoustic analysis revealed a distinct vocal signature specific to ASD. Controlling for covariates, ASD was significantly associated with increased mean pitch (ES=0.222), greater pitch variability (ES=0.261), and higher vocal loudness (ES=0.231) ( Fig. 4C ). Furthermore, participants with ASD exhibited higher dysphonia (ES=0.371, Fig. 4B ) and increased breathiness (ES=0.325). Notably, this vocal profile was highly specific: neither ADHD status was associated with altered vocal prosody. Age again proved to be a dominant predictor, with older participants naturally exhibiting reduced pitch (ES=-0.419), lower pitch/loudness variability (ES=-0.415 & -0.140), and decreased breathiness (ES=-0.104). Increased motor activity is uniquely associated with ADHD-Hyperactive diagnostic status Motor behaviors are relevant diagnostic features of ASD and ADHD, with hyperactivity being a key diagnostic criterion in ADHD, and stereotypical and repetitive motor behaviors being among the defining criteria for ASD. Yet motor behaviors are rarely quantified objectively in naturalistic settings. Here, we applied computer vision models to extract objective frame-to-frame displacement metrics for the upper body, head, eyes, and face ( Fig. 5A ). Consistent with typical maturation, age was a strong negative predictor of movement categories (ES ranging from -0.461 to -0.308). Download figure Open in new tab Figure 5. Associations of demographics and phenotype with body and face movements. A) We use Google’s Mediapipe computer vision models to obtain framewise locations for anatomical and facial landmarks. We then calculate the average framewise displacement for the different regions of interest. B) ADHD-Hyperactive youth show increased facial movements across the whole developmental spectrum, compared to non-ADHD Hyperactive. C) Effect Size matrix illustrating the effects of development and an ADHD-Hyperactive status on all measures of face and body movements. Effect Sizes in this matrix are estimated using multivariate regression, with p<0.01, after Bonferroni correction. See Supplementary Section S1.5 for detailed multivariate model reports. Once developmental decreases in movement were accounted for, we identified a highly specific motor signature: ADHD-Hyperactive status was robustly associated with increased movement across all measured body and facial regions (ES=0.225-0.301) ( Fig. 5B-C ). Conversely, neither ASD nor ADHD-Inattentive status were linked to significant changes in motor activity, confirming that objective video-derived kinematics isolate the hyperactive-impulsive domain of ADHD without generalizing to the inattentive presentation or ASD. Discussion This study applied automated digital phenotyping to a large community sample to isolate behavioral dimensions of ASD and ADHD, demonstrating that while narrative divergence and atypical speech uniquely characterize ASD, increased motor activity specifically marks ADHD-Hyperactivity. Crucially, language deficits often attributed to ADHD are fully explained by age, IQ, and co-occurring ASD. Previous standardized assessments suggest ADHD involves structural and pragmatic language impairments 59 , 60 , with community prevalence estimated at 40-45% 61 , 62 . However, standardized tests impose heavy demands on executive functions like sustained attention and working memory. Consistent with other naturalistic studies 63 our conversational analysis shows that once age, IQ, and ASD are controlled, structural, conversational language in ADHD is intact. This extends to higher-order social cognition. After accounting for age, IQ, and ASD status, ADHD was not independently associated with narrative ability or perspective-taking (ToM), indicating that previously reported differences may reflect developmental factors or diagnostic overlap rather than ADHD-specific effects reported previously 64 , 65 . Apparent communication difficulties in ADHD likely reflect executive dysfunction, such as poor inhibition, impacting speech organization, rather than fundamental deficits in linguistic comprehension or mentalizing 66 , 67 . Impaired ToM thus remains a critical differentiator indicating co-occurring ASD in youths with ADHD 68 . Conversely, our ASD cohort exhibited mostly typical structural language trajectories, supporting classic conceptualizations of the disorder as a core social communication deficit 5 . The primary structural deviation was increased self-referential speech via first-person pronouns, reflecting an asynchrony between linguistic and social development 69 . ASD communication deficits emerged strictly in higher-order domains: participants produced less typical narratives, missed overarching thematic ’gists’ 70 , and provided idiosyncratic emotion descriptions 71 . This naturalistic validation aligns with computational language analyses demonstrating diminished narrative coherence in demanding or emotive contexts for ASD 72 , 73 . Furthermore, we objectively quantified the long-observed atypical prosody of ASD 34 . Our analysis confirmed higher pitch and pitch variability 33 , while clarifying past inconsistencies by identifying increased vocal intensity, dysphonia, and breathiness as distinct, ADHD-independent ASD markers, validating clinical descriptions of \"hoarse\" voices 74 . Most importantly, these associations were absent in ADHD. Regarding motor behavior, unobtrusive computer vision effectively captured ADHD-related restlessness that typically requires specialized sensors 42 , 75 . Accounting for age-related decreases, widespread head, body, and facial movement specifically marked the ADHD-Hyperactive status. Neither ADHD-Inattentive nor ASD diagnoses affected the global displacement metrics we used here. Because motor challenges are core to ASD, 41 the absence of an ASD-associated signal in these global kinematic measures suggest that they may not capture the more specific, repetitive movement patterns often described in autism 40 which may require a more granular movement classification 40 , 76 – 81 . Developmental and demographic factors proved essential to the behavioral phenotype, with age robustly driving changes across motor, linguistic, and socio-cognitive domains alongside sex-specific profiles. Digital biomarkers must rigorously disentangle normative trajectories from pathology. Although cognitive ability varies within ASD, 47 , 48 and influences language 31 , 82 , and IQ remains controversial as a pure covariate 83 , our primary diagnostic associations remained robust whether Full-Scale or Verbal IQ were included or excluded (see Fig. S3 ). While our automated features correlate strongly with established clinical instruments (see Fig. S4 ), they can offer enhanced presentation-specificity. However, these computational tools are not intended to replace current diagnostic methods, as traditional instruments still align more closely with categorical DSM-5 definitions (see Fig. S5 ). As a cross-sectional study, we could identify developmental trends but could not establish causality or track individual trajectories over time. Furthermore, our cohort primarily included individuals with mild-to-moderate ASD traits (see Fig. S6 ). This likely explains why manual inspections revealed no overt behavioral stereotypies, our analysis focused on global kinematics instead. Future work on ADHD and ASD should validate these brief, video-based measures longitudinally using wearable sensors and apply them within standardized assessments like the ADOS 84 . More broadly, this work illustrates how brief, standardized clinical interactions can be transformed into multidimensional behavioral measures at scale. A similar approach may be relevant beyond developmental disorders, to include neurological and psychiatric conditions where patients present with altered movement, posture, speech or language. Methods Clinical Population This study used video-recorded conversations with 2,341 children (aged 