Large-Scale Assessment of Language, Speech, and Movement in Autism and ADHD with AI

This study demonstrates that large-scale AI analysis of language, speech, and movement in clinical interviews can effectively disentangle the overlapping symptoms of Autism and ADHD by identifying distinct behavioral signatures, such as narrative divergence and vocal pitch for Autism versus specific motor hyperactivity for ADHD.

The Gist

Imagine you are trying to tell the difference between two very similar-looking twins, Autism (ASD) and ADHD. For a long time, doctors and scientists have struggled to separate them because they often show up together, and their symptoms can look like a messy overlap. It's like trying to hear two different songs playing on the same radio station; the music blends together, and it's hard to tell which melody belongs to which song.

This paper is like a high-tech audio engineer stepping in with a massive soundboard to separate those tracks. The researchers used Artificial Intelligence (AI) to analyze video and audio recordings of over 2,300 children and young adults (ages 5 to 22) during a simple interview.

Here is the story of what they found, explained in everyday terms:

1. The Setup: A Movie and a Chat

Instead of a boring medical test, the kids watched a short, emotional cartoon called "The Present." It's a story about a boy who gets a puppy with three legs, but the boy himself also has a missing leg. After watching, a friendly interviewer asked the kids simple questions:

• "What happened in the movie?"
• "How did the puppy feel?"
• "How did you feel?"

The researchers didn't just listen to what the kids said; they used AI to measure how they said it, how they moved, and how they sounded.

2. The Big Revelation: It's Not About "Bad" Language

For years, people thought kids with ADHD had trouble with language or storytelling. But this study found that this isn't actually true.

• The Analogy: Imagine a young child trying to tell a story. They might use simple words or get distracted. Now imagine a teenager telling the same story. They use bigger words and stay on topic.
• The Finding: The study showed that when you account for age and intelligence (IQ), kids with ADHD speak just as well as anyone else. The "language problems" people saw before were mostly just because the kids were younger or because they also had Autism. ADHD itself doesn't break the language machine; it just makes it harder to sit still long enough to use it perfectly in a test.

3. The Autism Signature: The "Storyteller" and the "Voice"

While ADHD didn't change the language structure, Autism did leave a very specific fingerprint in two areas:

• The Storyteller (Perspective Taking):
  • The Metaphor: Imagine a movie director. A neurotypical person (or someone with just ADHD) can look at the movie and say, "The boy and the puppy are alike because they both lost a leg." That's seeing the big picture and understanding how others feel.
  • The ASD Finding: Kids with Autism often missed this "big picture." They might remember the facts (the puppy had a ball) but struggle to connect the emotional dots or explain why the characters felt a certain way. Their stories were unique, but they often lacked the shared "human connection" that others naturally make.
• The Voice (The Soundtrack):
  • The Metaphor: Think of a radio station. Most people speak with a steady volume and pitch.
  • The ASD Finding: Kids with Autism had a distinct "voice print." Their voices were often higher pitched, louder, and more variable (jumping up and down in tone). They also sounded a bit "breathier" or "hoarse." Crucially, kids with only ADHD did not have this voice signature. This is a new, objective way to spot Autism just by listening to how they talk.

4. The ADHD Signature: The "Wiggly" Factor

If Autism is about the story and the voice, ADHD is about the body.

• The Analogy: Imagine a car engine. A car with ADHD-Hyperactivity has an engine that is revving too high. It's constantly vibrating, shaking, and moving.
• The Finding: The AI tracked every tiny movement of the kids' faces, eyes, and bodies. They found that kids with the Hyperactive type of ADHD moved significantly more than everyone else. They fidgeted, shifted their eyes, and moved their heads constantly.
• The Twist: Kids with the Inattentive type of ADHD (the "daydreamers") and kids with Autism did not show this extra movement. This means "wiggling" is a specific sign of Hyperactive ADHD, not just "being a kid" or "being autistic."

