GuardRail: A Computational Approach to Early Prediction of Opioid Risk Behavior Using a Multimodal Machine Learning & LLM Framework
Year: 2026
Type: preprint
Abstract: Objective: The opioid crisis leads to approximately 100,000 deaths annually in the United States. Early prediction of clinical outcomes is essential for patient-risk evaluation and treatment allocation, yet current research lacks validated predictive tools to assess opioid-related aberrant behavior development early. To address this gap, GuardRail, a cost-effective and non-invasive computational framework, was developed to predict aberrant behavior development following initial opioid exposure, enabling earlier clinical intervention. Methods: Multiple classification algorithms (generated with scikit-learn) were trained on a cohort of 15,175 patients from the MIMIC-IV dataset. Data was divided into 70:30 training: test split. Within the training set, 5-fold stratified cross-validation was used for hyperparameter tuning. Models incorporated both structured data (pharmaceutical records) and unstructured data (clinical narratives) through LLM-based clinical record analysis. Results: All models achieved Area Under the Receiver Operating Characteristic Curve (AUROC) scores > 0.92, with gradient boosting demonstrating superior discriminative performance: AUROC of 0.9575 (test), recall of 93.3%, and precision of 84%. A permutation-based feature importance analysis identified current opioid prescription, Diagnosis Related Group (DRG) severity, and mental health history as major predictors of aberrant behavior formation. An ablation study demonstrated that clinical narrative features derived from LLMs contribute marginal predictive value beyond a model with solely structured features (ΔAUROC = -0.0003), suggesting that structured EHR data alone is sufficient to attain high performance in the present cohort. Conclusion: This performance was achieved using only structured EHR data, which makes GuardRail a cost-effective, noninvasive method for screening, compared to more resource-intensive intervention techniques, though this has not been formally evaluated from an economic perspective. The high sensitivity of GuardRail points to the potential for earlier identification of patients with recurrent opioid-related aberrant behavior. Further validation, including external and prospective assessments, is needed before decisions can be made on clinical implementation or effects on opioid-related negative outcomes.
Author: Sid Sreekanth
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7224885889