5-22, M=10.1, SD = 3.39 years) as part of the Healthy Brain Network (HBN) 23 . The cohort included 1,455 participants with clinician-confirmed ADHD: 653 with inattentive presentation, 93 with hyperactive-impulsive presentation, and 625 with combined presentation. An additional 84 showed ADHD traits without meeting diagnostic criteria for a specific ADHD presentation (DSM-5 ADHD-other/unspecified). The sample also comprised 349 individuals with mild to moderate ASD (mostly level 1 according to DSM-5). Of the 1455 participants with ADHD, 284 (19.5%) had a co-occurring ASD diagnosis, a rate similar to those reported in previous community-based cohorts 85 , 86 . The cohort included 168 Typically Developing individuals and 653 with other clinical diagnoses, mostly anxiety and learning disorders. This group did not exhibit significant differences in autistic, hyperactive, or inattentive traits compared to the Typically Developing controls ( Table 1 ), supporting their distinction from the primary target groups in specific ADHD and ASD traits. Furthermore, despite the wide developmental range of the study, there were no marked differences in age distributions between diagnostic and Typically Developing groups ( Table 1 and Fig. S1 ). The overall male-to-female ratio was 1.98:1 (1556/785). Detailed demographic and clinical characteristics per diagnosis are available in Table 1 . Of the 2,341 participants, 1143 (48.8%) were White/Caucasian, 311 (13.3%) Black/African American, 226 Hispanic (9.7%), 55 (2.3%) Asian, and 42 (1.8%) identified as other races. 382 (16.3%) identified as two or more races. 182 (7.8%) participants did not specify, or their race was unknown. Behavioral Task and Data Acquisition Naturalistic behavior was recorded during a semi-structured clinical interview following the viewing of an emotive animated short film (\"The Present\"). The film depicts a boy who initially rejects a three-legged puppy given to him by his mom while playing a videogame; in the final scene, it is revealed that the boy himself has a missing leg. The narrative therefore invites viewers to infer the characters’ emotional states and the similarity between them. The subsequent interview was structured to elicit responses related to narrative recall, emotional description, and perspective-taking (Theory of Mind) (see Table S1 for a list of the questions and Table S3 for an overview of related measures). High-definition video (Canon XC15) of the participant’s face and upper body and high-fidelity audio (Røde NT1) were captured simultaneously for automated analysis. Examples for the interview settings and camera view are available at Fig. S7 . Diagnostic Procedure The HBN is a large-scale biobank that utilizes a community-referred recruitment model to capture the wide heterogeneity inherent in developmental psychopathology. As such, the HBN sample is enriched for individuals meeting DSM-5 criteria. Diagnoses within the HBN are determined by licensed clinicians at the Child Mind Institute based on a consensus model that integrates multiple sources of information. The primary diagnostic instrument is the semi-structured Kiddie Schedule for Affective Disorders and Schizophrenia (K-SADS) 87 for DSM-5, which is administered to both participants and their parents. The final clinician consensus diagnosis also incorporates behavioral observations during extensive testing, parental and self-reported clinical questionnaires, and the participant’s developmental and educational history. For ASD, the primary evaluation relies on the KSADS autism module and questionnaires (e.g., SRS-2, SCQ, ASSQ). Participants with suspected autism received supplemental \"gold-standard\" assessments like the ADOS or ADI-R to confirm diagnosis by trained clinicians. Further details on the diagnostic process and clinician-administered assessments are available on the original HBN paper 23 . For our analysis, clinician-confirmed diagnoses were modeled to assess the impact of specific conditions. ADHD diagnoses were categorized as inattentive and/or hyperactive/impulsive diagnostic status. Participants with a clinician-confirmed diagnosis of ADHD-Combined presentation were considered positive for both hyperactive and inattentive. ASD diagnoses were modeled as a single binary variable (1 for positive, 0 for negative). The vast majority of these individuals (>95%) had mild to moderate presentations (only a few requiring significant support, DSM-5 level 1), consistent with the HBN’s inclusion criteria, which require participants to be verbal (see Fig. S6 for distribution of autism instrument scores). The control group, designated as the Typically Developing Population, consisted of individuals with a confirmed negative diagnostic clinician decision in all DSM-5 categories. Clinical Instruments Participants in the HBN are deeply phenotyped using a comprehensive battery of clinical instruments. For this study, we defined ADHD traits using the Strengths and Weaknesses of ADHD Symptoms and Normal Behavior Scale (SWAN) 88 , and ASD traits using the Autism Spectrum Screening Questionnaire (ASSQ) 89 . These scales were selected as they showed the strongest unique association with their respective clinician-confirmed diagnoses compared to other available scales (see Fig. S5 ). Other clinical instruments used in this study include the Mood and Feelings Questionnaire (MFQ, depressive traits) 90 , Screen for Child Anxiety Related Disorders (SCARED-P, anxiety traits) 91 , and the Child Behavior Checklist (CBCL) 92 , with two factors: internalizing and externalizing traits. A full description and statistical analysis of age, sex, IQ, and clinical instrument scores for each diagnostic group is provided in Tables 1 and S2 . IQ score Cognitive ability was assessed using a composite IQ score. Full-Scale IQ scores were primarily measured using the Wechsler Intelligence Scale for Children (WISC-V) 54 (N=1984, mean +- SD = 99.2 +- 16.7). Exceptions included early participants who were administered the Wechsler Abbreviated Scale of Intelligence (WASI) 55 (N=36, 97.8 +- 14.7), and children under age 6 or with known IQ below 70, for whom the Kaufman Brief Intelligence Test (KBIT) 56 was used (N=166, 100.4 +- 15.8). Full-Scale IQ Composite scores were available for 2186 out of the 2341 participants (93.3%). Similarly, verbal IQ scores were obtained by combining the WISC-VCI (N=1986, 103.6 +- 16.6), WASI-VCI (N=36, 99.3 +- 14.4), and KBIT-verbal (N=166, 100.6 +-15.9) scores. Verbal IQ scores were available for 2188/2341 participants (93.5%). The main analysis and results were computed using Full-Scale IQ as a covariate. However, the main associations between diagnoses and measured behavior were independent of including IQ as a covariate or not ( Fig. S3 ). Similarly, the results did not change significantly if modeling with Verbal IQ scores instead. Transcription, Diarization, & Question-Answer extraction Audio recordings were transcribed using the WhisperX model (v.3.3.0, Whisper model large-v2 ) 93 , generating a raw transcript with word-level timestamps. Validation against a manually transcribed and diarized ground-truth dataset (N=50, author AS) demonstrated a Word Error Rate (WER) of 3.75% +- 4.59% (M +- SD). To separate interviewer and participant speech, we utilized a custom prompt with Google’s large language model ( gemini-2.0-flash ). This context-aware approach yielded a Diarization Error Rate (DER) of 3.16% +- 3.97%, significantly outperforming standard acoustic-based methods ( pyannote/speaker-diarization-3.1 94 DER: 16.35% +- 8.78%). The LLM leveraged the semi-structured interview context to infer speaker turns, avoiding acoustic ambiguities caused by pitch and/or speech overlap between children and interviewers. A secondary LLM prompt subsequently isolated specific interview questions and the participant’s corresponding answers. Complete prompts are provided in the Supplementary Sections S2 and S3. The