An Explainable Hybrid of Classical and Quantum Support Vector Machine Models for Maternal Mental Health Risk Prediction Using Multicountry Data

Technical Summary: An Explainable Hybrid of Classical and Quantum Support Vector Machine Models for Maternal Mental Health Risk Prediction

Problem Statement
Maternal mental health disorders, particularly during pregnancy and the postpartum period, represent a critical global public health challenge, with prevalence rates significantly higher in low- and middle-income countries (LMICs). Despite the availability of effective interventions, underdiagnosis and delayed intervention persist due to limited resources, inadequate screening infrastructure, and a shortage of specialized personnel. While classical machine learning (ML) models have shown promise in predicting these risks using psychosocial and clinical data, they often struggle to capture highly complex nonlinear interactions present in heterogeneous, multicountry datasets. Furthermore, existing approaches rarely integrate quantum-enhanced learning or explainable artificial intelligence (XAI), limiting their interpretability and clinical trustworthiness. Additionally, standalone quantum machine learning models face performance constraints due to the limitations of current noisy intermediate-scale quantum (NISQ) hardware.

Methodology
The study proposes an Explainable Hybrid Classical-Quantum Support Vector Machine (SVM) model designed to leverage the discriminative power of classical algorithms and the nonlinear feature representation capabilities of quantum computing.

• Dataset: The research utilized a multicountry dataset comprising 14,008 maternal records. This integrated dataset included primary data from Ugandan hospitals (Kawempe National Referral Hospital and Mukono Hospital) and secondary data from a publicly available perinatal depression dataset from Pakistan. The data encompassed 51 variables covering demographics, obstetric history, socioeconomic factors, and psychosocial indicators (e.g., stress, hopelessness, suicidal ideation).
• Preprocessing: The pipeline involved data cleaning, handling missing values, and encoding categorical variables. To address class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied to the training set. Kernel Principal Component Analysis (KPCA) was employed for dimensionality reduction and to preserve nonlinear relationships prior to quantum feature embedding.
• Model Architecture:
  • Classical SVM: Utilized a Radial Basis Function (RBF) kernel to construct an optimal separating hyperplane.
  • Quantum SVM (QSVM): Implemented using the Pennylane library, this component mapped classical input features into a high-dimensional quantum Hilbert space using parameterized quantum circuits. A quantum kernel function ($K_Q$) was computed to measure similarity between samples in the quantum feature space.
  • Hybrid Integration: The final prediction was generated via a weighted ensemble approach, combining the probability outputs of the Classical SVM ($P_C$) and the Quantum SVM ($P_Q$) using a weighting coefficient ($\alpha$).
  • Explainability: SHapley Additive exPlanations (SHAP) were applied to the hybrid model to quantify the contribution of individual features to the risk prediction, ensuring clinical interpretability.
• Experimental Setup: Experiments were conducted in a Google Colaboratory environment using GPU acceleration. The dataset was split 80:20 for training and testing. The proposed model was benchmarked against Classical SVM, Random Forest, XGBoost, LightGBM, CatBoost, Stacking Ensemble, and standalone QSVM.

Key Results
The proposed Hybrid Classical-Quantum SVM demonstrated highly competitive performance, achieving:

• Accuracy: 99.86%
• F1-Score: 99.80%
• ROC-AUC: 99.92%
• PR-AUC: 99.91%

In comparative analysis, the hybrid model outperformed the standalone Quantum SVM, which achieved an accuracy of 77.82% and an ROC-AUC of 88.76%, highlighting the limitations of pure quantum models under current NISQ constraints. However, the hybrid approach achieved performance comparable to the best-performing classical models (Classical SVM and Stacking Ensemble, both at 99.93% accuracy).

SHAP Analysis Findings:
The explainability analysis identified key predictors contributing to maternal mental health risk, including:

• Suicidal ideation
• Stress levels
• Concentration difficulties
• Fatigue
• Poor appetite
• Feelings of hopelessness

Quantum kernel heatmap analysis confirmed that the quantum feature embedding successfully captured nonlinear patterns and hidden feature interactions within the dataset, despite the standalone model's lower overall classification accuracy.

Significance and Claims
The paper claims that the primary contribution of this study is the development of an explainable hybrid architecture that effectively combines classical discriminative learning with quantum-enhanced feature representation. The authors assert that this approach:

1. Enhances Robustness: By integrating quantum kernels with classical optimization, the model improves predictive stability and nonlinear separability in complex, heterogeneous healthcare data.
2. Bridges the Gap: It addresses the performance instability of standalone quantum models on NISQ hardware by leveraging the strengths of classical algorithms.
3. Ensures Interpretability: The integration of SHAP analysis provides necessary transparency for clinical decision-making, identifying specific psychosocial and behavioral risk factors.
4. Demonstrates Applicability: The study provides one of the few comprehensive evaluations of hybrid quantum-classical models in the context of maternal mental health, suggesting their practical viability for risk prediction in diverse, multicountry settings.

The authors remain modest regarding the current state of quantum hardware, acknowledging that the quantum component was simulated in a NISQ environment and that future work requires validation on larger, more diverse datasets and real quantum hardware.