# Learning Explainable Interventions to Mitigate HIV Transmission in Sex Workers Across Five States in India

**Authors:** Raghav Awasthi, Prachi Patel, Vineet Joshi, Shama Karkal, Tavpritesh Sethi  
**arXiv:** [2012.01930](https://arxiv.org/abs/2012.01930)  
**Date:** 2020-11-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2012.01930v1/

## Key points

- The study focuses on improving safe-sex practices in female sex workers (FSWs), who are vulnerable and stigmatized.
- Grassroot organizations face challenges in improving the effectiveness of interventions for FSWs due to a lack of quality access to care.
- The study combines structure learning, discriminative modeling, and grassroots-level expertise across five different Indian states to discover the influence of non-obvious factors for improving safe-sex practices in FSWs.
- A bootstrapped, ensemble-averaged Bayesian Network structure was used to quantify the factors that could maximize condom usage as revealed from the model.
- Financial literacy training was found to be the primary influence and predictor of condom use in FSWs through both generative and discriminative modeling approaches.
- The best model achieved 83% sensitivity, 99% specificity, and 99% area under the precision-recall curve for prediction.
- This work highlights how explainable models can transparently discover and prioritize anti-HIV interventions in FSWs in resource-limited settings.

## Summary

This study focuses on improving safe-sex practices in female sex workers (FSWs), who are one of the most vulnerable and stigmatized groups in society. Due to a lack of quality access to care, grassroot organizations engaged in improving health services often face challenges in improving the effectiveness of interventions for FSWs. To address this issue, the study combines structure learning, discriminative modeling, and grassroots-level expertise across five different Indian states to discover the influence of non-obvious factors for improving safe-sex practices in FSWs. The researchers used a bootstrapped, ensemble-averaged Bayesian Network structure to quantify the factors that could maximize condom usage as revealed from the model. The DAG G is defined as a triple, N = {X, G, P } over a set of random variables X. It encapsulates the underlying joint probability distribution P that can be factored as the product of probabilities of each node v conditioned upon its parents pa(v). The d-separation criterion allows directed edges E ⊑ V × V between vertices V to represent conditional independence relationships between random variables. A discriminative model was then constructed using XgBoost and random forest to predict condom use behavior. The best model achieved 83% sensitivity, 99% specificity, and 99% area under the precision-recall curve for prediction. Both generative and discriminative modeling approaches revealed that financial literacy training was the primary influence and predictor of condom use in FSWs. The study's insights have led to an ongoing field trial for assessing the real-world utility of this approach. This work highlights how explainable models can transparently discover and prioritize anti-HIV interventions in FSWs in resource-limited settings. Explainability and interpretability are key challenges in AI-based decision models, especially in public health settings. The expressive power of Bayesian networks combined with advances in learning structures directly make them well-suited to address complex problems that need transparency. The researchers learned the structure as a search problem over a constrained space of DAGs and evaluated it with the Akaike Information Criterion (AIC) as a metric. Since the DAG representation for a set of conditional independencies is not unique, an ensemble of such structures learned from bootstrapped samples of data was constructed using a majority voting criterion. For conducting inferences and queries, the DAG was parametrized with posterior probability distributions for all variables xvj ∈ X conditionalized upon evidence e = {e1, e2, ..em} from a set of variables X(e). The likelihood was then defined by P (e|xvj ) and inferred through application of Bayes rule. Finally, the posterior was always proportional to the estimated likelihood. In conclusion, this study provides valuable insights into improving safe-sex practices in FSWs through explainable models that transparently discover and prioritize anti-HIV interventions. By combining structure learning, discriminative modeling, and grassroots-level expertise across different Indian states, financial literacy training emerged as the primary influence and predictor of condom use in FSWs.