The Linguistic Effects of Context Specificity: Exploring Affect, Cognitive Processing, and Agency in Physicians’ Think-Aloud Reflections

Affect, cognitive processes, and agency in clinical reasoning

While much of the clinical reasoning literature focuses on the logical aspects of cognition, emotion (an affective state characterized by arousal that results from a specific stimulus in the environment [9]) is also an integral part of the reasoning process [9], [10], [18], [19]. The greater the magnitude (increase in level of arousal) of the emotion, the greater the possible effect on the clinical reasoning process [11]. But some emotional arousal is present in all reasoning [19], particularly in the high-stakes context of patient care, where anxiety and stress often exist [9]. Negative emotions like anxiety can cause a narrowing of attention and risk aversion (along with potentially resultant narrowing of cognitive capacity and increasing cognitive load), which, in turn, can increase the chance of medical error in the form of missed or delayed diagnoses [11], [20], [21], [22]. Positive emotions can often support reasoning, but they can also lead to overconfidence which can, in turn, result in less information gathering during a patient encounter [11], [20], [23]. Recent research has also linked emotions to contextual factors, with study participants voicing primarily negative emotional reactions to various contextual factors such as low English proficiency of the patient [24], [25]. One potential solution proposed by those interested in ameliorating diagnostic error is to increase explicit awareness of these emotions, taking them seriously and exploring how different contextual factors may trigger different emotional states [10], [11]. Linguistic analysis is one such path to increased awareness of emotions [12], [26], offering a novel way of exploring how various emotions are triggered by contextual factors.

While the major outcomes of clinical reasoning may be diagnostic and management plans, clinical reasoning itself is also a complex process of meaning making that scholars are only now beginning to fully understand [25], [27], [28], [29]. Contained under the umbrella of the clinical reasoning process are narrower cognitive processes like problem representation, hypothesis generation, hypothesis testing, and metacognition (which involves, among other things, being aware of, controlling, and managing one’s cognition in pursuit of a task [30], [31], [32]). Sometimes these cognitive processes are conscious and sometimes they are unconscious [32], and they appear to be inhibited by certain contextual factors, such as interruption and diagnostic suggestion [24], [28], [33]. The presence of these inhibiting contextual factors can increase cognitive load, defined as perceived mental effort [33], [34]. When the cognitive load is too high for a clinician, their reasoning can be negatively affected, leading to diagnostic error [33], [35]. In order to study cognitive processes in clinical reasoning and how they may be related to cognitive load, we can explore the patterns of distinct linguistic markers like think, know, or consider [36]. This allows us to examine whether and how expression of cognitive processes shifts under the influence of context specificity. Following Khawaja and colleagues, we predicted that higher cognitive load would be associated with more cognitive process markers as individuals worked to actively understand their situation [36].

While situated and context-dependent, the process of clinical reasoning is largely directed by the physician. It is the physician who marshals the necessary resources – some of which may involve other people (e.g. specialty consultation) and diagnostic artifacts (e.g. diagnostic imaging, laboratories) – to eventually reason to a decision. Yet, as discussed earlier, physicians may feel uncertain or anxious in the presence of contextual factors, letting their emotions guide their reasoning [20], [21], [24]. We approach this through the lens of agency (broadly defined as the capacity to produce an effect [7], [8]), exploring whether contextual factors affect how physicians talk about themselves as agents (or not) of the reasoning process. In particular, we examined the frequency of the first-person singular pronoun I, as it has been argued in prior work to be indicative of a feeling of individual, intentional causation, particularly in comparison to other pronouns like generic you [8], [15], [37]. Moreover, the first-person singular pronoun has been associated with greater depth of reflection in medical student essays [38] in one study and decreased cognitive load in team problem-solving in another [36]. These studies along with the broader cognitive load literature suggest additional reasons why we might expect to see decreased I usage in the presence of contextual factors.

In order to examine and describe potential effects of context specificity on affect, cognitive processes, and individual agency so that we can better support clinicians, we pose the following research questions:

1. Does the presence of contextual factors in cases lead to differences in linguistic measures of affect, cognitive processes, or individual agency?

2. If so, what are the patterns of different subtypes of affect, cognitive processes, and agency in cases with and without contextual factors?

Based on the literature reviewed earlier, we hypothesized that increased cognitive load in the condition with contextual factors would lead to a greater frequency of affect and cognitive process markers and a lower frequency of first-person singular pronouns.

Think-aloud procedure

For the think-aloud procedure, participants were asked to speak their thoughts out loud, without making judgments or offering insights, as they engaged with the task (e.g. a video of an event [41]). Past work has indicated that think-aloud transcripts represent a reasonable measure of thinking [41], [42], [43], as well as an effective way to assess clinical reasoning [44], [45], [46]. In this study, participants were given brief instructions and a warm-up exercise in the think-aloud method prior to engaging in the cases. Then, after either viewing the video case or participating in the live scenario and determining the diagnosis and management, they were prompted to think aloud about their thoughts leading to diagnosis and treatment. Participants were given up to 30 min to complete this and were allowed to stop or rewind the video.

