Machine learning accurately classifies age of toddlers based on eye tracking

Kirsten A. Dalrymple, Ming Jiang, Qi Zhao, Jed T. ElisonScientific 2019 · DOI

Abstract

How people extract visual information from complex scenes provides important information about cognitive processes. Eye tracking studies that have used naturalistic, rather than highly controlled experimental stimuli, reveal that variability in looking behavior is determined by bottom-up image properties such as intensity, color, and orientation, top-down factors such as task instructions and semantic information, and individual differences in genetics, cognitive function and social functioning. These differences are often revealed using areas of interest that are chosen by the experimenter or other human observers. In contrast, we adopted a data-driven approach by using machine learning (Support Vector Machine (SVM) and Deep Learning (DL)) to elucidate factors that contribute to age-related variability in gaze patterns. These models classified the infants by age with a high degree of accuracy, and identified meaningful features distinguishing the age groups. Our results demonstrate that machine learning is an effective tool for understanding how looking patterns vary according to age, providing insight into how toddlers allocate attention and how that changes with development. This sensitivity for detecting differences in exploratory gaze behavior in toddlers highlights the utility of machine learning for characterizing a variety of developmental capacities.

Introduction

In developmental psychology, eye-tracking serves as a critical window into the infant mind. Traditionally, these studies rely on "Areas of Interest" (AOIs)—predefined regions like "the face" or "the toy" manually selected by researchers. However, this approach introduces experimenter bias and may overlook subtle, complex features that drive a child's attention. This study sought to determine if a purely data-driven approach, utilizing Support Vector Machines (SVM) and Deep Learning (DL), could identify age-related differences in how toddlers explore complex, naturalistic scenes. The goal was to move beyond asking if toddlers look at different things to understanding what specific hierarchical image features—from raw pixels to semantic meaning—distinguish developmental stages.

Methods and Data

The researchers employed a two-pronged computational strategy. First, a linear SVM was used to weight the importance of 12 predefined semantic attributes (e.g., faces, touched objects, "watchability"). Second, a Deep Learning (DL) pipeline based on the SALICON architecture was trained to predict a "difference map"—the subtraction of the 18-month-old fixation map from the 30-month-old map. Finally, features extracted from the DL model were fed into a linear SVM to classify individual toddlers by age.

The study utilized the OSIE (Object and Semantic Images Eye-tracking) Database, selecting 100 naturalistic images. The final analysis included 41 toddlers: 19 eighteen-month-olds and 22 thirty-month-olds, after rigorous data quality filtering.

Results

The DL + SVM model achieved a high classification accuracy of 83% (AUC = 0.90), significantly outperforming a simple gaze-based linear SVM (68% accuracy). Differences in looking patterns emerged rapidly, often within the first 1 to 3 fixations.

Figure 1: Stimuli examples, fixation maps for 18- and 30-month-olds, difference maps, and quantitative comparisons of fixation number and duration.

Quantitatively, 30-month-olds exhibited a more rapid, exploratory scanning strategy, making significantly more fixations with shorter durations compared to 18-month-olds (see Figure 1B and 1C). 18-month-olds averaged ~5.96 fixations per 3-second trial, while 30-month-olds averaged ~6.81.

Developmental Shifts in Social Attention

A key finding was the shift from "face-centric" attention to "intentionality." While 18-month-olds prioritize faces early and often, 30-month-olds have begun to process more complex social cues, such as objects being gazed at or touched by others.

Figure 2: Probability of fixating on "TOUCHED" objects as a function of fixation order for different age groups.

This transition suggests that by 30 months, toddlers have "unlocked" more complex social processing. The differences between age groups were driven primarily by high-level semantic features (who is doing what) rather than low-level "bottom-up" features like color or brightness.

Conclusion

This research demonstrates that machine learning is an effective tool for understanding developmental changes in attention. While the study was limited by static images and a relatively small sample size, the high sensitivity for detecting differences in exploratory gaze behavior highlights the potential for clinical applications, such as the early detection of developmental delays like Autism Spectrum Disorder (ASD). Future work will likely extend this approach to dynamic video stimuli to capture spatio-temporal features of social interaction.

Summary updated

Abstract

How people extract visual information from complex scenes provides important information about cognitive processes. Eye tracking studies that have used naturalistic, rather than highly controlled experimental stimuli, reveal that variability in looking behavior is determined by bottom-up image properties such as intensity, color, and orientation, top-down factors such as task instructions and semantic information, and individual differences in genetics, cognitive function and social functioning. These differences are often revealed using areas of interest that are chosen by the experimenter or other human observers. In contrast, we adopted a data-driven approach by using machine learning (Support Vector Machine (SVM) and Deep Learning (DL)) to elucidate factors that contribute to age-related variability in gaze patterns. These models classified the infants by age with a high degree of accuracy, and identified meaningful features distinguishing the age groups. Our results demonstrate that machine learning is an effective tool for understanding how looking patterns vary according to age, providing insight into how toddlers allocate attention and how that changes with development. This sensitivity for detecting differences in exploratory gaze behavior in toddlers highlights the utility of machine learning for characterizing a variety of developmental capacities.
BibTeX
@inproceedings{dalrymple2019machine,
  author = {Kirsten A. Dalrymple and Ming Jiang and Qi Zhao and Jed T. Elison},
  title = {{Machine learning accurately classifies age of toddlers based on eye tracking}},
  booktitle = {Scientific},
  year = {2019},
  doi = {10.1038/s41598-019-42764-z},
}
Figure