Using Data Science To Center Patient Perspectives in Mechanism Discovery

Demand for cost-effective, novel, scalable trauma-focused interventions is high. Prevalence rates of posttraumatic stress disorder (PTSD) and Complex PTSD in US community mental health clinics are estimated to be as high as 50%. Yet response heterogeneity to PTSD interventions remains high, with non-response rates reaching 50-60% and dropout rates for traditional interventions (i.e., cognitive behavioral, exposure therapies) at 30-40%. Moreover, research populations in typical stepwise, efficacy-driven clinical research trials are often characterized by strict exclusion criteria and low representation from underrepresented communities. The homogeneous nature of efficacy-based research populations creates an incomplete picture, especially in public sector and community-based mental health facilities. Studies have suggested not only does this homogeneity limit effectiveness across diverse populations, but may contribute to exacerbating health disparities.

Engaging patient perspectives is crucial to research because it can provide insight into response heterogeneity and engagement, ultimately leading to an understanding of mechanisms and creating more patient-centered interventions. One way to center the patient's voice and increase the potential of identifying unique mechanisms of change for a novel therapy, is to use qualitative interviews because it directly accesses the lived experience and its context. Despite the potential benefits of utilizing qualitative data in stepwise randomized control trials, several obstacles persist, including resource constraints, the inability to quantify interactive elements, and concerns regarding the practical value of the gathered information. Innovative methods that reliably and rapidly extract value-laden, relevant themes, and discern non-verbal conversational elements may facilitate the integration of patient experience and inclusion of their perspectives in clinical intervention trials.

This single-arm study aims to evaluate the feasibility of using natural language processing (NLP) and deep learning to identify potential mechanisms of PTSD symptom change from patient interviews. The study will utilize ongoing cohorts from a clinical service that offers a 16-week, live-online group therapy model (and adjunct individual therapy sessions) called Program For Alleviating And Resolving Trauma and Stress (PARTS) that uses the IFS model. The investigators will use a convergent mixed methods approach applying machine learning and natural language processing to develop models that identify potential mechanisms of change.

Analysis: The investigators will use several different methods to develop our models including Latent Dirichlet Allocation, pre-trained language models, transfer learning (recurrent neural networks, generative adversarial network), and penalized regression-based models. These models will use data derived from patient perspectives of psychological change, as expressed in interviews, and will be compared to models derived from clinical measures. The study will use standard performance metrics and cross-validation scores to evaluate comparative performance of the models. As an exploratory aim, the study will evaluate the feasibility of using features derived from language processing models and clinical measures to predict individual therapy visits post-intervention. The exploratory data will also include structured clinical data, social determinants of health, and therapy-based utilization (dates, provider type, length).

Anticipated results: The development of two validated models: one derived from patient interview data and the other based on clinical measures to comprehensively identify mechanisms of change from group-based therapy models of IFS for PTSD.