Novel digital approaches to the assessment of problematic opioid use

Figures

• 

  Fig. 1. Conceptual image illustrating potential sources of POU risk using puzzle pieces. The right puzzle pieces (Medical History Data and Lifetime & Current Psychiatric Disease Comorbidity) represent phenotypes that are usually mined from electronic health records (EHR), clinical notes, or structured assessments/questionnaires. The left puzzle pieces (Digital Footprint and Environmental & Societal Data) represent mostly environmental data that can be obtained from EHR, clinical notes, structured assessments/questionnaires, social media, and biometrics. The middle puzzle piece (Omics Data) represents large-scale genetic, epigenetic, proteomic, and metabolomic data. This piece links the two sides of the image as there are likely synergistic and/or causal relationships with the environment and the greater phenotype

• 

  Fig. 2. Bar plot (orange) of the count of psychiatric disorders and substance use disorders that were significant indicators of problematic opioid use (POU) phenotypes in our literature search (see Review Methods for criteria) and bar plot (blue) of the shared gene/locus count between psychiatric disorders and substance use disorders with POU (see reference [28] for methodology). Shared gene/locus associations reflect the relative representation of each disorder as significant predictors of POU. Depression, nicotine and alcohol use disorders, and anxiety disorders show high shared genetic liabilities with POU and are the most significant indicators of POU. However, schizophrenia displays high shared genetic liability with POU despite lower POU prediction

• 

  Fig. 3. Conceptual Image illustrating the flow from data to knowledge to prediction. In A, the puzzle pieces from Fig. 1 are not connected, illustrating that each part of the puzzle is being treated separately. Three individual data extraction methods are shown with blue arrows passing through a filter. The filter represents feature cleanup and engineering. Data build a knowledge base (gray cylinder) before various methodologies (gears) create models for prediction. The area under the receiver operating characteristic curve (auROC) plots and feature importance plots (horizontal bar graphs) represent levels of model accuracy. In B, the puzzle pieces from Fig. 1 are connected. Individual data streams are filtered together to create a single source of data that contributes to a larger knowledge base. Multiple methodologies working together (interlinked gears) result in better models (as measured by higher auROC and greater feature importance)