Machine learning based phenotyping of the response to mindfulness for chronic low back pain

Akshit Soota¹Angela C Incollingo Rodriguez¹Benjamin C Nephew^1*Paula Gardiner²Jean King¹Natalia Morone³Carolina Ruiz¹

• ¹Worcester Polytechnic Institute, Worcester, MA, United States
• ²Cambridge Health Alliance, Cambridge, United States
• ³Boston Medical Center, Boston, United States

Background Millions of people each year suffer from chronic low back pain (cLBP), which adversely affects their physical and mental health. While non-pharmacological interventions such as mindfulness are known to be effective in treating cLBP, not all patients experience the same benefit. Determining who these treatments might work best for is difficult, as there are no reliable predictors of the response to mindfulness for cLBP. The objective of the current study was to apply predictive machine learning to data collected from a completed clinical trial of mindfulness for cLBP to identify phenotypes characterizing those who did and did not respond to the intervention. Methods The analyses here focused on 132 participants in the intervention arm of the clinical trial of mindfulness for cLBP. The Random Forest machine learning technique was used to identify key characteristics of responders (49) and non-responders (83). Results The top three responder phenotypes were able to identify 26 out of the 49 responders with 92-100% precision. The top three non-responder phenotypes were able to identify 36 out of 83 non-responders, all with 100% precision. Conclusions Results from this machine learning based phenotyping can guide clinician and patient decision-making to maximize clinical efficiency, patient outcomes, and resource use as well as inform research and development of mindfulness-based treatments for pain.