Harris Hawks Optimization for depression prediction in FSWs

Ensemble Feature Selection and Harris Hawks Optimization predicts depression in 3,005 female sex workers with 95.78% accuracy, according to a paper submitted 23 June 2026 by Ahnaf Atef Choudhury and four co-authors. The study pairs an ensemble feature selection strategy (ANOVA and mutual information) with Harris Hawks optimization–tuned logistic regression and uses explainable AI to surface trauma-related predictors.

What did the authors build and test?

The paper develops a hybrid predictive pipeline that combines ensemble feature selection using ANOVA and mutual information with Harris Hawks optimization–tuned logistic regression, and evaluates it on data from 3,005 female sex workers. The model achieved an accuracy of 95.78%, an F1 score of 95.77%, and an AUC of 0.96. The authors report the model outperformed traditional classifiers on this dataset.

The dataset and evaluation are central to the claim: the sample size is explicitly 3,005 FSWs, and the reported metrics give concrete performance benchmarks. The work was submitted to arXiv on 23 June 2026 and accepted and presented at the 2026 8th IEEE Symposium on Computers & Informatics (ISCI 2026).

How does the hybrid model work?

The pipeline first applies ensemble feature selection that merges ANOVA and mutual information, then uses Harris Hawks optimization to tune a logistic regression classifier, and finally applies explainable AI methods to interpret predictions. Ensemble selection reduces dimensionality and highlights candidate predictors; Harris Hawks optimization searches for tuning parameters for logistic regression; XAI methods then identify which features drive risk scores.

The paper frames this as a new application of swarm intelligence to mental health prediction in a marginalized group. The listed contributors to depression in the model’s explanations are post-traumatic stress, client-related violence, and occupational factors. The authors position the approach as bridging conventional and machine learning methods to enable targeted psychosocial care and health planning.

Why does this matter?

High reported performance on a 3,005-person cohort suggests these methods can capture complex, high-dimensional risk patterns that conventional approaches miss, and the use of XAI makes the outputs interpretable for care planning. For female sex workers, the study highlights concrete trauma-related drivers—post-traumatic stress, client-related violence, occupational factors—so predictions are tied to actionable domains rather than opaque scores.

The combination of ensemble feature selection and a swarm-optimized logistic model is notable because it pairs statistical and information-theoretic feature ranking (ANOVA and mutual information) with a metaheuristic optimizer (Harris Hawks) to tune a standard, interpretable classifier. The paper explicitly claims improved performance over traditional classifiers on the study cohort.

What are the limits and caveats in the paper?

The paper presents results for one dataset of 3,005 female sex workers and frames the contribution as an application to a vulnerable group; it does not, in the abstract, present external validation results or deployment outcomes. The reported metrics—95.78% accuracy, 95.77% F1 score, AUC 0.96—are promising but specific to the study data and methodology as described in the submission.

What to watch

Look for the conference proceedings from the 2026 8th IEEE Symposium on Computers & Informatics (ISCI 2026) where the paper will appear, and for any follow-up that publishes code, external validation, or clinical-implementation tests. Confirmation would come from replication on independent cohorts and release of model code or data that allow external benchmarking.

References and provenance: the analysis above is grounded in the arXiv submission titled "Ensemble Feature Selection and Harris Hawks Optimization for Explainable Mental Health Risk Prediction in Female Sex Workers," submitted 23 June 2026 by Ahnaf Atef Choudhury, Md. Parvej Hoque Palash, Shahriar Siddique Ayon, Ramkrishna Saha, and Abdullah Al Mamun, and accepted for presentation at ISCI 2026. Specific reported metrics and identified factors are taken from the paper abstract.