The potential of evaluating shape drawing using machine learning for predicting high autistic traits

Abstract

Background: Children with high autistic traits often exhibit deficits in drawing, an important skill for social adaptability. Machine learning is a powerful technique for learning predictive models from movement data, so drawing processes and product characteristics can be objectively evaluated. This study aimed to assess the potential of evaluating shape drawing using machine learning to predict high autistic traits.

Method: Seventy boys (5.03 ± 0.16) and 63 girls (5.06 ± 0.18) from the general population participated in the study. Participants were asked to draw shapes in the following order: equilateral triangle, inverted equilateral triangle, square, and the sun. A model for classifying participants as likely to have high autistic traits was developed using a support vector machine algorithm with a linear kernel utilizing 16 variables. A 16-inch liquid crystal display pen tablet was used to acquire data on hand-finger fine motor activity while the participants drew each shape. The X and Y coordinates of the pen tip, pen pressure, pen orientation, pen tilt, and eye movements were recorded to determine whether the participants had any problems with this skill. Eye movements were assessed using a webcam. These data and eye movements were used to identify the variables for the support vector machine model.

Data and results: For each shape, a model support vector machine was created to classify the high and low autistic trait groups, with accuracy, sensitivity, and specificity all above 85%. The specificity values across all models were 100%. In the inverted equilateral triangle model, specificity, accuracy, and sensitivity values were 100%.

Conclusions: These results demonstrate the potential of assessing shape characteristics using machine learning to predict high levels of autistic traits. Future studies with a wider variety of shapes are warranted to establish further the potential efficacy of drawing skills for screening for autism spectrum conditions.

Fig 1. The distribution of the SRS scores in high autistic group and low autistic group.

Fig 2. Drawing samples of (a) an equilateral triangle, (b) an inverted equilateral triangle, (c) a square, and (d) the sun in the LCD pen tablet. The numbers for each shape indicate the stroke order for each line to be drawn. These shapes were presented to the participants, and the drawing process was done via video during the data collection. LCD, liquid crystal display.

Fig 3. Tilt and orientation angles captured by the Wacom Cintiq 16.

Fig 4. (a) The impact of each feature on the equilateral triangle model analyzed by SHAP. (b) The impact of each feature on the inverted equilateral triangle model analyzed by SHAP. (c) The impact of each feature on the square model analyzed by SHAP. (d) The impact of each feature on the sun model analyzed by SHAP.