Intra-topic latency as an automated behavioral marker of treatment response in autism spectrum disorder

Elizabeth P. McKernan, Manoj Kumar, Adriana Di Martino, Lisa Shulman, Alexander Kolevzon, Catherine Lord, Shrikanth Narayanan, So Hyun KimScientific 2022 · DOI

Abstract

Data science advances in behavioral signal processing and machine learning hold the promise to automatically quantify clinically meaningful behaviors that can be applied to a large amount of data. The objective of this study was to identify an automated behavioral marker of treatment response in social communication in children with autism spectrum disorder (ASD). First, using an automated computational method, we successfully derived the amount of time it took for a child with ASD and an adult social partner (N pairs = 210) to respond to each other while they were engaged in conversation bits ("latency") using recordings of brief, natural social interactions. Then, we measured changes in latency at pre- and post-interventions. Children with ASD who were receiving interventions showed significantly larger reduction in latency compared to those who were not receiving interventions. There was also a significant group difference in the changes in latency for adult social partners. Results suggest that the automated measure of latency derived from natural social interactions is a scalable and objective method to quantify treatment response in children with ASD.

Introduction

Evaluating the effectiveness of treatments for Autism Spectrum Disorder (ASD) has historically been a slow, subjective process. Clinicians typically rely on manual behavioral coding, parent reports, or therapist observations, all of which are labor-intensive or prone to bias. Furthermore, many existing metrics are not sensitive enough to capture short-term changes in core social communication symptoms. There is a critical need for automated behavioral markers that are objective, scalable, and precise enough to capture the subtle temporal dynamics of human interaction—details that are nearly impossible for a human coder to measure accurately with a stopwatch.

Methods

The study utilized a dataset of 210 children and adolescents with ASD (ages 4–17) across four clinical sites. Participants were divided into a Treatment Group (n=105) receiving social skills interventions and a Treatment-as-Usual (TAU) Group (n=105).

Researchers developed an automated speech processing pipeline to analyze recordings from the Brief Observation of Social Communication Change (BOSCC), a 12-minute naturalistic social interaction. The pipeline consists of several sophisticated modules:

  1. Voice Activity Detection (VAD): Distinguishes speech from background noise using deep neural networks.
  2. Speaker Clustering & Role Assignment: Uses x-vectors to cluster audio and assigns roles (child vs. examiner) based on question frequency.
  3. Topic Segmentation: Segments transcripts into "topics" using dynamic programming and word embeddings.
  4. Latency Calculation: Measures "intra-topic latency"—the duration of non-speech regions between speaker turns within a specific topic.
The automated speech processing pipeline for measuring intra-topic latency, including voice activity detection, speaker clustering, and topic segmentation.

Results

The primary finding was a significant divergence in conversational response times between the groups. Children in the treatment group showed a significant reduction in latency from pre- to post-intervention ($p=0.042$), with average response times dropping from 2.94s to 2.69s. Conversely, the TAU group showed a significant increase in latency ($p=0.001$), moving from 2.63s to 3.05s.

Changes in intra-topic latency from pre- (Time 1) to post-intervention (Time 2) for children (a) and examiners (b).

A fascinating "reciprocity effect" was observed: examiners' behavior changed alongside the children. In the TAU group, examiners increased their latency over time, likely reflecting a dynamic adaptive exchange where the adult settles into a slower rhythm with a non-responsive partner. Notably, intra-topic latency correlated with social-communication severity but was independent of IQ, suggesting it is a "pure" marker of social coordination rather than general processing speed.

Discussion

This automated temporal measure proved sensitive enough to detect improvements over a relatively short 3-to-4-month period, whereas traditional severity scores often remain stable. However, limitations include the non-randomized design and the requirement for fluent phrase speech, which excludes minimally verbal individuals. Future work will focus on validating these tools in large-scale randomized controlled trials and expanding the feature set to include non-verbal cues like gaze and facial expressions.

Summary updated

Abstract

Data science advances in behavioral signal processing and machine learning hold the promise to automatically quantify clinically meaningful behaviors that can be applied to a large amount of data. The objective of this study was to identify an automated behavioral marker of treatment response in social communication in children with autism spectrum disorder (ASD). First, using an automated computational method, we successfully derived the amount of time it took for a child with ASD and an adult social partner (N pairs = 210) to respond to each other while they were engaged in conversation bits ("latency") using recordings of brief, natural social interactions. Then, we measured changes in latency at pre- and post-interventions. Children with ASD who were receiving interventions showed significantly larger reduction in latency compared to those who were not receiving interventions. There was also a significant group difference in the changes in latency for adult social partners. Results suggest that the automated measure of latency derived from natural social interactions is a scalable and objective method to quantify treatment response in children with ASD.
BibTeX
@inproceedings{mckernan2022intra,
  author = {Elizabeth P. McKernan and Manoj Kumar and Adriana Di Martino and Lisa Shulman and Alexander Kolevzon and Catherine Lord and Shrikanth Narayanan and So Hyun Kim},
  title = {{Intra-topic latency as an automated behavioral marker of treatment response in autism spectrum disorder}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-07299-w},
}
Figure