Ensemble Feature Selection and Harris Hawks Optimization for Explainable Mental Health Risk Prediction in Female Sex Workers

Technical Summary: Ensemble Feature Selection and Harris Hawks Optimization for Explainable Mental Health Risk Prediction in Female Sex Workers

Problem Statement
Female sex workers (FSWs) face a disproportionate burden of mental health disorders, particularly depression and post-traumatic stress disorder (PTSD), driven by structural violence, stigma, economic instability, and occupational risks. While machine learning (ML) offers potential for modeling these complex outcomes, existing approaches often struggle with high-dimensional, nonlinear, and imbalanced datasets. Traditional statistical methods and simplistic classifiers frequently fail to capture intricate interactions between violence exposure, socioeconomic variables, and health indicators. Furthermore, a lack of integrated feature selection, sophisticated optimization, and Explainable AI (XAI) limits the accuracy, stability, and clinical interpretability of current predictive models, hindering early intervention and targeted psychosocial care.

Methodology
The study proposes a hybrid predictive framework designed to enhance both performance and interpretability for depression prediction among FSWs. The methodology follows a structured pipeline:

1. Data Collection and Preprocessing: The model utilizes a community-based cross-sectional dataset comprising 3,005 adult FSWs from nine provinces in South Africa. The raw data included 20 columns; after handling missing values (mean imputation for numerical data, removal for categorical) and encoding, the final dataset consisted of 2,911 observations and 19 features. The target variable was binary (depression: yes/no).
2. Hybrid Ensemble Feature Selection: To identify the most significant predictors, the authors employed an ensemble strategy combining Analysis of Variance (ANOVA) and Mutual Information (MI). This approach ranked features by importance, filtering out less significant variables to create an optimized set of 11 features. Key drivers identified included PTSD outcomes, location, earning potential per client, and various forms of abuse (client, intimate partner, police).
3. Model Development and Optimization: Several classifiers were evaluated, including Random Forest (RFC), k-Nearest Neighbors (kNN), Support Vector Classifier (SVC), Light Gradient Boosting Machine (LGBM), Artificial Neural Networks (ANN), and Logistic Regression (LR). Among these, Logistic Regression was selected as the base model. Its hyperparameters were optimized using Harris Hawks Optimization (HHO), a swarm intelligence algorithm. The HHO configuration included a population size of 30, 50 maximum iterations, and an escape energy parameter ($E_0$) of 2, designed to efficiently search the parameter space and handle class imbalance.
4. Explainable AI (XAI): To ensure clinical applicability, the study integrated LIME (Local Interpretable Model-agnostic Explanations). LIME was used to generate instance-level explanations, clarifying how specific features (e.g., trauma history, occupational factors) contributed to individual predictions, thereby bridging the gap between "black box" models and clinical trust.

Key Contributions

• Novel Hybrid Architecture: The paper introduces a unified framework combining ensemble feature selection (ANOVA + MI) with HHO-tuned Logistic Regression, specifically applied to mental health prediction in marginalized populations.
• Swarm Intelligence Application: This work represents the first application of Harris Hawks Optimization to tune a Logistic Regression model for mental health prediction in FSWs, demonstrating superior performance over baseline models.
• Enhanced Interpretability: By integrating LIME, the study provides transparent, instance-level insights into model decisions, identifying specific traumatic and occupational variables that drive depression risk, thus facilitating evidence-based interventions.
• Robust Feature Identification: The ensemble selection method successfully isolated high-impact features, reducing dimensionality while retaining critical risk indicators such as client-related violence and PTSD history.

Results
The proposed HHO-optimized Logistic Regression model, utilizing the ensemble-selected features, achieved state-of-the-art performance on the test set:

• Accuracy: 95.78%
• F1-Score: 95.77%
• AUC: 0.96
• Precision: 95.60%
• Recall: 95.95%

These metrics outperformed all baseline models (including standard LR, ANN, and LGBM) and other optimization strategies (such as Grid Search). Specifically, the HHO-based approach improved recall by 1.20% and the F1-score by 1.22% over grid-searched LR, translating to approximately 36 additional correct depression detections in the cohort.

LIME analysis confirmed the stability of the model's logic, showing that PTSD outcomes and client abuse were the dominant predictors in 94.3% and 87.1% of cases, respectively. The feature importance rankings from LIME strongly correlated with the global ANOVA+MI rankings (Spearman $\rho = 0.91$), validating the model's consistency.

Significance and Claims
The authors claim that this work bridges the gap between conventional statistical approaches and advanced machine learning to create a tool capable of early detection and targeted support for vulnerable groups. The model's high accuracy and interpretability make it suitable for integration into existing health systems (e.g., REDCap-based systems) to assist Community Health Workers (CHWs) in conducting risk screenings and referrals.

The study emphasizes that while the model demonstrates high internal validity through cross-validation, its external generalizability to other geographic regions remains to be tested. The authors position the framework as a lightweight, real-time capable solution that can operate without cloud support, making it viable for low-resource settings. Future work is proposed to include mixed-methods validation, fairness-aware evaluation, and expansion to cohorts in East Africa and South Asia to address geographic biases.