In this project, a noval method is proposed based the self-similar property of pupillary dynamics for early detection of ADHD.
- Self-similarity is characterized in the wavelet domain by computing wavelet spectra.
- A distance variance based approach is proposed to compute wavelet spectra as it is more robust to noise and outliers.
- A set of localized discriminatory features are constructed and selected via a rolling window method to build classifiers.
- ADHD detection performance is evaluated by using three classifiers, Logistic Regression, Support Vector Machine, and k-nearrest neighbor.
- To compare performance of the proposed method, a feature engineering technique that uses features constructed in the original pupil diameter data domain is considered. More details about the feature engineering method can be found in Das et al., A Robust MAchine Learning based Framework for the Automated Detection of ADHD Using the Pupillometric Biomarkers and Time Series Analysis, Scientific Reports, 2021(https://www.nature.com/articles/s41598-021-95673-5).
Below a brief overview of the pupil data analysis procedure is presented.
-
The dataset (pupil_dataset.mat) used in this project is available at https://figshare.com/articles/dataset/ADHD_Pupil_Size_Dataset/7218725/3.
-
The dataset consists of pupil diameter data collected from 50 children (28 cases and 22 control) during a visuospatial working memory task. The dataset contains three pupil diameter data groups, namely ADHD-diagnosed children with and without medication and normal children.
-
More information about the dataset is given in the study, Rojas-Líbano et al, A pupil size, eye-tracking and neuropsychological dataset from ADHD children during a cognitive task, Scientific Data, 2019.
-
This project considered pupil diameter of the AHDHD-diagnosed children without medication and normal children.
MatlabFunctions folder contains matlab functions used in the project. You can run the project as follows.
-
ReadAndSaveData.m computes feature matrixes for self-similarity and feature engineering methods.
-
Demo_1.m is for computing the ROC curve and distribution of Hurst exponent for cases and controls.
-
Demo_2.m evaluating classification performance of the proposed self-similarity and feature engineering methods.