This entry is a part of the NYU Center for Data Science blog’s recurring guest editorial series. Angela Radulescu is a CDS Faculty Fellow.
Computational psychiatry brings quantitative tools to bear on the way mental illness is studied, classified and treated. A common aim across different cultures of computational psychiatry is to provide a set of features that, when inferred over time at the individual level, can predict both symptom outcomes, and the effect of different therapeutic interventions. The underlying assumption is that various behavioral and neurophysiological measures (e.g. choices, reaction times, speech patterns, gaze trajectories, etc.) can be summarized as individual data points in a useful feature space. The hope is that the structure of this space will map onto a diagnostic taxonomy; that process models can bridge the gap between behavior and its biological determinants; and that the dynamics of data points in this space can inform therapeutic interventions.
A question I always come back to when thinking about my own work in this area is, how can a computational approach account for the heterogeneity of biological and environmental factors specific to each individual story and subjective experience? Engaging with this question in a precise way is quite challenging, as it would require a record of a person’s significant past experiences, psychosocial context, subjective interpretation, and so on. Together, these form a personal narrative, and are often key pieces of information in clinical assessments.
Imagine for example two people who suffer from anxiety and depression. The cognitive and behavioral manifestations might be the same (persistent attention to negative information, avoidant behavior, etc.), rendering these two people indistinguishable in a feature space inferred from behavior and symptoms alone. But the root cause of their symptoms could be very different: one person might struggle with issues surrounding their sense of identity. The other may be overcoming past trauma. Therapy relies on the interaction between patient and therapist to examine and reshape such narratives during treatment.
Computational psychiatry can benefit from developing principled models of how different facets of personal narrative manifest in behavior. Exactly how to do this is a fundamentally interdisciplinary and collaborative research endeavor that would integrate various data science approaches with different areas within the psychological and clinical sciences.
For example, an important aspect of narrative is understanding how behavioral measures evolve over time. By leveraging readily available computing platforms accessible to diverse groups of individuals, it is possible to conduct dense sampling of a variety of measures, and study how they change in day-to-day life.
Aligning data collection with people’s daily environments can also be done by considering naturalistic study designs designed to assess the ecological validity of computational models. One area that holds promise in this regard is the “text-as-data” approach, which uses natural language processing (NLP) methods to extract meaningful information from text. For example, applying NLP to text transcripts could help understand the processes that occur during therapy. Closer integration of NLP techniques such as sentiment analysis (which can detect the emotional valence of complex utterances) with cognitive modeling (which aims to build precise process-oriented models of behavior) may enable predictive modeling of the interaction between narrative and behavioral outcomes in situations relevant for clinical practice, such as conversations and autobiographical recall.
This analysis approach is especially relevant for behavioral research in mixed reality platforms. Virtual reality has already enabled the piloting of immersive treatment of psychiatric conditions including anxiety and schizophrenia. Instantiating and testing computational models of behavior in virtual reality settings could advance the study of how subtle cognitive biases associated with mental illness manifest in people’s interaction with day-to-day environments; and if and how such biases factor into each individual’s own narrative of their experience.
Ultimately, the primary stakeholders in clinical research are the people whom the research aims to help. Their narratives have to be part of the conversation when asking questions and designing data-driven interventions. As my colleague Sarah Shugars put it in a previous post, “who asks and answers questions matters.” And our questions are only as good as our analytic tools. By engaging with personal narrative; by making narrative part of the research process in an ethically responsible way; and by ensuring narrative has a place in our analysis, our questions can only get better. And they may yield more actionable answers.
By Angela Radulescu