Unusual gastrointestinal histoplasmosis with jejunal perforation in a 57-year-old patient from Bogotá: Case report.

Psychiatric assessment is challenging – boundaries are blurry between normality and clinical impairment, and between different diagnoses. Making these distinctions involves extracting diagnostic information from large amounts of data about a person’s unique experiences within their psychosocial context. That is, it involves listening to idiosyncratic narratives and translating them into the vocabulary of psychopathology signs and symptoms that allow us to understand a person’s impairment, compare them to others, and facilitate clinical communication. Clinical interviews are the gold standard because they have been the only methods that allow such integration and translation. However, interviews are not scalable or portable because they are time-consuming, costly, and require considerable expertise or training. Large-scale and intensive assessments, in both clinical and research contexts, have therefore often relied on patient report scales. Despite their psychometric strengths, scales are rarely tailored to the individual’s issues, lack contextual information, and are narrow in scope. This tradeoff between assessments that are individualized and context-sensitive or scalable and portable has limited our knowledge of psychopathology. Natural language processing offers a solution to this tradeoff by automating psychopathology assessment from personal narratives. An ability to extract information from natural language is important because a person’s description of themselves, their circumstances, and their problems – in their own words – is arguably the richest source of diagnostic information we have (Pennebaker, Mehl, & Niederhoffer, ). However, for decades, automated language processing alternatives to resource-intensive human raters were limited to simple metrics like word counts. Word counts ignore context and cannot capture the nuance needed to assess psychopathology. Newly introduced transformer-based language models are completely different. Transformer-based models are trained on massive amounts of natural language data to learn language patterns. Then, unlike older models that process discrete, decontextualized words, transformer-based models process sequences of words, considering their interrelationships. Owing to this technological advance, these newer models encode the meaning of words within the broader of the narrative (Eichstaedt et al., ). Accounting for context in natural language analysis offers unprecedented capabilities for assessing psychopathology (Kjell, Kjell, & Schwartz, ). This is because symptoms are often implied rather than stated. For example, the meaning implied by a sentence like ‘I keep doing things I’m supposed to enjoy, but nothing lands anymore’ is the experience of anhedonia even though a specific symptom is not named. Perhaps more fundamentally, discerning normality from clinical impairment hinges on the context of behavior (Hopwood, Wright, & Bleidorn, ). A phrase like ‘I felt stressed’ could indicate impairment if the stress is in response to a minor hassle and becomes overwhelming but could be normal if it was manageable stress in anticipation of a high-stakes job interview. Likewise, context is necessary to distinguish different kinds of psychopathology. For example, distinct forms of psychopathology would be implied by about social evaluation versus signs of physical illness. Transformer-based models, in principle, can make such nuanced distinctions when analyzing text. Accordingly, these models can then compare rich, diagnostic information detected in text to its trained knowledge base and predict the kind and severity of psychopathology evident (if any). In essence, transformer-based models could solve the scalability problem by performing functions of trained human raters in seconds or minutes rather than the days or months needed to train raters and hand-code texts.