number of valid answers per question is available in Fig. S8 . Description of Behavioral Variables Various open-source tools were used to obtain objective measures of language, speech prosody, semantic content, and movement from the interview data. An in-depth description of the measures and their extraction methods is provided in Table S3 ; the distribution of the measures is available in Figure S9 . The code for the extraction of measures is available under the link in the Code Availability section. Language and Speech Prosody Measures We extracted a set of acoustic and linguistic features using the Openwillis toolkit (v.3.0.5) 58 . Openwillis is a Python wrapper developed by Brooklyn Health that automates feature extraction by integrating previously established open-source speech and natural language processing tools for the extraction of objective language and speech measures. See Table S3 for a complete description of all measures and their extraction methods. Age-Normed Semantic Measures To characterize the semantic content of participants’ answers, we utilized Google’s ’Gecko’ text embedding model ( text-embedding-004 ) to generate 768-dimensional numerical representations of each response that capture meaning in a high-dimensional semantic space 95 . To distinguish diagnosis-related divergence from typical developmental immaturity, we adopted an age-normed approach. First, we stratified the Typically Developing control sample into three developmentally distinct age bins: Middle Childhood (5-10 years), Early Adolescence (10-14 years), and Late Adolescence/Young Adulthood (14-22 years). Subsequently, for each of the 23 interview questions, we calculated a bin-specific semantic baseline by computing the median embedding of answers provided by TD participants within that specific age range. We then computed the cosine similarity between each participant’s response and the baseline of their corresponding age bin. This yielded a score representing how semantically ‘typical’ the response was. Participants with missing answers were assigned a missing score for that question. The interpretation of this score depends on the question’s content; for narrative-recount questions, higher similarity reflects greater \"typical narrative detail,\" whereas for questions about emotion, lower similarity may indicate more nuanced or idiosyncratic descriptions. Finally, these 23 individual typicality scores were averaged into conceptually meaningful domains based on the function of the questions. These domains included narrative detail, factual memory, thematic understanding, and perspective taking (e.g., describing one’s own or a character’s emotions, or the similarity between the dog and the kid, and how this relates to the kid’s behavior). For example, the ’self-emotion description’ score was the average semantic similarity score across answers to questions 14, 17, 20, and 23 (see Table S1 for details on these questions and examples of the ‘Typical Answers’ for each age bin). The Fisher-Z transform (inverse hyperbolic tangent, artanh) was applied to the semantic (cosine) similarity measures prior to regression analysis to approximate normal distributions. Movement measures We extracted anatomical and facial landmarks from video recordings at the native 30Hz sampling rate of the video recordings using Google’s Mediapipe Holistic model (v.0.10.11) on a local machine. The raw 3D landmark coordinates were filtered to remove low-confidence data based on two criteria applied over a sliding window. First, segments with more than 10% missing values in a 2-second window were excluded. Second, segments exhibiting excessive jitter, defined as a standard deviation of the x-coordinate exceeding empirically-defined thresholds (0.1 for face; 0.01 for pose and landmarks) in 10-frame windows, were also excluded. A landmark point was considered valid for analysis only if it passed both of these checks. To ensure comparability across participants, the filtered data were normalized. Facial landmarks were aligned to a canonical model using an affine transformation 96 to correct for head rotation, translation, and scale. Body landmarks were normalized by the frame-by-frame inter-shoulder distance to render movements invariant to body size and participant’s distance from the camera. Finally, movement was quantified as the frame-to-frame 3D Euclidean displacement for each landmark. These displacement values were then averaged within predefined anatomical regions (e.g., face, mouth, eyes, and body, see Fig. S7 ) and across all valid frames to yield a single mean movement score per region for each participant. Log of the average movement was used to model the effects of diagnosis on such measures. Analytical Approach & Statistical Modeling To test our primary research questions, we modeled the relationship between the measurements of interest and clinical diagnoses using multivariate linear regression. The model isolated the unique contribution of ASD and ADHD diagnostic status while controlling for key developmental and cognitive confounders. For each behavioral metric, the statistical model took the following form: with one model for each dependent variable (e.g., speech rate, lexical diversity), and the predictors were age (in years), a composite IQ score (Full-Scale), Sex (coded binary female=1, male=0), and binary variables for clinician-confirmed diagnoses of ASD, ADHD-Inattentive, and ADHD-Hyperactive (ADHD-Combined presentation was modeled as positive in both). Every participant counts as n=1, with a set of numbers describing them: Age, IQ, ASD, ADHD-Hyper, ADHD-Inatt. For instance, an 8-year-old boy with an IQ of 90 who presents with ASD and ADHD-Combined is coded as: 8, 90, 0, 1, 1, 1. Models were fitted using the bisquare ‘robust’ linear fitting option in MATLAB (2024b). Variance Inflation Factors (VIF) confirmed that multicollinearity was not a confounding factor (all VIFs < 1.32). To quantify and compare the effect of each predictor, we estimated its effect size from the t statistic as follows: Here n is the degrees of freedom, and n 1 and n 2 are the number of data points for each binary group. These effect size measures allow for comparison across different variable types, even when the number of observations varies due to missing data for a specific behavioral metric. For the binary variables (Sex and Disorders), equation (3) is the traditional Cohen’s d, and for continuous variables (Age and IQ), equation (2) captures the traditional Cohen’s f. In both instances, it is the effect on the outcome variable over the residual unexplained error. Statistical tables for each model of the behavioral outcome measures, including all beta coefficients, standard errors, t-statistics, effect sizes, and p-values, are provided in Supplementary Sections S1.2-S1.5. All multivariate model p-values reported in the main manuscript text and figures are corrected for multiple comparisons using Bonferroni correction by multiplying model p-values by the number of tested behavioral features. To assess diagnostic co-occurrence, we constructed three separate logistic regression models. For each model, one clinical diagnosis served as the binary outcome variable, while the remaining diagnoses were included as predictors, alongside age and sex. Crucially, the specific diagnosis used as the outcome was excluded from the predictor set for that model. We report odds ratios (OR) as the exponents of the beta coefficients ( Fig. 2A and black arrows in Fig. 3C , Supplement Section S1.1): Alternative analysis approaches Our primary and planned modeling approach was designed to maximize the utility of this large, heterogeneous sample, capturing the full spectrum of age and diagnostic complexity. To validate these methodological decisions, we conducted several post-hoc sensitivity analyses: First, although our models accounted for the significant overlap between ADHD and ASD, other