TL;DR
This paper introduces a novel feature selection pipeline (PCA-mFCBF) combined with automated machine learning (AutoML) to accurately detect negative mood states (depression, anxiety, stress) from handwriting and drawing data. The approach achieves near-perfect accuracy for binary mood detection and significantly improves upon existing methods for trinary classification, offering a promising tool for mental health monitoring.
What Problem Does This Paper Address?
Mental health disorders like depression, anxiety, and stress significantly impact individuals' well-being and are a major public health concern. Early detection and timely intervention are crucial for effective treatment and improved quality of life. While traditional methods rely on self-report questionnaires or clinical observation, this research explores the potential of behavioral biometrics, specifically analyzing the nuances of human handwriting and drawing patterns captured by digital tablets, as a non-invasive and objective indicator of these mood states. The challenge lies in extracting meaningful information from complex sensor data and developing robust models that can accurately classify these subtle emotional states.
Methods
The core of this research lies in a two-pronged approach: advanced feature engineering and selection, and automated machine learning.
- Feature Extraction: The study extracts a rich set of features from tablet sensor data, encompassing:
- Temporal Features: Related to time spent on/off the tablet, duration of strokes, etc.
- Kinematic Features: Describing pen movement, velocity, acceleration, and jerk.
- Statistical Features: Including range, median, mode, and standard deviation of various signals.
- Spectral Domain Features: Analyzing frequency components of the signals.
- Cepstral Domain Features: Derived from the log-energy of spectral features.
- Feature Selection Pipeline (PCA-mFCBF): To manage the high dimensionality of extracted features and identify the most discriminative ones, a novel pipeline is proposed:
- Principal Component Analysis (PCA): Used to orthogonalize the features, reducing dimensionality and decorrelating them. This step aims to capture the maximum variance in the data.
- Modified Fast Correlation-Based Filtering (mFCBF): Applied to the PCA-transformed features. Unlike traditional PCA where the first few components are selected, mFCBF selects principal components (PCs) that have a high correlation with the output (mood state), even if they are not the components explaining the most variance. This is a key innovation, aiming to balance dimensionality reduction with discriminative power.
- Automated Machine Learning (AutoML): The selected features are then fed into an AutoML framework (specifically H2O AutoML). This approach automates the process of model selection, hyperparameter tuning, and ensemble building, allowing for the rapid evaluation of numerous classification algorithms to find the best-performing model.
- Data Augmentation: To address the class imbalance in the dataset, a data augmentation technique was employed. This involved synthetically increasing the number of observations for under-represented mood states by adding Gaussian random noise to the extracted features.
Datasets
The primary dataset used in this research is the EMOTHAW database. This database contains handwriting and drawing data from 129 participants, along with their mood state assessments using the Depression, Anxiety, and Stress Scale (DASS) questionnaire. The study utilizes both binary (normal/above normal) and trinary (normal/mild/moderate) classifications of mood states derived from the DASS scores.
Key Insights and Ideas
- Handwriting as a Window to Mood: The study strongly reinforces the idea that subtle changes in handwriting and drawing patterns can serve as reliable indicators of a person's emotional state, particularly negative moods.
- PCA-mFCBF Synergy: The proposed PCA-mFCBF pipeline is a significant contribution. By orthogonalizing features with PCA first and then using mFCBF to select the most relevant principal components (based on correlation with the output, not just variance), the method achieves superior feature selection compared to using PCA or mFCBF alone. This suggests that focusing on discriminative power over pure variance explanation is key for this task.
- AutoML's Power for Complex Tasks: Leveraging AutoML significantly boosted performance. Automating the model selection and tuning process allowed for the exploration of a vast model space, leading to higher accuracy than traditional manual model selection.
- Kinematic and Statistical Features Matter: The inclusion of kinematic and statistical features, in addition to temporal, spectral, and cepstral features, demonstrably improved the detection accuracy, highlighting the importance of a comprehensive feature set.
- Binary vs. Trinary Classification: While the binary classification (normal vs. abnormal) achieved near-perfect accuracy, the trinary classification (normal, mild, moderate) showed a notable drop in performance, indicating the increased difficulty in distinguishing finer gradations of mood states.
Key Results
- Binary Classification: The PCA-mFCBF pipeline combined with AutoML achieved 100% accuracy in detecting two possible grades of mood severities for depression, anxiety, and stress. This is a remarkable result, significantly outperforming previous state-of-the-art methods.
- Trinary Classification: For detecting three mood states (normal, mild, moderate), the PCA-mFCBF pipeline with AutoML achieved accuracies of 82.5% for depression, 72.8% for anxiety, and 74.56% for stress. While lower than binary classification, these results are still highly competitive and represent a substantial improvement over baseline methods.
- Feature Set Impact: Adding kinematic and statistical features improved accuracy by up to 7.9% for binary classification.
- Pipeline Effectiveness: The PCA-mFCBF pipeline consistently outperformed using PCA or mFCBF individually for feature selection.
Limitations and Future Work
- Explainability Drawback: The authors acknowledge that while PCA helps minimize features and improve accuracy, it can negatively impact the model's explainability. Translating the selected principal components back into easily interpretable clinical insights remains a challenge.
- Trinary Classification Performance: The accuracy for trinary classification, while improved, is not as impressive as for binary classification. This suggests that distinguishing between subtle mood gradations is more challenging and may require further refinement of features or models.
- Dataset Size and Diversity: The EMOTHAW database, while valuable, has 129 participants. Future work could benefit from larger and more diverse datasets to further validate the findings and improve generalization.
- Comorbid Mood States: The study did not analyze comorbid mood states due to insufficient sample sizes. Investigating the detection of co-occurring mental health conditions would be a valuable future direction.
- Real-time Application: While the system is designed for online data capture, further research into real-time deployment and user experience would be beneficial.
Summary updated