Abstract
Antenatal depression (AD) is a significant public health concern, affecting maternal well-being and child development. Early detection is crucial but often hampered by the high dimensionality and inherent class imbalance of clinical questionnaire data. This paper presents a novel machine learning framework, Stacked Auto Encoder Support Vector Data Description (SAE-SVDD), designed to predict AD by treating it as an anomaly detection task. Validated on a dataset of 5,666 pregnant women, the model achieved an F-score of 87.96%, outperforming traditional classifiers. Furthermore, the study identifies key socio-environmental risk factors, such as work stress and marital status, through an improved TF-IDF analysis.
Introduction
In modern society, the prevalence of mental disorders is rising due to increased psychological pressure. Among these, antenatal depression (AD) affects between 10% and 30% of pregnant women. AD is not only a primary predictor of postpartum depression but also correlates with adverse physical health outcomes for the mother and developmental issues for the child.
While screening questionnaires are widely used, they present two significant computational challenges:
- High Dimensionality: Modern surveys collect a vast array of features (147 in this study), many of which are redundant or noisy.
- Data Imbalance: In clinical screening, the "depressed" group is significantly smaller than the "healthy" group. Standard binary classifiers often struggle with this "needle in a haystack" scenario, frequently biasing toward the majority class.
Methodology
The researchers developed the SAE-SVDD model, a hybrid architecture combining unsupervised feature extraction with a self-adaptive one-class classifier. The workflow consists of two primary phases:
1. Feature Embedding via Stacked Auto Encoders (SAE)
To manage the 147 raw features, the model employs a Stacked Auto Encoder. This unsupervised neural network learns to compress input data into a lower-dimensional "latent" space. This acts as a sophisticated filter, retaining only the most informative patterns while stripping away noise.
2. One-Class Classification (SA-SVDD)
Rather than training on both "healthy" and "depressed" labels simultaneously, the model uses Support Vector Data Description (SVDD). This technique constructs a "hypersphere" (a multi-dimensional bubble) around the majority (healthy) data. Any data point falling outside this boundary is flagged as an anomaly—in this context, a patient at risk for depression. To handle complex data distributions, the authors integrated Affinity Propagation (AP) clustering and Particle Swarm Optimization (PSO) to fine-tune the classifier's parameters.
Key Insights and Ideas
The success of the SAE-SVDD model rests on its ability to distill complex psychological data into a manageable form.
Dimensionality Reduction: The study found that compressing 147 features down to 90 (or even 30) dimensions did not degrade performance. This indicates that psychological surveys often contain significant redundancy. By focusing on a compressed feature set, the model becomes more efficient and less prone to overfitting.
Anomaly Detection for Imbalance: By framing depression as an "anomaly" relative to a healthy baseline, the researchers bypassed the common pitfalls of training on imbalanced data. This makes the model more robust for real-world clinical screening where healthy patients vastly outnumber those at risk.
Key Results
The model was first validated on three benchmark datasets (Adult, Madelon, and Protein) to ensure generalizability. When applied to the clinical dataset of 5,666 pregnant women from Jilin Women and Children Hospital, SAE-SVDD achieved an F-score of 87.96%.
This performance significantly exceeded that of traditional SVDD (85.26%) and deep learning architectures like CNNs (56.60%) and LSTMs (48.14%), which likely overfitted on the imbalanced tabular data. Beyond classification, the researchers used an improved TF-IDF (Term Frequency-Inverse Document Frequency) analysis to rank the importance of questionnaire items. The top indicators for AD included:
- Work stress
- Marital status
- Husband support
- Passive smoking
- Alcohol consumption
Conclusion
The SAE-SVDD model provides a robust, efficient, and interpretable framework for screening antenatal depression. By successfully addressing the hurdles of high-dimensional and imbalanced data, it offers a practical tool for early clinical intervention. Future work will focus on reducing the computational overhead of the optimization phase and validating the model across diverse cultural cohorts to ensure the universality of the identified risk factors.
Summary updated