Mental health diagnosis
Most mental disorders are highly complex and have high phenotypic variability, partially vague diagnostic criteria, and a significant overlap ratio. A single clinical examination is not representative and is susceptible to the framing effect and other cognitive biases. The patient’s subjective experience is often considered precious in diagnosis, but is often severely neglected due to its vagueness, variability and an idiosyncratic presentation of symptoms. The overwhelming majority of psychiatric disorders appear to be dimensional in nature. Consequently, their categorizations become artificial and debatable.
Subjectivity in mental health
The lack of systematic integration of patient lived experience and collaboration in clinical research hinders psychiatry’s advancement in correlating psychiatric symptoms with neural pathways and mechanisms. A fundamental issue lies in the unexplored notion of subjectivity, which is molded by prevailing common sense assumptions. Frequently, subjectivity is misconceived as the opposite of objectivity – hidden and unreliable. This misunderstanding is then paired with the belief that phenomenology merely uncovers the unique circumstances of an individual. This negative understanding of subjectivity as something solely private or elusive stems from an adherence to methodological traditions inherited from behaviorism.
Combining phenomenological, computational, and evolutionary approach
The research team takes an innovative approach to understanding mental health by combining phenomenological, computational, and evolutionary perspectives. The approach illuminates the cognitive biases in autism spectrum (AS) and schizophrenia spectrum (SS) not as mere anomalies, but as neurodiverse manifestations with potential adaptive value in ancestral environments. We explore how these biases – ranging from detail-focused cognition in AS to over-attributed intentionality in SS – might have served as compensatory mechanisms in specific ecological niches, offering advantages under certain conditions.
Mental health support: therapeutic chatbots
The team also studies therapeutic bots reliability in bias identification and rectification, effectiveness, and reaction to challenges, as well as how they handle emotional and affective dimensions of conversations. The bots, leveraging advanced machine learning technologies, interact with users often displaying biases such as anthropomorphism, overtrust, attribution, illusion of control, fundamental attribution error, and the just-world hypothesis, although their handling of affect might outweigh the cognitive restructuring advantages of general-use chatbots for some individuals.
Unveiling stereotypes in AI: autism representations
The team also explores how artificial intelligence handles stereotypes, focusing on autism. Unlike humans, whose stereotypes are complex and multifaceted, AI’s stereotypes mainly stem from the training data’s quality. Despite the safeguards against perpetuating stereotypes, biases are evident in the images LLMs generate, often relying on contentious symbols like the puzzle piece, meant to symbolize autism, and predominantly depicting individuals on the spectrum as white males. This not only reinforces harmful stereotypes in society but also impacts equal access to diagnostics.
Virtual expert by experience
The team creates of unique corpus comprising first-person narratives of individuals on the autism spectrum, the first globally to provide such specialized, preprocessed data for AI applications. It offers invaluable insights into their unique social interactions, interpersonal relationships, and the accompanying emotional and sensory experiences. Finetuning LLMs using such corpus can significantly contribute to the development of debiasing and educational tools thus to support the work of medical professionals. This model, adapted to reproduce the specific experiences of autistic individuals, offers broad developmental and application possibilities in the form of response-specific modules.
Cognitive Humility in Large Language Models
One of the most significant components influencing the reliability of expert decisions is the level of cognitive humility. In human experts, this parameter is measured using appropriate questionnaires, a methodology inadequate for evaluation of LLMs and decision support applications. The team conducts a comparative analysis of the decisions concerning mental health issued by human experts and AI, taking into account aspects such as the application of non-rational beliefs, succumbing to stereotypes, awareness of the potential for errors, and decision justification.
Google Scholar
Przemysław Kornatowski