Authors: Alexis Steinberg (1Department of Neurology, Critical Care Medicine, and Emergency Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.; 2Department of Critical Care Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.; 3Department of Emergency Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.), Nicholas Case (3Department of Emergency Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.; 4Department of Biostatistics and Health Data Science, University of Pittsburgh School of Public Health, Pittsburgh, PA, USA.), Yanran Yang (5Department of Engineering and Public Policy, Carnegie Mellon University, Pittsburgh, PA, USA.; 6Global Health Research Center, Duke Kunshan University, Suzhou, China.), Baruch Fischhoff (5Department of Engineering and Public Policy, Carnegie Mellon University, Pittsburgh, PA, USA.; 7Institute for Politics and Strategy, Carnegie Mellon University, Pittsburgh, PA, USA.), Clifton W. Callaway (3Department of Emergency Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.), Patrick J. Coppler (3Department of Emergency Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.), William G. Barsan (8Department of Emergency Medicine, University of Michigan, Ann Arbor, MI, USA.), Ramon Diaz-Arrastia (9Department of Neurology, University of Pennsylvania, Philadelphia, PA, USA.), Romergryko Geocadin (10Department of Neurology, Neurosurgery, Anesthesiology-Critical Care Medicine, Johns Hopkins University, Baltimore, MD, USA.), David O. Okonkwo (11Department of Neurosurgery, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.), Lori Shutter (1Department of Neurology, Critical Care Medicine, and Emergency Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.; 2Department of Critical Care Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.; 11Department of Neurosurgery, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.), Robert Silbergleit (8Department of Emergency Medicine, University of Michigan, Ann Arbor, MI, USA.), William J. Meurer (8Department of Emergency Medicine, University of Michigan, Ann Arbor, MI, USA.; 12Department of Neurology, University of Michigan, Ann Arbor, MI, USA.), Sharon D. Yeatts (13Department of Public Health Sciences, Medical University of South Carolina, Charleston, SC, USA.), Jonathan Elmer (1Department of Neurology, Critical Care Medicine, and Emergency Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.; 2Department of Critical Care Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.; 3Department of Emergency Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA.)
Categories: Article, Decision-making, Heuristics, Prognostication, Traumatic brain injury, Cardiac arrest
Source: Neurocritical care
Authors: Alexis Steinberg, Nicholas Case, Yanran Yang, Baruch Fischhoff, Clifton W. Callaway, Patrick J. Coppler, William G. Barsan, Ramon Diaz-Arrastia, Romergryko Geocadin, David O. Okonkwo, Lori Shutter, Robert Silbergleit, William J. Meurer, Sharon D. Yeatts, Jonathan Elmer
Understanding how clinicians prognosticate is important for creating interventions that improve current practice and clinical trials. We compared clinicians’ reported approaches to prognostication with traumatic brain injury (TBI) and comatose cardiac arrest (CA) for patients enrolled in a multicenter clinical trial.
We conducted semi-structured interviews with clinicians who treated patients with severe traumatic brain injury (TBI) enrolled in the Brain Oxygen Optimization in Severe TBI Phase-3 (BOOST-3) trial (NCT03754114). We compared these reports with ones in a previous study of patients who were comatose after cardiac arrest (CA) and were enrolled in the Influence of Cooling Duration on Efficacy in Cardiac Arrest Patients (ICECAP) trial (NCT04217551). We performed deductive coding using our codebook from CA interviews and then used inductive coding to add new topics raised in the TBI interviews. We looked specifically for reported reliance on initial “clinical gestalt” as observed in the CA interviews.
We interviewed 18 clinicians at 13 hospitals. Predicting poor outcomes was less common with TBI cases than in the CA study, consistent with records showing that final prognostication was determined later in the TBI cases (7 [interquartile range (IQR) 2–18.5] vs. 3 [IQR 2–7] days). Similar percentages of clinicians reported high confidence in their initial prognostic assessments in the two settings (TBI, 33%; CA, 40%). Fewer clinicians reported relying on initial clinical gestalt predictions with patients with TBI (22%) than with patients who had experienced CA (70%). With patients with TBI, more clinicians reported having used later subjective assessments to revise their initial uncertain prognostication.
