New research reports how AI analyzes patient clinical trial experiences and quantifies a placebo response in patients with chronic pain.
Placebos work, but we still don’t really know why. Knowing who would respond well to a placebo — typically just a sugar pill — could help doctors prescribe certain patients a placebo instead of a real drug to alleviate chronic pain, potentially helping them to save money and avoid side effects.
AI could help. Machine learning and natural language processing (NLP) analysis of patients’ verbal accounts of their experience are now starting to shed some light on the placebo response.
In a newly published paper,1 we report the first proof-of-concept that uses AI to analyze patients’ clinical trial experiences. The AI quantifies a placebo response in patients with chronic pain and distinguishes those who respond to placebo from those who do not.
The results show key differences in language use between patients who responded favorably to placebo — meaning their pain improved —versus those that did not, with patients’ language picking up on underlying personality traits and psychological factors. We hope that in future, clinicians could use speech-to-text AI to transcribe their conversations with patients and assess how likely a patient would be to respond to a placebo instead of a drug. Knowing ahead of time who would be good placebo responders, doctors may not have to prescribe them treatments at all, and still help them get better.
The results could also help improve the design of future clinical trials by either removing potential placebo responders from active treatment groups or by helping better balance placebo responders across different treatment arms to make statistics more robust.
Determining whether a patient’s placebo response is reliably predictable is tricky. It’s also difficult to pinpoint what features of a person’s pain experience, personality or cognitive processes factor most into the response. The placebo response has been observed across a variety of conditions and treatment types – including pills, patches, injections and surgeries — with some of the most significant placebo responses producing meaningful relief for those experiencing chronic pain.
Working with researchers from Northwestern University and McGill University in Montreal, we conducted a double-blinded study, meaning both the study participants and the facilitators were unaware of who had received a placebo pill during language collection and initial data analysis. We distinguished — with 79 percent accuracy using language features, alone — chronic back pain sufferers who were unknowingly helped by a placebo from those whose pain was not reduced.
To better understand why this response happens and how we could one day hopefully predict it, we used chronic pain patients’ speech collected via open-ended interviews to quantify hidden aspects of their thoughts and emotions. Based solely on what they said, we used AI to identify them as placebo responders (those who had a positive effect to a placebo pill, in the form of pain reduction) or non-responders (those who either showed no effect to the placebo pill or a negative effect — “a nocebo response”).
This is not the first time researchers have looked into the relationship between pain and voice. But most previous studies have concentrated on acute instead of chronic pain or have focused on the acoustics and quality of a subject’s voice (how they speak) instead of their language content (what they say).
Given that language is “a window into the mind,” we wanted to do it differently. We suspected that using a quantitative language methodology would provide a tool to directly tap into certain personality traits and psychological factors to help us objectively measure patients’ pain experience and response to treatment.
We conducted a monitored, registered clinical trial approved by an Institutional Review Board designated to review and monitor biomedical research under FDA regulations. We randomized patients with chronic low back pain into one of three groups: a placebo treatment group (sugar pills), an active treatment group (over-the-counter naproxen pain relief pills) or a no-treatment group (no pills).
We then followed these patients for eight weeks using a combination of clinical questionnaires, neuroimaging, and pain ratings collected twice a day on a smartphone. All participants signed an informed consent form, and data were de-identified prior to analysis.
At the end of the trial, we conducted an exit interview asking patients about various topics including their hobbies, medical history, and pain experiences. From this interview, we extracted more than 300 language features from what patients said. These features included how many words a person used in their interview, their use of different parts of speech (nouns, verbs, adjectives), how positive or negative their speech was, and how semantically similar their interviews were to different topics or concepts of interest (such as suffering, joy, pain, or disappointment). We entered the replies into a machine learning model to classify and identify if someone responded well to a placebo.
We were in for a surprise.
The patients who responded to the placebo talked more about their emotional experiences, themselves, and their personal relationships. And those who didn’t respond to the placebo talked more about taxing movements and doing physical activities in their interviews. Interestingly, how much their pain changed between treatment and either baseline or no-treatment periods was significantly linked to how much patients talked about their identity and their achievements. These language features ended up explaining 46 percent of the variance in pain relief between patient groups.
The translational potential of our methodology cannot be understated.
Clinicians regularly make use of conversations with patients to understand how much pain they are in, where the pain is located and how their quality of life is impacted. In the future, for example, a standardized list of pain-related questions given to patients during a visit could potentially be automatically transcribed with AI speech-to-text tools and analyzed in real time to provide a physician the likelihood of that patient’s response to a prescribed treatment.
Better understanding and quantifying pain might also have an impact on the ongoing opiate crisis, to help determine whether patients need to be prescribed strong pain medicine or if they might also respond to a placebo or a different drug. It may also help improve clinical trial efficacy and accuracy by providing tools that can identify placebo responders before randomization, allowing for more balanced study designs and treatment groups.
By mid- 2021, we aim to submit a new paper detailing the second set of results from this study, where AI identifies future placebo responders from non-responders before they take any pills. The second paper validates the findings from the first study and expands upon the utility of language by showing that not only can it be used to identify placebo response and quantify it, but also to predict it.
IBM Research’s use of AI to better understand pain management through speech is part of the company’s larger effort using AI and speech to analyze a host of neurodegenerative disorders, including Alzheimer’s, Parkinson’s and Huntington’s diseases, as well as psychiatric disorders such as schizophrenia and addiction. Chronic pain is likewise a neurodegenerative disease and is often associated with mental health issues, including depression, anxiety, and substance abuse.
IBM’s larger mission is to build a digital health platform that can analyze a range of biomarkers, including sleep, movement, and pain and use those metrics to help physicians better understand and treat diseases, using IT and AI to complement the clinical assessments, reaching patients with minimal burden as they go about their daily lives.
Berger, S. E. et al. Quantitative language features identify placebo responders in chronic back pain. PAIN Articles in Press, (2021) ↩