Cognitive behavioral therapy (CBT) is one of the most effective treatments for chronic low back pain. However, only half of Veterans have access to trained CBT therapists, and program expansion is costly. Moreover, VA CBT programs consist of 10 weekly hour-long sessions delivered using an approach that is out-of-sync with stepped-care models designed to ensure that scarce resources are used as effectively and efficiently as possible. Data from prior CBT trials have documented substantial variation in patients' needs for extended treatment, and the characteristics of effective programs vary significantly. Some patients improve after the first few sessions while others need more extensive contact. After initially establishing a behavioral plan, still other Veterans may be able to reach behavioral and symptom goals using a personalized combination of manuals, shorter follow-up contacts with a therapist, and automated telephone monitoring and self-care support calls. In partnership with the National Pain Management Program, we propose to apply state-of-the-art principles from "reinforcement learning" (a field of artificial intelligence or AI used successfully in robotics and on-line consumer targeting) to develop an evidence-based, personalized CBT pain management service (AI-CBT) that automatically adapts to each Veteran's unique and changing needs. AI-CBT will use feedback from patients about their progress in pain-related functioning measured daily via pedometer step-counts to automatically personalize the intensity and type of patient support, thereby ensuring that scarce therapist resources are used as efficiently as possible and potentially allowing programs with fixed budgets to serve many more Veterans.
The specific aims of the study are to: (1) demonstrate that AI-CBT has non-inferior pain-related outcomes compared to standard telephone CBT; (2) document that AI-CBT achieves these outcomes with more efficient use of scarce clinician resources as evidenced by less overall therapist time and no increase in the use of other VA health services; and (3) demonstrate the intervention's impact on proximal outcomes associated with treatment response, including program engagement, pain management skill acquisition, satisfaction with care, and patients' likelihood of dropout. We will use qualitative interviews with patients, clinicians, and VA operational partners to ensure that the service has features that maximize scalability, broad scale adoption, and impact.
278 patients with chronic low back pain will be recruited from the VA Connecticut Healthcare System and the VA Ann Arbor Healthcare System, and randomized to standard 10 sessions of telephone CBT versus AI-CBT. All patients will begin with weekly hour-long telephone counseling, but for patients in the AI-CBT group, those who demonstrate a significant treatment response will be stepped down through less resource-intensive alternatives to hour-long contacts, including: (a) 15-minute contacts with a therapist, and (b) CBT clinician feedback provided via interactive voice response calls (IVR). The AI engine will learn what works best in terms of patients' personally-tailored treatment plan based on daily feedback via IVR about patients' pedometer-measured step counts as well as their CBT skill practice and physical functioning. The AI algorithm we will use is designed to be as efficient as possible, so that the system can learn what works best for a given patient based on the collective experience of other similar patients as well as the individual's own history. Outcomes will be measured at three and six months post-recruitment, and will include pain-related interference, treatment satisfaction, and treatment dropout.
Ongoing analyses focus on evidence that the AI system implemented in the intervention arm is "learning" from patient feedback regarding how best to adapt its decision-making regarding the types of therapy session to recommend to maximize patient outcomes while minimizing costly therapist time: i.e., an IVR system message, a 15-minute live therapist booster call, or a 45-minute extended therapist counseling call. In making these decisions, the AI system is learning to what extent it should take into account patient feedback regarding various changes in their status (variables collectively called "state features). According to recent analyses, while making decisions for week 1, the AI system has learned to put the most weight on the patient's performance with their previous 15 minute live therapist calls, while spreading the remaining weight among the other patient-reported measures. For week 7, the AI system has learned that it should put the most weight on the Veteran's response to the IVR and 15 minute session types, but (in addition to the response to 45 minute sessions), that the system should apply added weight to the features related to change in pain intensity and CBT skill practice. Near the end of participants' exposure to the intervention, features corresponding to the patient's performance to each of the three session types carry the most weight, while the features representing pain related interference, pain intensity and skill practice have the next highest weighting. It is important to note that these results will likely change as the agent gains more data from making decisions about the patients enrolled in the program.
We hope to show that AI-CBT improves pain-related functional outcomes at least as much as VA's current evidence-based approach, and by scaling back unnecessary therapist contact, the AI-CBT approach will be significantly less resource-intensive. Secondary hypotheses are that AI-CBT will result in greater patient engagement and patient satisfaction.
- Piette JD, Krein SL, Striplin D, Marinec N, Kerns RD, Farris KB, Singh S, An L, Heapy AA. Patient-Centered Pain Care Using Artificial Intelligence and Mobile Health Tools: Protocol for a Randomized Study Funded by the US Department of Veterans Affairs Health Services Research and Development Program. JMIR research protocols. 2016 Apr 7; 5(2):e53.
- Piette JD. Computer algorithm for targeting support to VA patients with chronic back pain. 2016 Sep 30.