Artificial Intelligence (AI) to improve chronic pain care: Evidence of AI learning

Abstract

ObjectiveWe examined whether Cognitive Behavioral Therapy for chronic pain (CBT-CP) driven by artificial intelligence (i.e., AI-CBT-CP) increased its effectiveness through patient interactions. Materials and methodsData included 11,133 interactions with 168 patients receiving 10 weeks of AI-CBT-CP as part of a comparative effectiveness trial. Each week for each patient, AI-CBT-CP selected among three treatment recommendations: a 45-minute or 15-minute “live” therapy session, or asynchronous therapist feedback delivered via an interactive voice response (IVR) call. Recommendations were based on patients’ progress collected via IVR. AI-CBT-CP sought to optimize a “reward function” reflecting changes in patient-reported pedometer step counts and pain-related interference. We estimated changes over time in the frequency of each session type recommendation and the magnitude of reward scores. Simulations were used to predict increases in program effectiveness if AI-CBT-CP experienced more patient interactions than occurred during the trial. ResultsAI-CBT-CP had sufficient data to make recommendations 94% of the time. Recommendations evolved over time, with different patterns across patient subgroups and over patients’ course of therapy. As AI-CBT-CP gained experience via patient interactions, reward-scores increased significantly. In the model predicting reward based solely on calendar week, expected reward scores increased from 0.2949 at the beginning of the trial to 0.4637 in the 100th week of interaction (p = .002), with rewards doubling in one patient subgroup and increasing by more than 50% in another subgroup (p < .05 for both group-by-time interactions). Simulations indicated that if patient interactions continued, reward scores would continue to increase with up to 5000 interactions, at which point reward scores would stabilize. ConclusionsA program of CBT-CP driven by AI can learn from experience what treatment modalities work best to improve outcomes while conserving clinician time. As systems interact with more patients over longer periods of time, AI-driven disease management programs can become even more effective.