co-occurring conditions could have influenced the results. However, we ruled out the impact of child anxiety on our behavioral measures by repeating all analyses while modeling for co-occurring positive anxiety diagnoses, and the associations with ADHD or ASD remained consistent ( Fig. S10 ). Second, although our reference group included individuals with other non-target DSM-5 diagnoses (N=653), excluding these participants to allow for a direct comparison with Typically Developing youths did not alter the significant associations found for ADHD or ASD. Third, to mitigate potential bias from the skewed age distribution, we repeated the analysis, limiting it to the 5–14-year-old subset. All results remained the same, except that vocal loudness was no longer associated with ASD and ADHD-Hyperactive status. To clearly distinguish the variance related to Hyperactive versus Inattentive status, we did not separately code for a Combined presentation, unlike some prior studies 97 – 99 . When we repeated the analysis by coding with the Combined presentation instead of Hyperactive status, results on movement remained the same ( Fig. S11 ). Fourth, our planned analysis treated ASD, ADHD-Hyperactive, and ADHD-Inattentive as independent additive factors. This modeling choice is supported by the clinical instrument data in Table 1 , which shows that the only significant trait difference between ASD with co-occurring ADHD (ASD[ADHD+]) and without (ASD[ADHD-]) lies in hyperactive and inattentive traits, whereas the autistic traits and other clinical measures remain comparable. We nevertheless explored whether co-occurring ADHD and ASD represent a distinct phenotypic category rather than an additive condition. When we modeled the data using mutually exclusive groups (ADHD-only, ASD-only, and ADHD+ASD), as is often done in the literature 49 – 51 the distinct separation of effects observed in our primary additive model was lost ( Fig. S12 ), supporting our approach of treating these as independent, overlapping dimensions. Finally, the linear model (1) does not include interaction terms. For age in particular, this means that the effects of the disorders are modeled as a constant gap between a positive and a negative diagnosis. However, for some behavioral features, the gap appears to increase with age (e.g., Fig. 3D , 4B). Such an overall gap can be detected by model (1) even if it increases with age. In a post-hoc analysis, we included the interaction of age with all diagnostic variables. None of the interactions were statistically significant after multiple comparison correction, while the Bayesian Information Criterion increased over the reduced model for most of the measures, suggesting that we are lacking statistical power to warrant the increase in model complexity. Narrowing it to variables with a main age and diagnosis effect suggests that there may be an increasing gap with age for narrative detail in ASD ( Fig. 3D, p =0.034 uncorrected), and for body movement in ADHD-Hyperactivity ( Fig. 5C, p =0.044 uncorrected). Conversely, differences appear to diminish with age for pitch, pitch variation, and breathiness in ASD (p=0.004, 0.022 & 0.029, uncorrected). However, note that these are uncorrected and post-hoc analyses. Visualization of Distinctive Terms Semantic embedding analysis of the answers is quantitative and applies to all questions asked. To gain a sense of what it captures, we conducted post-hoc analyses on the significant ASD-question associations. To identify and visualize terms that distinguish ASD from non-ASD participant answers, we employed the scattertext library 57 . The horizontal axis represents the log frequency of each term in the corpus, while the vertical axis represents the term’s association with the ASD group, calculated using a Scaled F-Score (β=2) and normalized via the normal cumulative distribution function to a range of [-1, 1], as done previously for the comparison of suicidal vs non-suicidal terms 100 . This scoring metric was selected to prioritize recall, where values approaching +1 indicate broad usage within the ASD group, and values approaching -1 indicate broad usage within the non-ASD control group. See Fig. S2 for the visualization of answers with a significant ASD association. Ethics Statement The data used in this study were collected between 2016 and 2022. The HBN protocol was approved by the Chesapeake Institutional Review Board. Written informed consent was obtained from participants aged 18 or older, and from legal guardians for participants younger than 18. Access to the data was governed by a Data Transfer and Use Agreement (dated 6/27/2024) between the Child Mind Institute and The City University of New York. Data Availability Phenotypical and diagnostic data are available upon request and with the establishment of a Data Use Agreement with the Child Mind Institute at https://fcon_1000.projects.nitrc.org/indi/cmi_healthy_brain_network/ . Interview videos are not currently available due to the privacy of the children. Code Availability The full code pipeline used for speech transcription and diarization, as well as to extract language, vocal prosody, semantic embeddings, and body and face landmarks, is publicly available at https://github.com/asortubay/AI_behavioral_analysis . Supplementary Material View this table: View inline View popup Table S1. Interview Questions & Typical Answers View this table: View inline View popup Table S2. Detailed Statistical Test Results for Table 1 View this table: View inline View popup Table S3. Description of Behavioral Variables. Download figure Open in new tab Figure S1. Co-occurring diagnoses and age distributions. A) Co-occurring diagnoses (Top 12) in each of the three diagnostic groups of interest. B) Age-Sex distribution across the diagnostic groups. ADHD-Inatt and ADHD-Hyper contain participants with an ADHD-Combined presentation. Other Diagnosis grop contains participants with some positive diagnosis, excluding those with ADHD or ASD. Download figure Open in new tab Figure S2. Scattertext visualization of used terms when answering specific questions. We only did this post-hoc analysis to add interpretability to the found effects with the semantic analysis; we show here the questions that showed a significant effect of ASD. Download figure Open in new tab Figure S3. Results are independent of including IQ as a covariate. Effect size matrices for primary behavioral outcomes after excluding IQ as a covariate. The models (A through D) are identical to the primary multivariate models presented in the main text, but without modelling IQ as a covariate. Effect sizes are estimated using multivariate regression, with p<0.01 after Bonferroni correction. How do the features used here relate to established clinical instruments based on questionnaires? As one may expect, our measures of movement, language, and speech correlate strongly with these established metrics ( Figure S4A, B ). When using standard instruments, it appears that ADHD-Inattention is also associated with elevated motor activity ( Fig. S4C ). Thus, our movement measures appear to be more specifically associated with the ADHD-Hyperactive status (cf. Fig. 5C with S4C). On the other hand, when using standard instruments, it appears that ASD has a lesser effect on language use ( Fig. S4D ). Therefore, our features related to social communication appear to be more sensitive to the ASD presentation (cf. Fig. 3E with S4D). Download figure Open in new tab Figure S4. Comparing Clinical Instruments to measured behavior. Spearman correlation between automatically measured behaviors and the clinical instrument scores measuring movement (A), and language use/prosody (B). C-D) Effects of development and diagnosis on such clinical instrument scores. Effect sizes in C, D measured using multivariate models, significance at p<0.01 after Bonferroni correction, white otherwise. For A and B, significant at a cutoff of