Optimal and Non-Discriminative Rehabilitation Program Design for Opioid Addiction Among Homeless Youth

Amulya Yadav, Roopali Singh, Nikolas Siapoutis, Anamika Barman-Adhikari, Yu LiangIJCAI 2020 · DOI

TL;DR

This paper introduces CORTA, an AI agent that designs personalized rehabilitation programs for homeless youth struggling with opioid addiction. By leveraging causal inference models and Integer Linear Programming (ILP), CORTA aims to minimize the number of youth suffering from addiction while optimizing resource allocation and ensuring fairness.

What Problem Does This Paper Address?

Opioid addiction is a critical public health issue, particularly prevalent among homeless youth. Current rehabilitation efforts often rely on ad-hoc assessments and unprincipled heuristics to tailor programs, leading to suboptimal effectiveness. A key challenge is understanding the complex, underlying causes of addiction for each individual and matching them with the most effective rehabilitation strategies. Furthermore, rehabilitation centers operate under significant manpower and resource constraints, necessitating efficient assignment of limited resources. Finally, ensuring that these AI-driven systems do not discriminate against vulnerable populations based on sensitive characteristics is paramount.

Methods

CORTA employs a three-pronged approach:

  1. Causal Inference Models: Instead of traditional machine learning, CORTA uses causal inference models (specifically Bayesian Additive Regression Trees - BART) to predict the likelihood of opioid addiction. This allows for uncovering hidden causal relationships between various factors and addiction, crucial for personalized treatment. These models also generate "counterfactual estimates," which predict how a youth's addiction risk would change under different rehabilitation scenarios.
  2. Integer Linear Programming (ILP) Formulation: CORTA uses these counterfactual estimates to formulate an ILP. This optimization problem aims to find the optimal assignment of homeless youth to rehabilitation programs. The objective is to minimize the expected number of youth who remain addicted, while respecting program capacity constraints and ensuring that youth are assigned to programs that are predicted to be effective for them.
  3. Fairness Constraints: The ILP formulation can incorporate fairness constraints. This ensures that the assignment process does not disproportionately exclude individuals from protected minority categories, aiming for equitable distribution of rehabilitation resources.

Datasets

The paper introduces a novel, first-of-its-kind real-world dataset collected from approximately 1400 homeless youth. This dataset includes information on their educational background, exposure to violence, mental health history, STI status, and opioid misuse history. For training and evaluation, the dataset was pre-processed, resulting in 1426 data points, each with 212 features and a binary label indicating opioid addiction. The dataset exhibits class imbalance, with 80% of youth not suffering from opioid addiction.

Key Insights and Ideas

  • Causality Matters: The paper strongly advocates for causal inference over traditional ML for this problem. Understanding why someone is addicted (causal factors) is more important than just predicting if they will be addicted, especially for designing effective interventions.
  • Counterfactuals Drive Optimization: The ability to generate counterfactual predictions (what would happen if...) is key to optimizing program assignment. This allows for a more informed decision-making process than simply assigning based on current risk scores.
  • ILP for Optimal Resource Allocation: Integer Linear Programming provides a rigorous mathematical framework to balance multiple objectives: maximizing rehabilitation impact, minimizing wasted resources, and respecting capacity limits.
  • Fairness is Achievable: The research demonstrates that incorporating fairness constraints into the ILP does not significantly degrade the effectiveness of the rehabilitation assignment, suggesting that equitable outcomes can be achieved without sacrificing performance.

Key Results

  • CORTA significantly outperforms baseline methods, rehabilitating 110% more opioid addicts on average.
  • The system achieves minimal resource wastage, with ~4% unused capacity, compared to baselines that can reach up to ~75% wastage.
  • The importance of both causal modeling and ILP optimization is highlighted, with CORTA outperforming methods that use causal models but lack explicit optimization, and vice-versa.
  • Incorporating fairness constraints showed minimal impact on the number of addicts rehabilitated and resource wastage, suggesting that fairness can be achieved without a significant performance trade-off.

Limitations and Future Work

  • Ground Truth Treatment Effects: The paper acknowledges that actual ground-truth heterogenous treatment effects of rehabilitation programs are not available in their dataset. They rely on their BART model's counterfactual probabilities as a proxy for this.
  • Generalizability: While the dataset is novel, its specific characteristics might influence the generalizability of the findings to other populations or contexts.
  • Future Work: The authors suggest exploring more sophisticated fairness metrics and potentially incorporating dynamic aspects into the rehabilitation program design, as needs can change over time. Further research could also investigate the impact of different types of rehabilitation programs and their specific causal links to addiction.

Summary updated

BibTeX
@inproceedings{yadav2020optimal,
  author = {Amulya Yadav and Roopali Singh and Nikolas Siapoutis and Anamika Barman-Adhikari and Yu Liang},
  title = {{Optimal and Non-Discriminative Rehabilitation Program Design for Opioid Addiction Among Homeless Youth}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/605},
}
Figure