PROJECT SUMMARY
Chronic pain is still a clinical diagnosis based on location, symptom report, and clinical expertise. Despite recent
efforts to delineate specific and evidence-based criteria to diagnose different chronic pain conditions, substantial
heterogeneity persists among chronic pain patients often within the same clinical pain syndrome (e.g., low-back
pain). The lack of quantitative and reliable measures to diagnose chronic pain and the related heterogeneity
that ensues are major obstacles to medical care for patients and for research studies. Chronic pain patients are
often managed using a “trial and error” approach as targeted and precise treatment is not possible without
quantitative biomarkers, like glucose levels for diabetes. In addition, patient-related variability in analgesic
response is thought to be one of the main reasons why the current therapeutic interventions for chronic pain are
unsatisfactory, as 20% of US adults live in chronic pain and 8% of US adults are disabled from chronic pain.
Natural language processing analyzes semantic and emotional content, syntactic structure, and complexity of
speech; audio-visual processing analyzes voice acoustics and facial expressions. These tools have recently
been shown to be powerful quantitative and reliable biomarkers for discriminating between patients with
psychiatric conditions like schizophrenia and major depression, and in predicting long-term outcomes, like the
development of psychosis in high-risk groups. A parallel can be drawn between chronic pain and chronic mental
illness like major depressive disorder, as both conditions are diagnosed based on subjective report of symptoms,
diagnostic criteria, and clinical expertise. In addition, both conditions are closely associated with negative affect
which has been corroborated by preclinical research and brain imaging data showing a critical role of the limbic
brain in the pathophysiology of these conditions. Therefore, it stands to reason that natural language and audio-
visual processing may serve as biomarkers to phenotype different types of chronic pain patients and to measure
patients' responses to treatment.
This proposal will study the ability of language analysis and audio-visual processing tools in discriminating
between different types of patients with chronic pain (i.e., discriminant validity) in Aim1, and the ability of these
tools to predict analgesic response of chronic low-back pain (CLBP) patients receiving spinal cord stimulation
(SCS) (i.e., predictive validity) in Aim 2. In both aims patients will be video recorded during an interview where
they speak about their pain or mood (for major depressive disorder patients). Language, speech, and facial
expression features will be extracted from the recordings and used in multivariate machine learning models. In
Aim 1 natural language and audio-visual processing patterns will be compared between patients with 3
conditions: (1) musculoskeletal CLBP, (2) musculoskeletal CLBP with clinically significant negative affect, and
(3) moderate major depressive disorder. In Aim 2, natural language and audio-visual processing patterns will be
used to identify responders and non-responders to SCS.