Introduction

Chronic low back pain (CLBP) is one of the leading causes of disability and chronic pain among adults. Despite evidence against effectiveness in CLBP, opioids are often prescribed. It has been shown that approximately 20% of individuals receiving long-term opioid therapy develop an opioid use disorder. Given the prevalence of CLBP in the United States and the debilitating effects of opioid dependency, it is important to identify individuals who will benefit from opioid therapy and those who may not. We used the novel network phenotyping strategy (NPS) to identify unique CLBP subpopulations with characteristic probabilities for continued opioid use at follow-up. We then determined how self reported pain scores at the initial visit were distributed among the CLBP subpopulations to gain further insight into which individuals may benefit from opioid therapy.

Methods

We used a UPMC database with 100 clinical variables from 20,000 CLBP patients prescribed opioids at the initial visit as input for NPS which is an explainable artificial intelligence approach. NPS gave us NPS-X and NPS-Y which are related to probabilities for continued opioid use and discontinued opioid use at 60-day follow-up respectively.

Self-reported pain scores ranged from 0-10 and were not included in the input list of 100 variables used to make NPS predictions. Pain-scores at initial visit for each patient were superimposed back onto each patient’s NPS-X and NPS-Y predictions for further analyses.

Results

NPS classified patients with CLBP into 27 unique subpopulations with characteristic probabilities for opioid use (Fig1). In the map, each patient is a point defined by their NPS-X and NPS-Y descriptors. Patients with higher value of X have higher probability of discontinuing opioids and those with higher Y have higher risk of staying on opioids.   

Clinical burden is the difference between probabilities for continued opioid use and stopped opioid use at follow-up. We determined that clinical burden and probability for a particular pain score are negatively correlated for pain scores between 0 and 5 at index visit. For pain score 6 to 10 at index visit, the clinical burden and probability for pain score were positively correlated (Fig 2).

For each subpopulation, we categorized pain scores at index visit into 2 groups - above 5 and below or equal to 5. We then compared 2 of the 27 subpopulations at a time using a fisher test to determine whether there was any difference in the distribution of pain scores above and below 5. We found that several subpopulation-subpopulation comparisons (48 out of 351 comparisons) were statistically significant (p<0.01) demonstrating there was a significant difference between distribution of pain scores above 5 vs. below 5 among subpopulations with characteristic probabilities for opioid use (Fig3). 

Discussion

Subpopulations with higher probabilities for continued opioid use have fewer pain scores in the range of 0-5 and more pain scores in the range of 6-10. Thus, patients with higher self-reported pain scores may need to be given more attention clinically and provided with alternatives to opioid therapy when possible.