GRACE: GRaph-based Addiction Care prEdiction

TL;DR

GRACE introduces a graph-based framework for locus-of-care prediction in addiction treatment, addressing severe class imbalance with a meta-learning anchored GNN and an unbiased meta-graph sampler. It combines rich note-based node features, a patient similarity network, and reasoning pathways extracted from LLMs to improve predictive power across binary, risk-stratified, and multilabel tasks. The approach achieves state-of-the-art minority-class performance, demonstrates strong results on TA, TB, and TC tasks, and provides insights into feature contributions and meta-graph design, highlighting practical potential for AI-assisted clinical decision support in resource-constrained addiction care.

Abstract

Determining the appropriate locus of care for addiction patients is one of the most critical clinical decisions that affects patient treatment outcomes and effective use of resources. With a lack of sufficient specialized treatment resources, such as inpatient beds or staff, there is an unmet need to develop an automated framework for the same. Current decision-making approaches suffer from severe class imbalances in addiction datasets. To address this limitation, we propose a novel graph neural network (GRACE) framework that formalizes locus of care prediction as a structured learning problem. Further, we perform extensive feature engineering and propose a new approach of obtaining an unbiased meta-graph to train a GNN to overcome the class imbalance problem. Experimental results in real-world data show an improvement of 11-35% in terms of the F1 score of the minority class over competitive baselines. The codes and note embeddings are available at https://anonymous.4open.science/r/GRACE-F8E1/.

Figures (6)

• Figure 1: An example illustrating Grace framework as a black box, given an input clinical note with the reasoning pathways highlighted.
• Figure 2: Workflow of Grace framework. Recall that $\mathcal{N}_p$ represents all the visit notes of patient $p$ concatenated together. $\mathcal{F}_B$, $\mathcal{F}_L$, $\mathcal{F}_E$, and $\mathcal{F}_R$ represents the components of node features as base, lexical, emotive, and reason embeddings, respectively.
• Figure 3: Word clouds of frequently occurring words in IP and OP classes.
• Figure 4: Side-by-side comparison of $t$-SNE plots of node embeddings before and after training with Grace. The rows represents embeddings used as $\mathcal{F}_B$, $\mathcal{F}_B$+$\mathcal{F}_L$, $\mathcal{F}_B$+$\mathcal{F}_E$, $\mathcal{F}_B$+$\mathcal{F}_E$+$\mathcal{F}_L$ and $\mathcal{F}_B$+$\mathcal{F}_E$+$\mathcal{F}_L$+$\mathcal{F}_R$ respectively. Here blue and orange dots indicate IP and OP classes respectively.
• Figure 5: Prompts used to get reasoning and classification.