In interviews with clinicians practicing at multiple institutes, we found that clinicians were less likely to report relying on initial gestalt impressions with patients with TBI than with patients who had experienced CA and were more likely to report relying on later subjective assessments to refine uncertain initial prognostic judgments. Fewer clinicians reported high confidence in initial assessments of patients with TBI.
For patients with acute brain injury, prognostication influences life-altering treatment decisions and can determine patient outcomes [1, 2]. For example, surrogate decision-makers who believe that recovery potential is poor may be more likely to choose withdrawal of life-sustaining therapies (WLST). Because prognostication is challenging [3], with no gold standard tools, prognoses are often inaccurate [4–8], potentially leading to the death of patients who may have recovered [9, 10], or prolonged care of patients who never achieve an acceptable recovery. Variable WLST decisions can also confound the results of clinical trials [11].
In a qualitative interview study of prognostication after cardiac arrest (CA), many clinicians reported relying heavily on subjective “clinical gestalt” when formulating their initial prognostication [12]. Those clinicians typically also reported having early, confident predictions and rarely seeking objective prognostic tests to challenge their initial assessments. We argued that such reliance solely on clinical gestalt to prognosticate could reduce the quality of patient care [13].
Here, we ask whether a similar pattern is observed when prognosticating after traumatic brain injury (TBI), using a similar open-ended, semi-structured interview protocol and coding scheme. Participants in both studies were clinicians in multiple hospitals participating in a large clinical trial [14]. The interview protocol asks participants to describe how they formulated their prognostication for trial patients and how their prognostication evolved over the course of the patient’s hospital course.
We conducted a qualitative study to elucidate clinicians’ cognitive approaches to prognostication for patients with severe TBI enrolled in the Brain Oxygen Optimization in Severe TBI Phase-3 (BOOST-3) trial (NCT03754114). We compared their reports with the reports of clinicians treating patients comatose after CA and enrolled in the Influence of Cooling Duration on Efficacy in Cardiac Arrest Patients (ICECAP) trial (NCT04217551). Both trials were part of the Strategies to Innovate Emergency Care Clinical Trials (SIREN) network. For this study, we specifically define prognostication as the clinician’s cognitive process of estimating a patient’s potential future outcome. We did not address communications of prognosis between clinicians and surrogates.
We recruited clinicians who cared clinically for at least one trial participant. We did not recruit members of the trial research team, some of whom treated patients. We did not interview clinicians who participated in our initial CA qualitative study. The University of Pittsburgh Institutional Review Board (IRB) approved the study (STUDY21080091), which followed institutional and regional ethical standards of the responsible committee on human experimentation and the Helsinki Declaration. We previously published the results of the study of prognostication with patients who had experienced CA [12].
We recruited participants through convenience sampling, seeking a sample that was diverse in terms of clinical specialty and geography. To ensure diverse participant interviews, later in the recruitment process, we used purposive sampling to enroll participants from specific specialties and locations of practice that were not represented by our previous participants. We contacted participating sites prospectively after a suitable patient had enrolled in BOOST-3. We defined a “clinician” as an attending, fellow, resident physician, or advanced practice provider (nurse practitioner or physician assistant). In both studies, we recruited and performed interviews from March 2022 to May 2023. To avoid influencing clinical care or the primary trial outcome, we recruited clinicians after that patient had either died or been discharged from the hospital. We interviewed each participant once, focusing the interview on the one specific patient.
We planned to enroll participants until we achieved thematic saturation, that is, until we heard no new themes. We anticipated recruiting 20–30 participants, as is typically sufficient with a relatively homogeneous population [15, 16]. Ultimately, we conducted 18 interviews, having reached thematic saturation. In qualitative research, 20 participants per group is a typical sample size for thematic analysis comparisons [17]. After each interview, we asked clinicians to complete a structured survey with demographic information, practice location, and local practices related to neuroprognostication after TBI. All participants provided verbal consent before the interviews.