p<0.05, uncorrected, white otherwise. To select the most specific and robust trait measures for our analyses in Figures 1 - 2 , we evaluated which of these scales best and most uniquely reflected the corresponding gold-standard, clinician-confirmed diagnosis. We ran separate multivariate regression models for each clinical instrument, entering the diagnoses as predictors. This approach allowed us to see the instruments most uniquely associated with each diagnosis. As shown here, the SWAN Inattention and Hyperactivity subscales and the ASSQ emerged as the strongest and most specific associations for their corresponding clinician-confirmed diagnoses. Therefore, we selected these scales for all subsequent analyses where continuous measures of ADHD and ASD traits were required: Download figure Open in new tab Figure S5. Unique association of clinical questionnaire scales with clinician-confirmed diagnoses. The effect size matrices reveal which clinical instruments are more uniquely explained by specific positive diagnoses while accounting for development and demographic effects. Download figure Open in new tab Figure S6. Distribution of Autism Scores in the ASD-positive population. The Healthy Brain Network’s deep phenotyping protocol provides multiple parent- and self-report questionnaires to measure traits associated with ADHD and ASD. These included the Strengths and Weaknesses of ADHD Symptoms and Normal Behavior Scale (SWAN), the Autism Spectrum Screening Questionnaire (ASSQ), the Social Communication Questionnaire (SCQ), the Social Responsiveness Scale-2 (SRS-2), the Child Behavior Checklist (CBCL), and the Conners 3 Self-Report (C3SR). The ASD population is mainly composed of mild-to-moderate cases. [removed to meet bioRxiv policy regarding personal pictures] Figure S7. Interview video examples, face landmarks, and validity. The interviews were recorded in different locations and under different lighting conditions. However, all the videos showed the face and upper torso. B) 468 face landmarks tracked by Mediapipe. We characterized total movements for the whole face (blue), only the eyes (magenta), and the mouth (green). C) Most participants showed their face and torso for most part of the video (N=2108 and N=2226 participants were missing 10% or less of the data for face and pose, respectively); however, the amount of available data was less overall for the face landmarks than for the pose landmarks, as participants covered their face or looked away from the camera. Download figure Open in new tab Figure S8. Validity of interview questions. A) Not all questions were asked to every participant, or the answer was not available. In average, 67.7% of participants answered each question. B) Each participant answered 19.9 questions in average. C) The number of questions answered by each participant is correlated with the LLM’s rating of the conversation completion (ρ(2339) = 0.79, p<0.001), see Supplementary Section S3 for rating prompt. Download figure Open in new tab Figure S9. Distribution of Behavioral Variables. A complete description of variables and their units is available in Table S3 . Given the high rates of anxiety in both ADHD and ASD, and the potential for a semi-structured interview with an unfamiliar clinician to elicit anxious behaviors, we modeled this as a key potential confound. Behaviors associated with anxiety, such as motor restlessness, social reticence, or altered vocal tone, could overlap with the objective behavioral markers under investigation and potentially obscure the specific effects of ADHD and ASD. To address this and confirm the robustness of our primary findings, we repeated our primary multivariate regressions, adding a binary variable for any clinician-confirmed DSM-5 anxiety diagnosis as an additional covariate (N = 789). This allowed us to test whether the observed associations between our objective behavioral markers and ADHD/ASD diagnoses remained significant after statistically controlling for the influence of anxiety. We find that all the associations between ASD and/or ADHD with our behavioral measures are independent of whether we model anxiety or not: Download figure Open in new tab Figure S10. Results are independent of anxiety diagnoses. Effect size matrices for primary behavioral outcomes after controlling for co-occurring anxiety. The models (A through D) are identical to the primary multivariate models presented in the main text, with the addition of a binary covariate for any clinician-confirmed anxiety diagnosis (N=789). Effect sizes are estimated using multivariate regression, with p<0.01 after Bonferroni correction. Throughout the work presented in this manuscript, clinician-diagnosed ADHD-Combined presentations were modelled as positive in both ADHD-Hyperactive and ADHD-Inattentive status. This was done to maximize the number of participants in the modelled group, while maintaining the distinction of the two main inattentive and hyperactive diagnostic statuses. Previous studies on the HBN population have modelled this overlap by considering the ADHD-Combined separate from the ADHD-Inattentive presentations 97 – 99 . We repeated the analyses on the behavioral features by following this approach, and the results were mostly consistent our associations with the ADHD-Hyperactive presentation: Download figure Open in new tab Figure S11. Results are independent of how the ADHD presentations are coded. Effect size matrices for primary behavioral outcomes after modelling ADHD-Combined and ADHD-Inattentive presentations separately. The models (A through D) are identical to the primary multivariate models presented in the main text, but ADHD was modelled separately as the ADHD-Combined (N=625) and ADHD-Inattentive (N=653) presentations. Effect sizes are estimated using multivariate regression, with p<0.01 after Bonferroni correction. Download figure Open in new tab Figure S12. Modeling ASD(ADHD+) and ASD(ADHD-) separately does not paint a clear picture. Effect size matrices for primary behavioral outcomes after treating co-occurring ADHD and ASD as a separate category. The models (A through D) are identical to the primary multivariate models presented in the main text, but without modelling IQ as a covariate. Effect sizes are estimated using multivariate regression, with p<0.01 after Bonferroni correction. Supplementary Section S1 Multivariate Model Reports p-values reported in the tables below appear uncorrected for multiple comparisons unless otherwise specified. In the main manuscript, and in the reported corrected p-values below, all p-values are corrected for multiple comparisons using Bonferroni correction by multiplying these p-values by the number of behavioral features tested. Specifically, in Figures 2 - 5 , Effect Size matrices show significant effect sizes in color with a cutoff of p<0.01 on the corrected p-value, and are left blank otherwise. See Methods for the explanation on analytical models. View this table: View inline View popup Download powerpoint Supplementary Section S1.1. Diagnoses Detailed model reports for panel A in Figure 2 and black arrows in panel C in Figure 3 . View this table: View inline View popup Supplementary Section S1.2. Language Use Measures Detailed model reports for panels D & E in Figure 2 . View this table: View inline View popup Supplementary Section S1.3. Semantic Measures Detailed model reports for panel E in Figure 3 . View this table: View inline View popup Supplementary Section S1.4. Vocal Prosody Measures Detailed model reports for panel B in Figure 4 . View this table: View inline View popup Supplementary Section S1.5. Movement Measures Detailed model reports for panel B in Figure 5 . Download figure Open in new tab Download figure Open in new tab Supplementary Section S2. Diarizing Prompt for LLM Download figure Open in new tab Download figure Open in new tab Download figure Open in new tab Download figure Open in new tab Download figure