Autism Spectrum Disorder (ASD) affects 1 in 77 children in Italy, but the diagnostic process is slow and costly. As autistic individuals exhibit different gaze patterns from healthy controls in visual exploration of images and semantic interpretation, these are promising biomarkers to exploit in diagnosis. This study aims at developing a model to assist in the diagnosis of ASD using gaze data when static images are presented to the subjects. We first propose a set of features, each one motivated by psychological studies and findings. Then we apply a feature selection mechanism based on Boruta algorithm with SHAP values. Finally we use CatBoost to perform binary classification, and a strategy to optimize model hyperparameters using a multivariate Tree Parzen Estimator. We validated our model on the popular Saliency4ASD dataset, outperforming state of the art models tested with the same protocol by more than 3% in accuracy. We also provide an in-depth analysis of the feature importance and we show how these results are in line with the psychological literature.
Autism Spectrum Disorder Identification from Visual Exploration of Images
Marco Bolpagni;Francesco Setti
2023-01-01
Abstract
Autism Spectrum Disorder (ASD) affects 1 in 77 children in Italy, but the diagnostic process is slow and costly. As autistic individuals exhibit different gaze patterns from healthy controls in visual exploration of images and semantic interpretation, these are promising biomarkers to exploit in diagnosis. This study aims at developing a model to assist in the diagnosis of ASD using gaze data when static images are presented to the subjects. We first propose a set of features, each one motivated by psychological studies and findings. Then we apply a feature selection mechanism based on Boruta algorithm with SHAP values. Finally we use CatBoost to perform binary classification, and a strategy to optimize model hyperparameters using a multivariate Tree Parzen Estimator. We validated our model on the popular Saliency4ASD dataset, outperforming state of the art models tested with the same protocol by more than 3% in accuracy. We also provide an in-depth analysis of the feature importance and we show how these results are in line with the psychological literature.File | Dimensione | Formato | |
---|---|---|---|
978-3-031-43148-7_7.pdf
solo utenti autorizzati
Licenza:
Non specificato
Dimensione
945.75 kB
Formato
Adobe PDF
|
945.75 kB | Adobe PDF | Visualizza/Apri Richiedi una copia |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.