- Derick A. Lagunes-Ramírez
- Gabriel González-Serna
- Leonor Rivera-Rivera
- Nimrod González-Franco
- María Y. Hernández-Pérez
- José A. Reyes-Ortiz
This repository contains the files needed to replicate the results presented in the paper Through the Youth Eyes: Training Depression Detection Algorithms with Eye Tracking Data.
This dataset file in CSV format contains data from eye-tracking tests involving 172 young participants aged 15 to 21 years (mean age = 16.53) randomly selected from two high schools in Mexico. After excluding 33 participants based on inclusion criteria (age between 15 and 21, no medication use, no eye-related issues, and consent to participate), the dataset comprises 139 participants. They are classified into two groups based on their DASS-21 depression scale scores: Healthy Control (n=82) and Depressed (n=57).
The data were cleaned, and the most significant features were selected using U-tests. To balance classes, data augmentation techniques were applied. The final dataset includes 21 selected features, including:
- Pupil size metrics: Maximum values, skewness, and kurtosis for both the left and right pupils.
- Eye fixation metrics: Fixation count, variance and standard deviation of both X and Y coordinates, maximum and minimum values of the X coordinate (side-to-side view), and mean, skewness, and kurtosis of the Y coordinate (up-and-down view).
- Additional metrics: Blink count, blink rate per minute, saccade velocity, and saccade latency.
The target variable Class (0 for the control group, 1 for the depressed group) is also included, resulting in a dataset with 228 data vectors (114 for each group).
This dataset was used to train supervised machine learning models for binary classification, aiming to predict whether a participant is classified as "depressed" or "healthy control" based on their eye-tracking data.
This Python 3 script implements and evaluates four machine learning algorithms using the selectedFeaturesDataset.csv dataset.
The following machine learning algorithms were chosen for this study:
- Support Vector Machine (SVM): Selected for its strong performance in high-dimensional spaces.
- Random Forest (RF): Known for its robustness and ability to reduce overfitting.
- Multilayer Perceptron (MLP): A neural network model chosen for its ability to capture complex, non-linear patterns.
- Gradient Boosting (GB): Selected for its high accuracy and sequential error-correcting nature.
These models were evaluated and compared in terms of accuracy, generalization, and interpretability.
Hyperparameter tuning was performed using GridSearchCV from Scikit-learn to find the optimal parameters for each algorithm. The dataset was split into training and testing sets (80% for training, 20% for testing), and stratified sampling was used to maintain the same class proportions in both sets.
A portion of the training data was set aside for cross-validation. The best hyperparameters were selected based on training set performance, and the final model was trained with the optimal parameters and evaluated on the testing set.
The fine-tuned hyperparameters for each algorithm are as follows:
-
SVM:
- Regularization parameter ( C = 0.1 )
- Polynomial kernel, degree = 4
- Coefficient ( \text{coef0} = 1 )
- Gamma = 0.1
-
Random Forest:
- max_depth = None (unlimited)
- min_samples_leaf = 1
- min_samples_split = 2
- n_estimators = 200
- random_state = 11
-
Gradient Boosting:
- Learning rate = 0.1
- max_depth = 3
- min_samples_split = 3
- n_estimators = 200
-
MLP:
- Batch size = 16
- Epochs = 30
- Dropout rate = 0.25
- Hidden layer sizes = (64, 32, 16, 8)
- Optimizer = rmsprop
- Learning rate = 0.001
To use the script, ensure you have the necessary Python libraries specified at the start of the .py file, then execute the script. After processing, two .txt files will be generated, reporting the results of all ML algorithms with performance metrics, including accuracy, recall, F1 score, and MCC.