Open in new tab Download figure Open in new tab Download figure Open in new tab Download figure Open in new tab Download figure Open in new tab Supplementary Section S3. Question-Answer Extraction Prompt for LLM Acknowledgments & Funding NIH-NIMH P50 MH109429, CA Department of Health, grant to Nathan Kline Institute, Milham (PI). Gemini 3.1 Pro was used to assist in shortening our original text in the Introduction, Results and Discussion. Footnotes Updated Figures and shortened main manuscript text. Added some post-hoc analyses on alternate analysis methods and updated the semantic analysis to compare against age-matched controls. References 1. ↵ Kogan , M. D. et al. Prevalence of parent-reported diagnosis of autism spectrum disorder among children in the US, 2007 . Pediatrics 124 , 1395 – 1403 ( 2009 ). OpenUrl CrossRef PubMed Web of Science 2. Polanczyk , G. , De Lima , M. S. , Horta , B. L. , Biederman , J. & Rohde , L. A . The worldwide prevalence of ADHD: a systematic review and metaregression analysis . Am. J. Psychiatry 164 , 942 – 948 ( 2007 ). OpenUrl CrossRef PubMed Web of Science 3. Polanczyk , G. V. , Willcutt , E. G. , Salum , G. A. , Kieling , C. & Rohde , L. A . ADHD prevalence estimates across three decades: an updated systematic review and meta-regression analysis . Int. J. Epidemiol . 43 , 434 – 442 ( 2014 ). OpenUrl CrossRef PubMed 4. ↵ Yang , Y. et al. Prevalence of neurodevelopmental disorders among US children and adolescents in 2019 and 2020 . Front. Psychol . 13 , 997648 ( 2022 ). 5. ↵ American Psychiatric Association . Diagnostic and Statistical Manual of Mental Disorders: DSM-5 . ( American psychiatric association , 2013 ). 6. ↵ Grzadzinski , R. et al. Examining Autistic Traits in Children with ADHD: Does the Autism Spectrum Extend to ADHD? J. Autism Dev. Disord . 41 , 1178 – 1191 ( 2011 ). OpenUrl CrossRef PubMed 7. Hattori , J. et al. Are pervasive developmental disorders and attention-deficit/hyperactivity disorder distinct disorders? Brain Dev . 28 , 371 – 374 ( 2006 ). OpenUrl CrossRef PubMed Web of Science 8. Hollingdale , J. , Woodhouse , E. , Young , S. , Fridman , A. & Mandy , W . Autistic spectrum disorder symptoms in children and adolescents with attention-deficit/hyperactivity disorder: a meta-analytical review . Psychol. Med . 50 , 2240 – 2253 ( 2020 ). OpenUrl PubMed 9. Holtmann , M. , Bölte , S. & Poustka , F . Attention deficit hyperactivity disorder symptoms in pervasive developmental disorders: association with autistic behavior domains and coexisting psychopathology . Psychopathology 40 , 172 – 177 ( 2007 ). OpenUrl CrossRef PubMed Web of Science 10. Nijmeijer , J. S. et al. Attention-deficit/hyperactivity disorder and social dysfunctioning . Clin. Psychol. Rev . 28 , 692 – 708 ( 2008 ). OpenUrl CrossRef PubMed Web of Science 11. Reiersen , A. M. , Constantino , J. N. , Volk , H. E. & Todd , R. D . Autistic traits in a population-based ADHD twin sample . J. Child Psychol. Psychiatry 48 , 464 – 472 ( 2007 ). OpenUrl CrossRef PubMed Web of Science 12. ↵ Simonoff , E. et al. Psychiatric disorders in children with autism spectrum disorders: prevalence, comorbidity, and associated factors in a population-derived sample . J. Am. Acad. Child Adolesc. Psychiatry 47 , 921 – 929 ( 2008 ). OpenUrl CrossRef PubMed Web of Science 13. ↵ Mulligan , A. et al. Autism symptoms in Attention-Deficit/Hyperactivity Disorder: A Familial trait which Correlates with Conduct, Oppositional Defiant, Language and Motor Disorders . J. Autism Dev. Disord . 39 , 197 – 209 ( 2009 ). OpenUrl CrossRef PubMed Web of Science 14. Reiersen , A. M. , Constantino , J. N. , Grimmer , M. , Martin , N. G. & Todd , R. D . Evidence for shared genetic influences on self-reported ADHD and autistic symptoms in young adult Australian twins . Twin Res. Hum. Genet . 11 , 579 – 585 ( 2008 ). OpenUrl CrossRef PubMed 15. Rommelse , N. N. , Franke , B. , Geurts , H. M. , Hartman , C. A. & Buitelaar , J. K . Shared heritability of attention-deficit/hyperactivity disorder and autism spectrum disorder . Eur. Child Adolesc. Psychiatry 19 , 281 – 295 ( 2010 ). OpenUrl CrossRef PubMed Web of Science 16. ↵ Ronald , A. , Simonoff , E. , Kuntsi , J. , Asherson , P. & Plomin , R . Evidence for overlapping genetic influences on autistic and ADHD behaviours in a community twin sample . J. Child Psychol. Psychiatry 49 , 535 – 542 ( 2008 ). OpenUrl CrossRef PubMed Web of Science 17. ↵ Aoki , Y. et al. Association of White Matter Structure With Autism Spectrum Disorder and Attention-Deficit/Hyperactivity Disorder . JAMA Psychiatry 74 , 1120 – 1128 ( 2017 ). OpenUrl PubMed 18. Brieber , S. et al. Structural brain abnormalities in adolescents with autism spectrum disorder and patients with attention deficit/hyperactivity disorder . J. Child Psychol. Psychiatry 48 , 1251 – 1258 ( 2007 ). OpenUrl CrossRef PubMed Web of Science 19. Christakou , A. et al. Disorder-specific functional abnormalities during sustained attention in youth with attention deficit hyperactivity disorder (ADHD) and with autism . Mol. Psychiatry 18 , 236 – 244 ( 2013 ). OpenUrl CrossRef PubMed Web of Science 20. Di Martino , A. et al. Shared and Distinct Intrinsic Functional Network Centrality in Autism and Attention-Deficit/Hyperactivity Disorder . Biol. Psychiatry 74 , 623 – 632 ( 2013 ). OpenUrl CrossRef PubMed Web of Science 21. Bedford , S. A. et al. Brain-Charting Autism and Attention-Deficit/Hyperactivity Disorder Reveals Distinct and Overlapping Neurobiology . Biol. Psychiatry 97 , 517 – 530 ( 2025 ). OpenUrl CrossRef PubMed 22. ↵ Norman , L. J. et al. Cross-sectional mega-analysis of resting-state alterations associated with autism and attention-deficit/hyperactivity disorder in children and adolescents . Nat. Ment. Health 3 , 709 – 723 ( 2025 ). OpenUrl PubMed 23. ↵ Alexander , L. M. et al. An open resource for transdiagnostic research in pediatric mental health and learning disorders . Sci. Data 4 , 170181 ( 2017 ). 24. ↵ Geurts , H. M. & Embrechts , M . Language Profiles in ASD, SLI, and ADHD . J. Autism Dev. Disord . 38 , 1931 – 1943 ( 2008 ). OpenUrl CrossRef PubMed Web of Science 25. ↵ Bishop , D. V. & Baird , G . Parent and teacher report of pragmatic aspects of communication: use of the Children’s Communication Checklist in a clinical setting . Dev. Med. Child Neurol . 43 , 809 – 818 ( 2001 ). OpenUrl CrossRef PubMed Web of Science 26. Geurts , H. M. et al. Can the Children’s Communication Checklist differentiate between children with autism, children with ADHD, and normal controls? J. Child Psychol. Psychiatry 45 , 1437 – 1453 ( 2004 ). OpenUrl CrossRef PubMed Web of Science 27. ↵ Geurts , H. M. & Embrechts , M . Language profiles in ASD, SLI, and ADHD . J. Autism Dev. Disord . 38 , 1931 – 1943 ( 2008 ). OpenUrl CrossRef PubMed Web of Science 28. ↵ Kim , Y.-S. G. , Dore , R. , Cho , M. , Golinkoff , R. & Amendum , S. J . Theory of mind, mental state talk, and discourse comprehension: Theory of mind process is more important for narrative comprehension than for informational text comprehension . J. Exp. Child Psychol . 209 , 105181 ( 2021 ). 29. ↵ Baron-Cohen , S. , Leslie , A. M. & Frith , U . Does the autistic child have a “theory of mind” ? Cognition 21 , 37 – 46 ( 1985 ). OpenUrl CrossRef PubMed Web of Science 30. ↵ Baron-Cohen , S. , Leslie , A. M. & Frith , U . Does the autistic child have a “theory of mind” ? Cognition 21 , 37 – 46 ( 1985 ). OpenUrl CrossRef PubMed Web of Science 31. ↵ Tager-Flusberg , H. , Paul , R. & Lord , C. Language and communication in autism . Handb. Autism Pervasive Dev. Disord . 1 , 335 – 364 ( 2005 ). OpenUrl 32. ↵ Kuijper , S. J. M. , Hartman , C. A. , Bogaerds-Hazenberg , S. T. M. & Hendriks , P . Narrative production in children with autism spectrum disorder (ASD) and children with attention-deficit/hyperactivity disorder (ADHD): Similarities and differences . J. Abnorm. Psychol . 