We used the same interview guide as in our previous study of prognostication after CA in the ICECAP trial [12]. It uses a short, semi-structured interview guide with five broad questions on aspects of clinicians’ cognitive approach to making predictions about the future outcome (prognostication) of the patient that they cared (1) When you first met the patient, how did you predict that he/she would do? (2) What informed this prediction? (3) How did your prediction evolve throughout the case? (4) How certain were you about your initial predictions? (5) How did your certainty change throughout the case? After hearing participants’ initial responses to these questions, we used nondirective probes, asking them to expand on each topic. We did not probe on communication, but instead we asked them about their process of estimating a potential future outcome of the patient. We did not ask about specific prognostic tests, but allowed the clinician to describe the tests that they used. Depending on these initial responses, we asked participants about neurologic and non-neurologic factors that influenced their predictions. We did not ask whether patient values influenced their prognostication, because we were exclusively interested in their cognitive process. We did not want the interviews to focus on how end-of-life decisions were made, because we believed that this is an entirely different process. Before conducting the interview, participants were made aware that we were conducting a research study and the purpose of the study was to determine how clinicians make decisions for patients who have had a traumatic brain injury. We performed all interviews in English.
One investigator (A.S.) conducted interviews either by phone, in person, or via online video conferences. A.S. had no prior relationship with the participants before she conducted the interviews. Interviews lasted between 10 and 30 min. Only A.S. and the participant were present during the interview. We audio-recorded, transcribed, and deidentified all interviews before analysis.
We adapted the codebook for the prognostication after CA to analyze the TBI interviews [3]. We used a constant comparative approach, which is a cyclic, iterative process, to organize the data [17]. We first used deductive coding [18] as far as possible to capture constructs observed in the previous interviews. Deductive coding uses a predefined codebook, where parts of the interviews can be organized into pre-defined themes (ideas) and subthemes (more specific ideas) [18–20]. We (A.S. and NC) read the transcribed interviews and iteratively organized them on the basis of the predefined codes in our codebook. One such construct that we focused on was “clinical gestalt,” which we defined as prognostication based on a patient’s overall appearance or a subjective impression in the absence of objective data [12, 21]. We then used inductive coding to add new codes capturing recurrent themes that emerged when participants described their approach to prognostication after TBI. We (A.S. and N.C.) reread the interviews. We iteratively added new codes (themes) that we found in the interviews to the codebook. Then, we coded all interviews with this new, expanded codebook [17, 18, 20]. This combination of deductive and inductive coding enabled us to compare findings in the two studies.
Two authors (A.S. and N.C.) individually coded each interview using MAXQDA 2022 (VERBI Software, Berlin). The two coders then met to review their codes, discuss disagreements, and reach consensus. A.S. is a neurointensivist and physician-scientist who studies clinical decision-making regarding neurologic prognostication. Participants were aware of A.S.’s job and experience. N.C. is a research assistant, flight paramedic, and statistics graduate student. Throughout the coding process, a senior investigator (J.E.) reviewed the analysis for internal consistency and external (medical) validity [22, 23]. J.E. is a neurointensivist and physician-scientist with expertise in neuroprognostication. We applied any coding changes to all previously coded interviews. We sought to maximize the trustworthiness of the data by triangulation, which includes reviewing and debriefing among the three main researchers in the study (A.S., N.C., and J.E.) [17]. We double-coded every interview. We reached thematic saturation (i.e., no new themes were identified) after coding approximately 75% of the interviews. A.S. and N.C. developed Fig. 2 as a graphical representation and visual summary of the dominant themes described by respondents, without adhering to any specific methodology for figure development.
We used descriptive statistics to summarize survey responses. Because BOOST-3 is a blinded, ongoing clinical trial, we did not have access to patient identifiers, the randomized treatment arm, or the actual patient outcomes. All patient outcomes studied here are based on clinicians’ memories. During coding, we defined patient outcomes as (1) good patient awakened or was discharged home or to a rehabilitation center; (2) indeterminate either unknown outcome, minimally awakened (minimally conscious state or persistent vegetative state), or discharged to nursing facility or long-term acute care facility; and (3) poor never awakened during the time that clinician followed the patient or died. Our outcome definition is similar to using the Glasgow Outcome Scale Extended (GOSE) [24], where GOSE 1–2 is a poor outcome and GOSE 4–5 is a good outcome. GOSE 3 is indeterminant because this outcome is dependent on patients’ views of what is acceptable.