5. Why This Matters

Think of the old way of diagnosing these conditions like trying to sort a pile of mixed-up Legos by color, but the colors are all muddy and blended.

This study is like a magnetic sorter. It uses AI to find the specific "magnetic signature" of each condition:

• Autism = Unique voice + Trouble connecting the emotional dots in a story.
• ADHD (Hyperactive) = Constant body movement.
• ADHD (Inattentive) = No specific movement or voice signature found in this study (they were more like the "control" group in terms of movement).

The Bottom Line

The researchers proved that we can use simple video interviews and AI to untangle these two conditions. We don't need to rely on guesswork or subjective opinions anymore.

• If a kid is wiggling a lot, it points to ADHD-Hyperactivity.
• If a kid has a unique, high-pitched voice and struggles to connect the emotional dots in a story, it points to Autism.
• If a kid seems to have "language issues," it's likely just because they are young or have Autism, not because they have ADHD.

This is a huge step forward because it means we can give kids a more accurate diagnosis, leading to better support that actually fits their specific needs, rather than treating a blurry mix of symptoms.

Technical Summary

1. Problem Statement

Attention-Deficit/Hyperactivity Disorder (ADHD) and Autism Spectrum Disorder (ASD) frequently co-occur, sharing genetic factors and overlapping symptoms, which complicates differential diagnosis. Current diagnostic methods rely heavily on subjective clinical observation and parent/teacher reports, which can be inconsistent.

• The Gap: Previous studies using automated behavioral analysis have been limited by small sample sizes (often <50 participants), analyzed conditions in isolation, and failed to disentangle the effects of developmental maturation (age), cognitive ability (IQ), and sex from true psychopathology.
• The Goal: To determine if large-scale, automated analysis of naturalistic clinical interviews can identify distinct, objective behavioral signatures that separate ASD and ADHD, even when they co-occur.

2. Methodology

Data Source & Cohort

• Dataset: The study utilized the Healthy Brain Network (HBN) cohort, a large community sample of 2,341 youths (ages 5–22).
• Diagnoses: The cohort included clinician-confirmed diagnoses of ADHD (Inattentive, Hyperactive, and Combined presentations) and ASD. Notably, 19.5% of the ADHD group had co-occurring ASD.
• Task: Participants watched a short, emotive animated film ("The Present") and underwent a semi-structured clinical interview. The interview probed narrative recall, factual memory, emotional interpretation, and perspective-taking (Theory of Mind).

AI Pipeline & Feature Extraction
The researchers employed a state-of-the-art computational pipeline to extract objective metrics from audio-video recordings:

1. Transcription & Diarization: Used WhisperX for transcription and a custom Large Language Model (LLM) prompt (Google Gemini) for speaker diarization (separating interviewer vs. subject). The LLM approach significantly outperformed acoustic-based methods in handling pitch overlap and context.
2. Language Analysis:
   • Structural: Extracted metrics like lexical diversity (MATTR), speech rate, and filler word usage using the OpenWillis toolkit.
   • Semantic/Pragmatic: Used Google's Gecko text embedding model to convert answers into 768-dimensional vectors. "Answer typicality" was calculated by comparing a participant's response to the median embedding of age-matched Typically Developing (TD) peers.
3. Vocal Prosody: Extracted acoustic features (pitch, intensity, breathiness, dysphonia) using Parselmouth (Praat tool).
4. Motor Movement: Used Google's MediaPipe Holistic computer vision model to track 3D anatomical and facial landmarks. Movement was quantified as frame-to-frame Euclidean displacement for the face, eyes, mouth, and body.

Statistical Modeling

• Multivariate Regression: The core analytical approach modeled behavioral metrics as a function of:
  $$Metric \sim Age + IQ + Sex + ADHD_{Inatt} + ADHD_{Hyper} + ASD$$
• Key Innovation: Instead of treating co-occurring ADHD+ASD as a single categorical group, the study modeled ADHD-Inattentive, ADHD-Hyperactive, and ASD as independent additive binary variables. This allowed the isolation of unique variance contributed by each dimension while controlling for developmental (age) and cognitive (IQ) confounders.