126 , 63 – 75 ( 2017 ). OpenUrl PubMed 33. ↵ Fusaroli , R. , Lambrechts , A. , Bang , D. , Bowler , D. M. & Gaigg , S. B . “ Is voice a marker for Autism spectrum disorder? A systematic review and meta-analysis ”. Autism Res . 10 , 384 – 407 ( 2017 ). OpenUrl PubMed 34. ↵ Trayvick , J. et al. Speech and language patterns in autism: Towards natural language processing as a research and clinical tool . Psychiatry Res . 340 , 116109 ( 2024 ). 35. ↵ Hamdan , A.-L. et al. Vocal Characteristics in Children With Attention Deficit Hyperactivity Disorder . J. Voice 23 , 190 – 194 ( 2009 ). OpenUrl CrossRef PubMed 36. ↵ Barona-Lleo , L. & Fernandez , S . Hyperfunctional Voice Disorder in Children With Attention Deficit Hyperactivity Disorder (ADHD). A Phenotypic Characteristic? J. Voice 30 , 114 – 119 ( 2016 ). OpenUrl PubMed 37. ↵ Briend , F. et al. Voice acoustics allow classifying autism spectrum disorder with high accuracy . Transl. Psychiatry 13 , 250 ( 2023 ). 38. Eni , M. et al. Reliably quantifying the severity of social symptoms in children with autism using ASDSpeech . Transl. Psychiatry 15 , 14 ( 2025 ). 39. ↵ Godel , M. et al. Prosodic signatures of ASD severity and developmental delay in preschoolers . Npj Digit. Med . 6 , 99 ( 2023 ). 40. ↵ de Belen , R. A. J. , Bednarz , T. , Sowmya , A. & Del Favero , D . Computer vision in autism spectrum disorder research: a systematic review of published studies from 2009 to 2019 . Transl. Psychiatry 10 , 1 – 20 ( 2020 ). OpenUrl PubMed 41. ↵ Zampella , C. J. , Wang , L. A. L. , Haley , M. , Hutchinson , A. G. & de Marchena , A . Motor Skill Differences in Autism Spectrum Disorder: a Clinically Focused Review . Curr. Psychiatry Rep . 23 , 64 ( 2021 ). 42. ↵ De Crescenzo , F. et al. The use of actigraphy in the monitoring of sleep and activity in ADHD: A meta-analysis . Sleep Med. Rev . 26 , 9 – 20 ( 2016 ). OpenUrl CrossRef PubMed 43. ↵ Hartman , C. A. , Geurts , H. M. , Franke , B. , Buitelaar , J. K. & Rommelse , N. N. J . Changing ASD-ADHD symptom co-occurrence across the lifespan with adolescence as crucial time window: Illustrating the need to go beyond childhood . Neurosci. Biobehav. Rev . 71 , 529 – 541 ( 2016 ). OpenUrl PubMed 44. ↵ Larsson , H. , Dilshad , R. , Lichtenstein , P. & Barker , E. D . Developmental trajectories of DSM-IV symptoms of attention-deficit/hyperactivity disorder: genetic effects, family risk and associated psychopathology . J. Child Psychol. Psychiatry 52 , 954 – 963 ( 2011 ). OpenUrl CrossRef PubMed 45. ↵ Loomes , R. , Hull , L. & Mandy , W. P. L . What is the male-to-female ratio in autism spectrum disorder? A systematic review and meta-analysis . J. Am. Acad. Child Adolesc. Psychiatry 56 , 466 – 474 ( 2017 ). OpenUrl CrossRef PubMed 46. ↵ Ramtekkar , U. P. , Reiersen , A. M. , Todorov , A. A. & Todd , R. D . Sex and age differences in Attention-Deficit/Hyperactivity Disorder symptoms and diagnoses: Implications for DSM-V and ICD-11 . J. Am. Acad. Child Adolesc. Psychiatry 49 , 217 – 28 .e1–3 ( 2010 ). OpenUrl CrossRef PubMed Web of Science 47. ↵ Joseph , R. M. & Fein , D . The significance of IQ and differential cognitive abilities for understanding ASD. Neuropsychol . Autism 281 , 294 ( 2011 ). 48. ↵ Charman , T. et al. Defining the cognitive phenotype of autism . Brain Res . 1380 , 10 – 21 ( 2011 ). OpenUrl CrossRef PubMed Web of Science 49. ↵ Craig , F. et al. Overlap Between Autism Spectrum Disorders and Attention Deficit Hyperactivity Disorder: Searching for Distinctive/Common Clinical Features . Autism Res . 8 , 328 – 337 ( 2015 ). OpenUrl CrossRef PubMed 50. Rao , P. A. & Landa , R. J . Association between severity of behavioral phenotype and comorbid attention deficit hyperactivity disorder symptoms in children with autism spectrum disorders . Autism 18 , 272 – 280 ( 2014 ). OpenUrl CrossRef PubMed 51. ↵ Scandurra , V. et al. Neurodevelopmental Disorders and Adaptive Functions: A Study of Children With Autism Spectrum Disorders (ASD) and/or Attention Deficit and Hyperactivity Disorder (ADHD) . Front. Psychiatry 10 , ( 2019 ). 52. ↵ Simonoff , E. et al. Psychiatric Disorders in Children With Autism Spectrum Disorders: Prevalence, Comorbidity, and Associated Factors in a Population-Derived Sample . J. Am. Acad. Child Adolesc. Psychiatry 47 , 921 – 929 ( 2008 ). OpenUrl CrossRef PubMed Web of Science 53. ↵ Menyuk , P. & Brisk , M. E. Language Development in Middle Childhood: Ages 9–13 . in Language development and education: Children with varying language experiences 118 – 141 ( Springer , 2005 ). 54. ↵ Wechsler , D. Wechsler Intelligence Scale for Children-Fifth Edition . ( Pearson , Bloomington, MN ). 55. ↵ Wechsler , D. Wechsler abbreviated scale of intelligence-- . ( 1999 ). 56. ↵ Kaufman , A. S. & Kaufman , N. L . Kaufman brief intelligence test . Encycl. Spec. Educ. Ref. Educ. Child. Adolesc. Adults Disabil. Except. Individ . ( 2013 ). 57. ↵ Kessler , J. Scattertext: a Browser-Based Tool for Visualizing how Corpora Differ . in Proceedings of ACL 2017, System Demonstrations (eds Bansal , M. & Ji , H. ) 85 – 90 ( Association for Computational Linguistics , Vancouver, Canada, 2017 ). 58. ↵ Worthington , M. , Efstathiadis , G. , Yadav , V. & Abbas , A . 172. OpenWillis: An Open-Source Python Library for Digital Health Measurement . Biol. Psychiatry 95 , S169 – S170 ( 2024 ). OpenUrl 59. ↵ Korrel , H. , Mueller , K. L. , Silk , T. , Anderson , V. & Sciberras , E . Research Review: Language problems in children with Attention-Deficit Hyperactivity Disorder–a systematic meta-analytic review . J. Child Psychol. Psychiatry 58 , 640 – 654 ( 2017 ). OpenUrl CrossRef PubMed 60. ↵ Bellani , M. , Moretti , A. , Perlini , C. & Brambilla , P . Language disturbances in ADHD . Epidemiol. Psychiatr. Sci . 20 , 311 – 315 ( 2011 ). OpenUrl CrossRef PubMed 61. ↵ Sciberras , E. et al. Language problems in children with ADHD: A community-based study . Pediatrics 133 , 793 – 800 ( 2014 ). OpenUrl CrossRef PubMed Web of Science 62. ↵ Tirosh , E. & Cohen , A . Language deficit with attention-deficit disorder: A prevalent comorbidity . J. Child Neurol . 13 , 493 – 497 ( 1998 ). OpenUrl CrossRef PubMed Web of Science 63. ↵ Redmond , S. M . Conversational profiles of children with ADHD, SLI and typical development . Clin. Linguist. Phon . 18 , 107 – 125 ( 2004 ). OpenUrl PubMed 64. ↵ Jepsen , I. B. , Hougaard , E. , Matthiesen , S. T. & Lambek , R . A systematic review and meta-analysis of narrative language abilities in children with Attention-Deficit/Hyperactivity Disorder . Res. Child Adolesc. Psychopathol . 50 , 737 – 751 ( 2022 ). OpenUrl PubMed 65. ↵ Kuijper , S. J. , Hartman , C. A. , Bogaerds-Hazenberg , S. & Hendriks , P . Narrative production in children with autism spectrum disorder (ASD) and children with attention-deficit/hyperactivity disorder (ADHD): Similarities and differences . J. Abnorm. Psychol . 126 , 63 ( 2017 ). 66. ↵ Geurts , H. M. , Broeders , M. & Nieuwland , M. S . Thinking outside the executive functions box: Theory of mind and pragmatic abilities in attention deficit/hyperactivity disorder . Eur. J. Dev. Psychol . 7 , 135 – 151 ( 2010 ). OpenUrl 67. ↵ Tannock , R. & Schachar , R. Executive dysfunction as an underlying mechanism of behavior and language problems in attention deficit hyperactivity disorder . in Language, learning, and behavior disorders: Developmental, biological, and clinical perspectives 128 – 155 ( Cambridge University Press , New York , NY, US, 1996 ). 68. ↵ Taurines , R. et al. ADHD and autism: differential diagnosis or overlapping traits? A selective review. ADHD Atten. Deficit Hyperact. Disord . 4 , 115 – 139 ( 2012 ). OpenUrl 69. ↵ Naigles , L. R. et al. “You’re telling me!” The prevalence and predictors of pronoun reversals in children with autism spectrum disorders and typical development . Res. Autism Spectr. Disord . 27 , 11 – 20 ( 2016 ). OpenUrl PubMed 70. ↵ Diehl , J. J. , Bennetto , L. & Young , E. C . Story recall and narrative coherence of high-functioning children with autism spectrum disorders . J. Abnorm. Child Psychol . 