During our study period, clinicians were contacted for 77 consecutive BOOST-3 patients across 26 sites. We interviewed 18 (23%) clinicians who were willing to participate in our study and who cared for patients with TBI from 13 (50%) sites before reaching thematic saturation. As with the CA study, most sites were academic medical centers with high enrollments in the clinical trial. The previous study involved 30 interviews with clinicians caring for patients after CA enrolled in 19 ICECAP sites [12]. Clinician participants in both studies were mostly attending physicians and intensivists (Table 1). The time from the patient’s initial brain injury event to the interview date was longer in TBI interviews compared with CA interviews (70 days vs. 20 days; Table 1) because patients with TBI spent longer in the hospital, and interviews could only be conducted once the patient was discharged or died. Clinicians who cared for patients with TBI had more years of practice (19 vs. 8), were more likely to be a neurosurgeon (38% vs. 0%), cared for more patients annually (150 TBI vs. 30 CA), and were more often male (76% vs. 51%).
Clinicians more often initially predicted a poor outcome in CA cases than in TBI cases (54% vs. 33%) (Fig. 1). Based on clinicians’ memories, many patients had a poor outcome in both groups (Fig. 1). Poor outcome was mainly driven by WLST and not by progression to brain death. Clinicians reported high confidence in their initial prognostic assessments in 33% of TBI cases and 40% of CA cases. Final prognostication, defined as the clinician’s last prognostication that would not change further, was determined later in patients with TBI than in those who had experienced CA (7 [interquartile range (IQR) 2–18.5] vs. 3 [IQR 2–7] days). A summary of prognostic tests used for cases wherein the patient underwent WLST is in Supplemental Table 1.
We found different approaches to prognostication in TBI and CA (Fig. 2, Table 2), with both approaches relying on subjective assessments. Clinicians caring for patients after CA more often reported using initial clinical gestalt to make their initial predictions compared with clinicians caring for patients with TBI (21 [70%] vs. 4 [22%]). Representative quotes outlining differences in their prognostication approach are in Table 2, with a visualization of the overall difference in Fig. 2. In most interviews, clinicians described using subjective assessments (clinical gestalt), but the timing of when these subjective assessments influence their prognostication was different between the two groups (Table 2). Many clinicians described an initial judgment after CA that was influenced by clinical gestalt, resulting in highly confident predictions of poor outcome, WLST, and death (Table 2; initial clinical gestalt). By contrast, after TBI, most clinicians stated that it was difficult to prognosticate initially and that they had low confidence in their initial judgments (Table 2; initial uncertainty of TBI), except for the few cases wherein the patient had clear signs of an extremely good (e.g., awake or localizing) or poor (e.g., already herniated radiographically) outcome (Table 2; using objective data for extremes in TBI). Typically, early prognostic uncertainty in TBI led to patients undergoing aggressive initial treatment (e.g., neurosurgery, neuromonitors). Clinicians explained that they used subjective assessments (e.g., severe extracranial organ failure, older age with prolonged ventilator dependence, and refractory intracranial pressure [ICP] elevations) to prognosticate poor outcomes later in the hospital course, resulting in delayed WLST and death (Fig. 2, Table 2; later subjective assessment used to prognosticate).
A representative interview exemplifies this last case. A neurosurgeon reported caring for a young patient with severe TBI who had a “poor” initial exam and developed refractory ICP, requiring surgery. The clinician described their initial prognostication as indeterminate, without relying on clinical gestalt. On hospital day 3, the patient underwent surgical intervention for refractory ICP management. In the operating room, the patient had a brief hypoxic cardiac arrest. Afterwards, the clinician concluded with high certainty that they now predicted a poor outcome for the patient, without additional testing, despite probing of the participant about using any additional tests. Instead, they relied on subjective assessment. The clinician communicated their prediction of a poor outcome to the family, who chose WLST, and the patient died. The neurosurgeon
In our multicenter qualitative interview study, we found that clinicians reported different approaches to prognostication after TBI than in our previous study with CA. In both disease processes, clinicians used subjective assessments to make predictions about patients’ likely outcomes, but the timing of when these subjective assessments influence their prognostication differs. In CA, clinicians relied on their initial clinical gestalt of patient presentation, whereas for patients with TBI, clinicians’ prognostic decisions were influenced by gestalt (subjective assessment) later in the case. Perceived poor prognosis led to WLST and death in both groups, but occurred earlier after CA than TBI.
