A Joint Survival Modeling and Therapy Knowledge Graph Framework to Characterize Opioid Use Disorder Trajectories

TL;DR

The paper addresses the heterogeneity of opioid use disorder (OUD) trajectories by modeling three clinically meaningful time-to-event transitions—onset, remission, and relapse—using large-scale All of Us EHR and survey data. It introduces a two-pronged framework: (i) stage-specific regularized survival models to derive compact, interpretable risk-factor sets, and (ii) a therapy knowledge graph that maps these risk factors to evidence-based treatments and potential drug candidates via a PPR-based ranking. The study identifies consistent patterns across stages, notably the roles of chronic pain, mental health conditions, and polysubstance use, and demonstrates that recent clinical activity and MOUD exposure are linked to transition timing. By integrating predictive modeling with an evidence-weighted, mechanism-informed knowledge graph, the work supports interpretable decision support for personalized OUD prevention and treatment strategies.

Abstract

Motivation: Opioid use disorder (OUD) often arises after prescription opioid exposure and follows transitions among onset, remission, and relapse. Linked EHR-survey resources such as the All of Us Research Program enable stage-specific risk modeling and connection to intervention options. Results: We built a multi-stage framework to model time-to-onset, time-to-remission, and time-to-relapse after remission using All of Us EHR and survey data. For each participant we derived longitudinal predictors from clinical conditions and survey concepts, including recent (1/3/12-month) event counts, cumulative exposures, and time since last event. We fit regularized Cox models for each transition and aggregated selection frequencies and hazard ratios to identify a compact set of high-confidence predictors. Pain, mental health, and polysubstance use contributed across stages: chronic pain syndromes, tobacco/nicotine dependence, anxiety and depressive disorders, and cannabis dependence prominently predicted onset and relapse, whereas tobacco dependence during remission and other remission-coded conditions were strongly associated with transition to remission. To support therapeutic prioritization, we constructed a therapy knowledge graph integrating genetic targets, biological pathways, and published evidence to map identified risk factors to candidate treatments in recent OUD studies and clinical guidelines.

Figures (5)

• Figure 1: Predictors of OUD onset. Hazard ratios (HRs) from time-to-event models for OUD onset. HR$>1$ indicates increased hazard (earlier onset) and HR$<1$ indicates decreased hazard. Predictors shown are the 24 prioritized features selected by the modeling/feature-selection pipeline; Selected indicates recurrence across resampling/validation runs. Abbreviations: OUD, opioid use disorder; EHR, electronic health record; HR, hazard ratio.
• Figure 2: Top predictors of OUD remission. Hazard ratios (HRs) summarize associations with time to OUD remission; HR$>1$ indicates a higher hazard of remission (i.e., faster remission), and HR$<1$ indicates a lower hazard. To improve readability, the x-axis is shown on a log scale; values above 10 are plotted at the axis limit and annotated as off-scale. Predictors shown are the 10 prioritized features in the remission model, including recent-condition counts (last 30 days; last 1 year) and longitudinal summaries (cumulative counts; days since last occurrence). Abbreviations: OUD, opioid use disorder; HR, hazard ratio.
• Figure 3: Top predictors of OUD relapse (hazard-ratio point estimates). Hazard ratios (HRs) summarize point-estimate associations with time to relapse after remission; HR$>1$ indicates a higher hazard of relapse. Predictors shown are the top-ranked features from the relapse screening/prioritization pipeline. The x-axis is log-scaled to accommodate extreme HR values driven by sparse counts.
• Figure 4: Relapse model significant predictors (FDR $<0.05$). Forest plot of adjusted hazard ratios (HR) and 95% confidence intervals (CIs) from the relapse transition model. Points denote HR estimates and horizontal bars indicate 95% CIs; the dashed vertical line marks $\mathrm{HR}=1$. Predictors are modeled per unit increase in $\log(1+x)$-transformed counts for utilization/count features (e.g., MAT, prescriptions, diagnoses, opioid-related counts) and per unit increase in “days since last” variables. Only predictors passing FDR $<0.05$ are shown. The case indicator is excluded because it encodes cohort labeling rather than a clinical predictor.
• Figure 5: Knowledge graph drug repurposing results for OUD. The graph connects OUD (red) to GWAS/meta-analysis prioritized genes (orange; including direct seeds and pathway-bridged genes) and to drugs via curated gene--drug target relationships. Green nodes indicate established medication-assisted treatments (MAT/MOUD) for OUD; blue nodes indicate non-MAT drugs prioritized by Personalized PageRank starting from the OUD node. Recovery of known MAT/MOUD drugs provides face validity. Non-MAT candidates cluster around mechanistic axes such as opioid receptor signaling (OPRM1-linked), PDE4/cAMP signaling (PDE4B-linked), and ion-channel modulation (KCNN1-linked), suggesting hypotheses for adjunctive or alternative therapeutic exploration. Duplicate drug entries reflect ingredient versus salt forms in source databases.