School of Arts, Sciences and Humanities (EACH) of the University of Sao Paulo (USP), Rua Arlindo Béttio, 1000 - Ermelino Matarazzo, São Paulo, 03828-000, São Paulo, Brazil.
Department of Psychiatry, University of Sao Paulo's School of Medicine (FMUSP), Rua Doutor Ovídio Pires de Campos, 785 - Cerqueira César, São Paulo, 05403-010, São Paulo, Brazil.
Comput Biol Med. 2024 Nov;182:109184. doi: 10.1016/j.compbiomed.2024.109184. Epub 2024 Sep 30.
Diagnosing Autism Spectrum Disorder (ASD) remains a significant challenge, especially in regions where access to specialists is limited. Computer-based approaches offer a promising solution to make diagnosis more accessible. Eye tracking has emerged as a valuable technique in aiding the diagnosis of ASD. Typically, individuals' gaze patterns are monitored while they view videos designed according to established paradigms. In a previous study, we developed a method to classify individuals as having ASD or Typical Development (TD) by processing eye-tracking data using Random Forest ensembles, with a focus on a paradigm known as joint attention.
This article aims to enhance our previous work by evaluating alternative algorithms and ensemble strategies, with a particular emphasis on the role of anticipation features in diagnosis.
Utilizing stimuli based on joint attention and the concept of "floating regions of interest" from our earlier research, we identified features that indicate gaze anticipation or delay. We then tested seven class balancing strategies, applied seven dimensionality reduction algorithms, and combined them with five different classifier induction algorithms. Finally, we employed the stacking technique to construct an ensemble model.
Our findings showed a significant improvement, achieving an F1-score of 95.5%, compared to the 82% F1-score from our previous work, through the use of a heterogeneous stacking meta-classifier composed of diverse induction algorithms.
While there remains an opportunity to explore new algorithms and features, the approach proposed in this article has the potential to be applied in clinical practice, contributing to increased accessibility to ASD diagnosis.
自闭症谱系障碍（ASD）的诊断仍然是一个重大挑战，尤其是在专家资源有限的地区。基于计算机的方法为实现更便捷的诊断提供了有前景的解决方案。眼动追踪已成为辅助 ASD 诊断的一种有价值的技术。通常，在个体观看根据既定范式设计的视频时，会监测他们的注视模式。在之前的研究中，我们开发了一种使用随机森林集成处理眼动追踪数据的方法，通过关注联合注意范式，将个体分类为 ASD 或典型发育（TD）。
本文旨在通过评估替代算法和集成策略来增强我们之前的工作，特别关注预测特征在诊断中的作用。
利用基于联合注意的刺激以及我们早期研究中的“浮动感兴趣区域”概念，我们确定了表示注视预测或延迟的特征。然后，我们测试了七种类别平衡策略、七种降维算法，并将它们与五种不同的分类器归纳算法相结合。最后，我们使用堆叠技术构建了一个集成模型。
与我们之前工作中 82%的 F1 分数相比，通过使用由不同归纳算法组成的异构堆叠元分类器，我们的研究结果显示出了显著的改进，达到了 95.5%的 F1 分数。
虽然仍有机会探索新的算法和特征，但本文提出的方法有可能应用于临床实践，有助于增加 ASD 诊断的可及性。
