Poster Presentation 50th International Society for the Study of the Lumbar Spine Annual Meeting 2024

PHENOTYPING OPIOID CHRONICITY AND PREDICTING CHANGE IN OPIOID USAGE FROM PHYSICAL THERAPY ELECTRONIC HEALTH RECORDS (#47)

Naoki Takegami 1 , Kazuhito Morioka 1 , Abel Torres Espin 2 , Sachin Umrao 3 4 , Adam Ferguson 1 , Jeffrey Lotz 3 , Sang Pak 5 , Thomas Peterson 3 4 , Aaron Fields 3
  1. Department of Neurosurgery, Brain and Spinal Injury Center, Weill Institutes for Neurosciences, University of California, San Francisco, CA, United States
  2. School of Public Health Sciences, University of Waterloo, ON, Canada
  3. Department of Orthopaedic Surgery, University of California, San Francisco, CA, United States
  4. Bakar Computational Health Sciences Institute, University of California, San Francisco, CA, United States
  5. Department of Physical Therapy and Rehabilitation Science, University of California, San Francisco, CA, United States

Introduction

Chronic low back pain (cLBP) is a major global health issue, costing the US $90B/yr. The misuse of painkillers, particularly opioids, is rising among cLBP patients, which contributes to the opioid crisis. Currently, there's a significant lack of effective models to predict the chronicity of LBP. This situation results in biological, psychological, and social consequences including depression, increased medical expenses, and especially the misuse of analgesics, hindering early intervention and determination of appropriate treatment strategies. This study aims to create predictive models using data from the UCSF Physical Therapy (PT) department.

Methods

We analyzed 906 patients aged ≥18 from UCSF with diagnosis of cLBP within three months of beginning PT since 2016. Data were obtained from electric health records (EHRs) and the Chronic Opioid Registry. For the entire cohort, we analyzed opioid chronicity. After preprocessing, we used models (LightGBM, XGBoost, and Logistic Regression) with stratified K-fold cross-validation. For 394 individuals having opioid data, opioid prescription change was determined by comparing the pre-ISS and post-ISS visits. Regression models (LightGBM) were applied.

Results

Regarding demographics, the average age of the entire cohort was 50.91 years (SD: 17.25). The gender distribution was 40.77% male and 58.9% female. In terms of race, the breakdown was as follows: White or Caucasian (48.51%), Asian (21.33%), Black or African American (8.18%). For ethnicity, 12.38% identified as Hispanic or Latino. The average BMI was 25.48 (SD: 7.48). The CCI Score averaged at 0.42 (SD: 1.06), STarTBack Screening tool (SBT) 4.19 (SD: 3.22), PROMIS PHYSICAL at 42.87 (SD: 8.32), and PROMIS MENTAL at 47.96 (SD: 10.17).

In phenotyping opioid chronicity from the EHR, 260 patients (28.7%) were determined to have chronic opioid administration. Variables associated with opioid chronicity included the CCI Score (correlation coefficient 0.24), BMI (0.17), and depression (0.14). Regarding the prediction model for opioid chronicity, the LightGBM model had an AUC of 0.69, and the logistic regression model had an AUC of 0.67. Our ensemble model, composed of XGBoost and TabPFN, achieved an ROC of 0.75. When compared to the SBT, our ensemble model demonstrated a significantly stronger correlation with opioid chronicity, as indicated by a Cramér's V of 0.36, in contrast to SBT's 0.10.

For assessing changes in opioid prescription, our regression model surpassed the predictive utility of SBT for this purpose, with a Root Mean Square Error of 2.37. There were significant differences using ANOVA between the actual values of groups after grouped by the predicted values (p< 0.001, Figure 1A, B).

Discussion

We used a dataset of 906 individuals, with manually annotated and machine-extracted data, to accurately phenotype and create models to predict changes in drug dosage. Considering few studies aimed at predicting opioid misuse risk, our findings not only highlight the limitations of current clinical scores in predicting opioid misuse but also demonstrate the potential to outperform these scores. We aim to improve the accuracy of our models through feature engineering and to further refine and validate them using larger and more detailed datasets available.

65558371586b7-Figure_1a.jpg65558371586b7-Figure_1b.jpg