3. Key Contributions

1. Scale and Generalizability: This is the largest study of its kind (N=2,341), moving beyond the small-sample limitations of previous AI-based behavioral phenotyping studies.
2. Disentangling Overlap: The study successfully separated the behavioral signatures of ASD and ADHD, demonstrating that many previously attributed "ADHD language deficits" are actually driven by age, IQ, or co-occurring ASD.
3. Dimensional Modeling: By treating diagnoses as overlapping dimensions rather than mutually exclusive categories, the study provides a more nuanced view of neurodevelopmental phenotypes.
4. Open Science: The authors released the full code pipeline for transcription, diarization, and feature extraction, enabling reproducibility.

4. Key Results

A. Language: Development Dominates Diagnosis

• Finding: Structural language abilities (vocabulary, syntax, speech rate) are overwhelmingly driven by Age and IQ, not by ADHD or ASD status.
• Nuance: Once age and IQ were controlled, ADHD showed no significant independent association with structural language deficits.
• ASD Specificity: The only significant language marker for ASD was an increase in self-referential speech (first-person pronouns), suggesting an asynchrony between linguistic and social development.

B. Social Cognition & Narrative: ASD-Specific Deficits

• Finding: ASD was uniquely characterized by divergent narrative production and perspective-taking (Theory of Mind).
• Metrics: Participants with ASD produced less "typical" narratives, failed to grasp overarching thematic "gists" (e.g., the shared disability between the boy and puppy), and provided idiosyncratic emotional descriptions.
• ADHD: Neither ADHD-Inattentive nor ADHD-Hyperactive statuses were independently associated with deficits in narrative coherence or perspective-taking once ASD and age were controlled.

C. Vocal Prosody: Distinct ASD Signature

• Finding: ASD exhibited a distinct vocal profile independent of ADHD.
• Metrics: ASD was associated with higher mean pitch, greater pitch variability, increased vocal loudness, dysphonia, and breathiness.
• ADHD: Neither ADHD presentation showed significant alterations in vocal prosody.

D. Motor Activity: Specific to ADHD-Hyperactivity

• Finding: Increased motor activity was a robust, specific marker for ADHD-Hyperactive status.
• Metrics: ADHD-Hyperactive youth showed significantly increased movement across all measured regions (face, eyes, mouth, body).
• Specificity: Neither ASD nor ADHD-Inattentive status showed significant global kinematic differences. This suggests that "restlessness" is a specific marker for the hyperactive-impulsive domain, not a general feature of neurodivergence or ASD.

5. Significance and Implications

• Refining Diagnostic Criteria: The study challenges the notion that ADHD inherently involves structural language deficits, suggesting these are often confounded by developmental factors or co-occurring ASD. Conversely, it validates that social-cognitive and narrative deficits are core to ASD, even in those with high verbal ability.
• Digital Biomarkers: The research demonstrates that brief, standardized clinical interviews can be transformed into scalable, multidimensional digital phenotypes. These AI-derived measures offer higher specificity than traditional questionnaire-based instruments (e.g., distinguishing ADHD-Hyperactive motor restlessness from general anxiety or ASD-related movement).
• Clinical Utility: While not intended to replace clinical diagnosis, these tools can assist clinicians in identifying specific behavioral dimensions (e.g., distinguishing a child with pure ADHD from one with ASD+ADHD) and tracking developmental trajectories.
• Future Directions: The authors advocate for longitudinal validation of these measures and their integration into standardized assessments like the ADOS, as well as expansion to other neurological and psychiatric conditions.

In summary, this paper establishes that scalable AI analysis of naturalistic interactions can effectively disentangle the overlapping phenotypes of ASD and ADHD, revealing that motor hyperactivity is specific to ADHD, while narrative and prosodic anomalies are specific to ASD, with structural language largely driven by normal development.