Leveraging What We Know: Knowledge Discovery and Knowledge Exploitation as Complementary Aspects of Explainable AI

Natural Language Processing is experiencing a phenomenal capabilities acceleration since the introduction of Large Language Models (LLMs). This success stems in part from the ability to ingest diverse and vast amounts of unstructured text to quantify normative, unlabeled language. It is not obvious however how to make use of LLMs to analyze language produced in the context of psychiatric evaluation and intervention, and how to incorporate decades-worth of neuropsychiatric knowledge in the form of questionnaires, inventories, constructs, manuals, papers, etc., not to mention the personal experience of psychiatrists and patients. We will discuss recent efforts to leverage this knowledge in several applications, including analysis of interviews for differential diagnosis of psychosis and its risk states, discrimination of fibromyalgia and neuropathic pain, prediction of treatment outcomes in cocaine addiction, and inference of progression in psychotherapy. We will also discuss similar efforts in modeling multimodal data streams, e.g., the combination of qualitative assessments and neuroimaging in the clinic with at-home digital data in ongoing clinical trials.

Learning Objectives:

1.  Identify how current language analytics can be applied in different mental health assessment tasks.

2. Generalize the concept of Interpretable AI to models arising from different data modalities and analytic tasks.

3. Evaluate the advantages and drawbacks of multimodal classification/regression models.