Opioid overdose is now one of the most concerning public health crises of the 21st century, with a sharp increase in misuse of opioids, unintentional overdoses and opioid use disorder (OUD) over the last two decades. Prescription opioids have become the predominant substance of addiction, and their consumption can result in both physical and psychological adverse effects. Opioid overdose is the leading cause of accidental mortality among young adults. Opioid-related aberrant behaviors (ORABs) are patient behaviors indicative of prescription medication misuse (Fleming et al., 2008). In the dataset used for this study, ORABs are defined as confirmed aberrant behaviors, meaning behaviors with clear evidence of opioid misuse, or addiction (Portenoy, 1996, National Institute on Drug Abuse, 2023). Prior literature has shown that ORABs are strongly associated with patients exhibiting drug misuse problems (Webster and Webster, 2005). Accordingly, the identification of ORABs has been recognized as an important component in assessing the risk of opioid misuse, and opioid overdose (OOD) (Maumus et al., 2020, Lingeman et al., 2018). In 2018, forecasts indicated that the United States would have about 500,000 opioid-related fatalities in the subsequent decade. The consequences of opioid usage are numerous. In the surgical context, opioid usage correlates with a heightened risk of postoperative complications, mortality, and falls and fractures among the aged (Brott et al., 2022). is now essential to advance toward the implementation of practical strategies for identifying and managing at-risk patients. This represents a crucial component of the broader challenge, as opioid-related aberrant behaviors (ORABs) serve as key indicators of prescription medication misuse and early warning signs for patients at high risk of overdose and opioid use disorder (OUD) (Kwon et al., 2024, Kwon et al., 2023, Goldberger et al., 2000). Aberrant behaviours, including falsification of prescriptions, diversion of medications, and non-adherence to prescriptions, are confirmed abnormalities that are clear evidence of loss of control of opioid use, increasing risks to patient safety (St Marie, 2019). Therefore, early detection of these behaviours is required in order to manage safely on opioid therapy, to prevent adverse outcomes, and to individualize interventions (Brott et al., 2022). However, confirming ORABs is a problematic, underexplored task due to the multifaceted nature of the opioid misuse problem, as well as to current risk assessment tool limitations.

Standard tools used to assess the risk of opioid misuse, including the Opioid Risk Tool (ORT) and Screener and Opioid Assessment for Patients with Pain-Revised (SOAPP-R), are based primarily on self-reported or clinician estimates of patient response (Brott et al., 2022, Claxton and Arnold,). These tools have a varying degree of sensitivity [53-99% for ORT, 79-91% for SOAPP-R] and specificity (16-96% for ORT, 52-69% for SOAPP-R) and exhibit significant shortcomings (Maumus et al., 2020, Claxton and Arnold,, Butler et al., 2009). However, these approaches are also limited by their inability to capture information rich in clinical narratives regarding extensive behavioural patterns, the propensity to bias in patient reporting, and reliance on structured data, ignoring the available textual information in unstructured electronic health records (EHR) notes. These notes may hold subtle clues to aberrant behaviour that can prove vital for accurate risk prediction, for instance, inconsistent statements and attempts to lie to clinicians. Currently, state-of-the-art (SOTA) approaches to opioid misuse detection are also using natural language processing (NLP) and machine learning techniques to process clinical data. However, these enhancements have mainly dealt with overall opioid misuse instead of the genuine expectation of affirmed ORABs, leaving a gap in the examination (Kwon et al., 2023).

Furthermore, the existing predictive models are not interpretable, providing another constraint to their adoption in the clinic. Although black box models can be accurate while remaining uncertain, they offer no actionable explanation for how predictions are derived, which raises ethical concerns like biased decision-making or person with substance-use disorder (Claxton and Arnold,, Butler et al., 2009). However, for the ORAB prediction, techniques such as SHAP (Shapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) have been applied to enhance the transparency of the model, although their application for the ORAB prediction is limited (Claxton and Arnold). Eventually, combining clinical structured variables with unstructured text may provide enriched ORAB risk profile information, although this has not occured for confirmed ORABs.

However, these gaps are addressed by this study, which tackles the problem of the prediction of confirmed aberrant behaviours on opioid therapy in patients with opioids. Instead, we propose a novel approach that uses the multimodal data and explainable machine learning to address the limitations of traditional tools. Using the Opioid Related Aberrant Behavior Detection Dataset (ODD) (Kwon et al., 2023), a recently developed dataset containing both structured clinical information and unstructured EHR text, we aim to enhance the detection of subtle behavioral indicators associated with confirmed ORABs. In our framework, we use classifiers such as the Opioid Risk Ensemble, ensuring the classifier is interpretable with SHAP analysis and performing an ablation study to determine the contributions of features. This study aims to take a step towards this end by pioneering a multimodal and explainable approach for the early identification of ORAB to provide clinicians with a practical and transparent tool for safer opioid prescribing practices and eventually reduce the disastrous effects of the opioid epidemic. Our approach uses multimodal, clinically standardized information alongside free-text EHR documentation to identify concealed indicators that are missing from numeric inputs alone. Precise monitoring of aberrant behaviours depends heavily on the improved data representation, which standard prescription systems cannot detect correctly. The manuscript discusses the related work in Section 2 and the proposed methodology in Section 3. Section 4 presents the results, followed by the discussion in Section 5. Section 6 presents the limitations of the study. Finally, Section 7 provides the conclusion and outlines directions for future work.