Data analysis

To understand how the process of clinical reasoning is affected by context specificity, we used Linguistic Inquiry and Word Count (LIWC) software. LIWC is a transparent (i.e. coded words and phrases are accessible to researchers) text analysis program that codes for affect, cognitive processes, and agency, among other psychological processes [12]. We coded all transcripts with LIWC for the broad categories of affect and cognitive processes and the subcategory of first-person singular pronouns (i.e. I and me). To control for the potential effect of varying word counts, LIWC calculates a percentage of coded categories per 100 words (e.g. if there were 10 affect-related words in a 200-word transcript, LIWC assigns that transcript a value of 5% for affect). We then conducted a repeated-measures multivariate analysis of variance (MANOVA) and follow-up univariate analyses with affect, cognitive process, and first-person pronouns as the dependent variables, comparing participants’ language in the cases with and without contextual factors.

To explore patterns in affect, cognition, and agency, we examined descriptive statistics of the subcategories making up affect [positive emotions (for which LIWC has no subcategories) and negative emotions (for which LIWC denotes the subcategories of anxiety, anger, and sadness)] and cognitive processes (insight, causal processes, certainty, tentativeness, discrepancy, and difference – described in greater detail in the Results section). We also examined descriptive statistics of other personal pronouns (we, you, he/she, and they) to better understand how individual actions interacted with the actions of others in these data.

The effects of context specificity: affect, cognitive processing, and individual agency

Repeated-measures MANOVA results revealed significant differences in cases with and without contextual factors [Pillai’s trace=0.22, F=5.6, df=(3, 61), p<0.01]. Follow-up univariate analyses indicated that participant language contained statistically significantly more cognitive process markers in think alouds of cases with contextual factors. There were also more affective markers in cases with contextual factors, but this difference was not statistically significant. Additionally, in think alouds without contextual factors, participants used more first-person singular I/me pronouns, suggesting perhaps a greater expression of individual agency, although this difference was not statistically significant (see Table 2).

In order to better understand the differences in affect in cases with and without contextual factors, we explored LIWC’s affect subcategories: positive emotions, negative emotions, and the three subcategories of negative emotions, anxiety, anger, and sadness (see Table 3). The difference between conditions resulted from more negative emotions in cases with contextual factors, where participants thought aloud about the standardized patient’s emotions (stress, anxiety) and their own thought processes (e.g. thinking “that’s ridiculous” about a potential diagnosis of a coal worker’s lung). LIWC also identified some medical terms (e.g. stress test, head trauma, resolves with rest), but these uses appear in both conditions (with and without contextual factors).

Next, LIWC’s cognitive process category derives from six subcategories: insight, causal processes, certainty, tentativeness, discrepancy, and difference (see Table 4 ). The greatest contrast appears to be in terms of insight (terms associated with learning or understanding like think, explain, evaluate, or consider [12], [47]): participants talked more about their learning or understanding when contextual factors were present, more often explicitly reflecting on their thinking or considering. While the other differences were not as great, it is notable that, in the presence of contextual factors, participants seemed to use fewer markers of certainty (terms indicating a certain level of conviction like clear, sure, certainly, or namely [12]) and more markers of tentativeness (terms indicating a hedging or uncertain stance like kind of, may, if, or anything [12]). Similarly, participants made more discrepancies (terms indicating a difference between an actual and possible state like should, would, could, and need [12]) in the presence of contextual factors, often conveying a speculation about what could or would be the case, given some condition (e.g. “Her HCTZ [dose] could be improved….[so] her lifestyle could improve.”). Finally, markers of causal process (terms implying that one thing gives rise to another like how, based, because, or why [12]) and difference (terms of distinction, including negation, like but, really, not, or other [12]) appeared to be similar across conditions.

Finally, examining agency beyond first-person pronouns (discussed earlier), participants appeared to use third-person pronouns in the presence of contextual factors to focus on the actions of others, often the patient or, in the video cases, the doctor depicted in the video (Table 5). In fact, thinking aloud about the actions of third-person singular others was the most common pronominal use across conditions (between 4.8% and 5.5% of the word count), but these exploratory analyses suggest that he/she/him/her usage goes up in the presence of contextual factors as I/me goes down. Despite the increase in cognitive processing words (which often have I subjects) in the presence of contextual factors, the overall focus on the self’s actions yielded to thinking aloud about the patient and video doctor, often with reference to a contextual factor. For example, here a participant reflects on the case with a patient who is not a native speaker of English: “She [patient] asks him [doctor] about speaking Spanish and he says he only speaks English”. This participant only referred to herself eight times (0.9% of words) in this case, while she referred to herself 45 times (4.3% of words) in the non-contextual factors case.