Grace : GR aph-based A ddiction C are pr E diction

Subham Kumar Prakrithi Shivaprakash Koustav Rudra, Lekhansh Shukla, Animesh Mukherjee · 2025

Read Grace : GR aph-based A ddiction C are pr E diction by Subham Kumar Prakrithi Shivaprakash Koustav Rudra, Lekhansh Shukla, Animesh Mukherjee (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-20671.txt.

Introduction

Subham Kumar, Lekhansh Shukla, Animesh Mukherjee, Koustav Rudra, and Prakrithi Shivaprakash address a hard clinic choice, deciding the right level of care for people seeking help with addiction. Beds and staff for live in care are scarce, so a wrong call wastes resources or leaves a patient with too little support.

Their system, GRACE, frames the choice as a graph learning task and builds a fair training graph to fight the severe imbalance in past records. Rare care levels are swamped by common ones. On real world data, scores for the rare class rise by wide margins over rival methods. Tuning the input features together with the graph model adds a further clear boost in accuracy.

Catalog record

CREATORSubham Kumar Prakrithi Shivaprakash Koustav Rudra, Lekhansh Shukla, Animesh Mukherjee
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-20671

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Grace : GR aph-based A ddiction C are pr E diction Subham Kumar Prakrithi Shivaprakash Koustav Rudra, Lekhansh Shukla, Animesh Mukherjee 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/ . 1 Introduction The healthcare industry is experiencing a transition to AI-driven approach Sutton et al. 2020 in solving the major issues in patient care management and resource allocation. To contextualize the objectives of this work, it is important to understand the domain of addiction psychiatry and the unsolved challenges. One such challenge is in determining the appropriate decision for patient care based on clinical condition. Substance use disorders (SUD), especially alcohol use disorders (AUD), cause brain damage and premature death Volkow and Blanco 2023 . For a person with AUD to stop drinking is often fraught with danger, characterized by psychosis and complicated withdrawal symptoms (seizures, delirium, etc.). Half of those who suddenly stop or reduce their drinking tend to experience alcohol withdrawal syndrome, although the severity varies Goodson et al. 2014b . The significance of this problem was acutely realized during COVID-19, when a large number of patients developed complicated withdrawal syndrome in India Narasimha et al. 2020 . In addition, Gururaj et al. 2017 showed 90% unmet need for treatment of SUDs in low-resource countries. The process of safely stopping alcohol use requires medical treatment, which is commonly called as detoxification. Healthcare providers must classify patients seeking alcohol detoxification into inpatient (IP) and outpatient (OP) triage. In medical systems, this decision point is referred to as the ‘locus of care’. Computationally, this task of binary classification is challenging due to the highly imbalanced nature of the addiction dataset. In the real world, inpatient cases are significantly lower than outpatients, primarily due to two reasons. First, most AUD patients actually do not develop severe illness. Second, there are resource constraints in terms of the number of experts to effectively manage patients or the availability of hospital infrastructure to monitor admitted patients. An additional layer of complication stems from another type of uncertainty in hospital admission – patients who actually need admission ( high-risk patients) may not get it either due to lack of beds, or because they do not consent while highly motivated but anxious patients who can get treated safely at home ( low-risk patients) may insist on admission. Note that high-risk patients can be further categorized into subclasses. There have been attempts to model risk stratification using laboratory investigations such as platelet count, blood alcohol levels, and liver function tests Goodson et al. 2014a . Laboratory tests are often unavailable in low-resource centres, and when available, have a turn-around time of 3-4 hours at least, precluding efficient and quick decision-making in a busy outpatient setup. However, information collected during a clinical consultation itself (mostly recorded digitally as a clinical note by medical experts) can provide signals to assess the risk of complicated withdrawal. This work is motivated by the growing need to develop an accurate and practically applicable clinical decision support systems in addiction treatment to predict the locus of care. To address these gaps Lamb S 1998 , in this paper, we propose a GNN-based unified framework – Grace . The novelty lies in (a) obtaining initial node features (section 4.1 ) representing a patient, (b) expressing the ‘semantic’ similarity between patients based on these features, (c) feeding this network to a meta-learning anchored GNN model to make accurate predictions (section 4.2 ). A second line of novelty is the inclusion of reasoning pathways hidden in the clinical notes into the patient representation, enhancing the predictive power of Grace . Research questions. RQ1. Can we use a network formulation and a GNN based architecture for the prediction of the locus of care? To answer this question, we introduce the patient similarity network (PSN), where patients are the nodes and edges connect ‘semantically’ similar patients. The features representing a patient node are obtained from the clinical notes recorded about the patient over his/her trajectory of hospital visits. This PSN is then treated as an input to train a GNN model for locus of care prediction. RQ2. How do we resolve the class imbalance problem? To address the inherent class imbalance prevalent in the addiction care dataset, we integrate meta-learning into the GNN framework. The meta-GNN Mohammadizadeh et al. 2023 framework dynamically adjusts the training sample weights, using an unbiased meta-data set to minimise the bias toward the majority class. We further propose novel heuristics and use genetic algorithm to guide the selection of meta-nodes such that meta-data set maintains similar properties as that of the base training graph (section 4.2 ). RQ3. Does integrating reasoning pathways as additional node features be effective? When a clinical note is drafted about a patient, the medical expert typically follows a reasoning pathway to arrive at the present locus of care decision. Some of this reasoning might be explicit in the notes, while others might be implicit. We make use of the SOTA reasoning-based large language models (LLMs) to extract these reasoning pathways ‘hidden’ in the clinical notes (see Figure 1 for an example). We fuse this silver data as an additional node feature with the objective of improving the predictive power of our model. Figure 1: An example illustrating Grace framework as a black box, given an input clinical note with the reasoning pathways highlighted. Contributions. This paper makes the following contributions. 1. To our knowledge, Grace is the first model to predict locus of care decisions for addiction patients, supporting binary, risk-stratified, and fine-grained multilabel classification. 2. We conduct extensive evaluations on Grace using different GNN architectures, including GCN Kipf and Welling 2017 ( Grace GCN ), GraphSAGE Hamilton et al. 2018 ( Grace Gsage ), GAT Veličković et al. 2018 ( Grace GAT , and GraphTransformer Shi et al. 2021 ( Grace Gtran ). Grace Gsage demonstrates the best performance, in terms of F-1 score (0.74 for the minority class), while Grace Gtran (0.73) is the second best (see Table 2 ). Grace also beats the baseline SOTA LLMs by 11-35%. 3. We perform ablation of node features and heuristics involved in the construction of meta-graph (section 5 and Appendix E ). Sequentially enriching node features led to consistent improvements, while omitting any one of the heuristics in genetic algorithm reduced the performance by 7-9%. 4. Finally, we show that the predictive power of Grace goes

Read the complete text