34 , 83 – 98 ( 2006 ). OpenUrl CrossRef 71. ↵ Baron-Cohen , S. , Leslie , A. M. & Frith , U . Does the autistic child have a “theory of mind” ? Cognition 21 , 37 – 46 ( 1985 ). OpenUrl CrossRef PubMed Web of Science 72. ↵ Losh , M. & Gordon , P. C . Quantifying Narrative Ability in Autism Spectrum Disorder: A Computational Linguistic Analysis of Narrative Coherence . J. Autism Dev. Disord . 44 , 3016 – 3025 ( 2014 ). OpenUrl CrossRef PubMed 73. ↵ Lee , M. et al. What’s the story? A computational analysis of narrative competence in autism . Autism 22 , 335 – 344 ( 2018 ). OpenUrl PubMed 74. ↵ Fusaroli , R. et al. Toward a cumulative science of vocal markers of autism: A cross-linguistic meta-analysis-based investigation of acoustic markers in American and Danish autistic children . Autism Res . 15 , 653 – 664 ( 2022 ). OpenUrl PubMed 75. ↵ Pecoraro , P. M. , Marsili , L. , Espay , A. J. , Bologna , M. & di Biase , L . Computer Vision Technologies in Movement Disorders: A Systematic Review . Mov. Disord. Clin. Pract . ( 2025 ). 76. ↵ Cook , A. , Mandal , B. , Berry , D. & Johnson , M . Towards Automatic Screening of Typical and Atypical Behaviors in Children With Autism . in 2019 IEEE International Conference on Data Science and Advanced Analytics (DSAA) 504 – 510 ( 2019 ). doi: 10.1109/DSAA.2019.00065 . OpenUrl CrossRef 77. Kojovic , N. , Natraj , S. , Mohanty , S. P. , Maillart , T. & Schaer , M . Using 2D video-based pose estimation for automated prediction of autism spectrum disorders in young children . Sci. Rep . 11 , 15069 ( 2021 ). 78. Lemler , C. et al. Semi-Automated Multi-Label Classification of Autistic Mannerisms by Machine Learning on Post Hoc Skeletal Tracking . Autism Res . 18 , 833 – 844 ( 2025 ). OpenUrl PubMed 79. Vyas , K. et al. Recognition Of Atypical Behavior In Autism Diagnosis From Video Using Pose Estimation Over Time. in 2019 IEEE 29th International Workshop on Machine Learning for Signal Processing (MLSP) 1 – 6 ( 2019 ). doi: 10.1109/MLSP.2019.8918863 . OpenUrl CrossRef 80. Zhang , Y. , Tian , Y. , Wu , P. & Chen , D . Application of Skeleton Data and Long Short-Term Memory in Action Recognition of Children with Autism Spectrum Disorder . Sensors 21 , 411 ( 2021 ). 81. ↵ Koehler , J. C. et al. Machine learning classification of autism spectrum disorder based on reciprocity in naturalistic social interactions . Transl. Psychiatry 14 , 76 ( 2024 ). 82. ↵ Faerman , A. et al. Language Abilities are Associated with Both Verbal and Nonverbal Intelligence in Children on the Autism Spectrum . Dev. Neuropsychol . 48 , 248 – 257 ( 2023 ). OpenUrl PubMed 83. ↵ Dennis , M. et al. Why IQ is not a covariate in cognitive studies of neurodevelopmental disorders . J. Int. Neuropsychol. Soc . 15 , 331 – 343 ( 2009 ). OpenUrl CrossRef PubMed Web of Science 84. ↵ Lord , C. et al. The Autism Diagnostic Observation Schedule—Generic: A standard measure of social and communication deficits associated with the spectrum of autism . J. Autism Dev. Disord . 30 , 205 – 223 ( 2000 ). OpenUrl CrossRef PubMed Web of Science 85. ↵ Stevens , T. , Peng , L. & Barnard-Brak , L . The comorbidity of ADHD in children diagnosed with autism spectrum disorder . Res. Autism Spectr. Disord . 31 , 11 – 18 ( 2016 ). OpenUrl 86. ↵ Zablotsky , B. , Bramlett , M. D. & Blumberg , S. J . The Co-Occurrence of Autism Spectrum Disorder in Children With ADHD . J. Atten. Disord . 24 , 94 – 103 ( 2020 ). OpenUrl PubMed 87. ↵ Kaufman , J. et al. Schedule for affective disorders and schizophrenia for school-age children-present and lifetime version (K-SADS-PL): initial reliability and validity data . J. Am. Acad. Child Adolesc. Psychiatry 36 , 980 – 988 ( 1997 ). OpenUrl CrossRef PubMed Web of Science 88. ↵ Swanson , J. M. et al. Categorical and Dimensional Definitions and Evaluations of Symptoms of ADHD: History of the SNAP and the SWAN Rating Scales . Int. J. Educ. Psychol. Assess . 10 , 51 – 70 ( 2012 ). OpenUrl PubMed 89. ↵ Ehlers , S. , Gillberg , C. & Wing , L . A screening questionnaire for Asperger syndrome and other high-functioning autism spectrum disorders in school age children . J. Autism Dev. Disord . 29 , 129 – 141 ( 1999 ). OpenUrl CrossRef PubMed Web of Science 90. ↵ Angold , A. , Costello , E. J. , Messer , S. C. & Pickles , A . Development of a short questionnaire for use in epidemiological studies of depression in children and adolescents . Int. J. Methods Psychiatr. Res . ( 1995 ). 91. ↵ Birmaher , B. et al. Psychometric properties of the Screen for Child Anxiety Related Emotional Disorders (SCARED): a replication study . J. Am. Acad. Child Adolesc. Psychiatry 38 , 1230 – 1236 ( 1999 ). OpenUrl CrossRef PubMed Web of Science 92. ↵ Achenbach , T. M. Integrative Guide for the 1991 CBCL/4-18, YSR, and TRF Profiles . ( Department of Psychiatry, University of Vermont , 1991 ). 93. ↵ Bain , M. , Huh , J. , Han , T. & Zisserman , A. WhisperX: Time-Accurate Speech Transcription of Long-Form Audio . in INTERSPEECH 2023 4489 – 4493 (ISCA, 2023). doi: 10.21437/Interspeech.2023-78 . OpenUrl CrossRef 94. ↵ Bredin , H. et al. Pyannote. audio: neural building blocks for speaker diarization . in 7124 – 7128 ( IEEE , 2020 ). 95. ↵ Lee , J. , et al. Gecko: Versatile Text Embeddings Distilled from Large Language Models . Preprint at doi: 10.48550/arXiv.2403.20327 ( 2024 ). OpenUrl CrossRef 96. ↵ Silvan , A. , Parra , L. C. & Madsen , J . Real-Time Estimation of Overt Attention from Dynamic Features of the Face Using Deep Learning. in 2025 47th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) 1 – 5 ( 2025 ). doi: 10.1109/EMBC58623.2025.11254013 . OpenUrl CrossRef 97. ↵ Conway , K. P. et al. Association between COVID-19 risk-mitigation behaviors and specific mental disorders in youth . Child Adolesc. Psychiatry Ment. Health 17 , 14 ( 2023 ). 98. Restrepo , A. et al. Problematic internet use in children and adolescents: associations with psychiatric disorders and impairment . BMC Psychiatry 20 , 252 ( 2020 ). 99. ↵ Kramer , E. et al. Diagnostic Associations of Processing Speed in a Transdiagnostic, Pediatric Sample . Sci. Rep . 10 , 10114 ( 2020 ). 100. ↵ Zuromski , K. L. et al. Detecting suicide risk among U.S. servicemembers and veterans: a deep learning approach using social media data . Psychol. Med . 54 , 3379 – 3388 ( 2024 ). OpenUrl View the discussion thread. Back to top Previous Next Posted February 25, 2026. Download PDF Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Large-Scale Assessment of Language, Speech, and Movement in Autism and ADHD with AI Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Large-Scale Assessment of Language, Speech, and Movement in Autism and ADHD with AI Aimar Silvan , Lucas C. Parra , Adriana Di Martino , Michael P. Milham , Jens Madsen bioRxiv 2025.10.20.682864; doi: https://doi.org/10.1101/2025.10.20.682864 Share This Article: Copy Citation Tools Large-Scale Assessment of Language, Speech, and Movement in Autism and ADHD with AI Aimar Silvan , Lucas C. Parra , Adriana Di Martino , Michael P. Milham , Jens Madsen bioRxiv 2025.10.20.682864; doi: https://doi.org/10.1101/2025.10.20.682864 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Bioengineering Subject Areas All Articles Animal Behavior and Cognition (7633) Biochemistry (17680) Bioengineering (13888) Bioinformatics (41927) Biophysics (21445) Cancer Biology (18585) Cell Biology (25491) Clinical Trials (138) Developmental Biology (13373) Ecology (19897) Epidemiology (2067) Evolutionary Biology (24308) Genetics (15606) Genomics (22494) Immunology (17736) Microbiology (40385) Molecular Biology (17175) Neuroscience (88583) Paleontology (666) Pathology (2830) Pharmacology and Toxicology (4822) Physiology (7641) Plant Biology (15149) Scientific Communication and Education (2045) Synthetic Biology (4293) Systems Biology (9822) Zoology (2271)","source_license":"CC-BY-4.0","license_restricted":false}