There are several reasons why clinicians may employ different cognitive approaches to prognostication after TBI compared with CA. For each acute brain injury, the pathophysiology causing coma, the recovery potential, and time to awakening are distinct [25–27]. The heterogeneity of TBI can make prognostication more challenging compared with CA [28]. Cultural and training norms that influence prognostication differ between the two [29, 30]. For example, in CA, the resuscitation community emphasizes following post-cardiac-arrest guidelines [31–33] that instruct clinicians about the timing of prognostic tests and what combination of test results determines a poor prognosis, whereas, in TBI, the neurotrauma community recommends delaying prognostication as much as possible while treating no specific test, combination of tests, or patient demographic as reliable for prognostication [13]. Follow-up is also organized differently in the two domains [3]. For example, neurosurgeons follow patients with TBI into the outpatient clinic, allowing them to observe their long-term outcomes [34]. In contrast, the intensivists who treat patients after CA rarely see them after discharge. Another example is that prognostication training for neurologists and neurosurgeons emphasizes avoiding the self-fulfilling prophecy, teaching clinicians not to have overly certain initial predictions of poor outcome. It is unclear whether the same is true in other specialties. The greater complexity of the TBI prognostication process, seen in Fig. 2, may reflect the design of our study, which added new themes (codes) to the TBI study from our analysis of the CA interviews.
One consequence of the uncertain initial prognoses common to the TBI cohort is allowing more time before WLST, giving patients an additional chance for functional recovery. At least one recent study has found that some patients with TBI who have undergone WLST might otherwise have survived with a degree of functional independence [35], highlighting that excessive death may occur in patients with TBI based on potentially poor prognostication processes. Among the clinicians whose initial prognostication was uncertain, most described reaching a final, certain prognosis at 7 days. Yet, in TBI cases without WLST, the median time to awakening was day 12 [25], with many patients continuing to improve years after their initial event [27]. An overly optimistic, uncertain prognosis can also have negative consequences. Clinicians might put patients through unnecessary procedures or have surrogates make choices for patients that are incongruent with their values and preferences.
The accuracy and quality of subjective assessments used by TBI clinicians later in the hospital course is unclear. Given the minimal research on prognostic tests in TBI, it is reasonable that clinicians rely on subjective assessments, which may be rooted in well-established experience and decision-making. However, subjective assessments are likely influenced by cognitive biases and heuristics, which could result in inaccurate prognostication [1, 34, 36]. The effects of these biases are compounded when clinicians do not receive feedback on their decisions, such as when WLST occurs, and the patient’s counterfactual outcome remains unknown [34]. A potential consequence of cognitive biases in TBI prognostication is the presence of disparities in patients undergoing WLST on the basis of the patients’ demographics (i.e., age or ethnicity) or insurance-related factors (i.e., higher rates of WLST are found in patients with TBI who are uninsured) [37, 38]. For instance, clinicians may overweigh a patient’s age or insurance status when making their predictions.
Our paper is the first study to explore the cognitive approach to prognostication in TBI. Elucidating the mechanism of prognostication is the first step in improving the way clinicians prognosticate. Next steps to further define the mechanism of clinician prognostication in patients following acute brain injury would be to use a think-aloud [39] or ethnographic approach [40], where we prospectively observe clinicians in real time and explore how they make their prognostic decisions. We could better identify how and when clinicians use subjective assessments in their prognostication, which is important for creating targeted interventions that improve the way clinicians prognosticate.
Prognostication and WLST can bias clinical trial results when WLST determines a patient’s outcome [11]. WLST is influenced by both measured and unmeasured factors, including the way clinicians prognosticate. Inaccurate prognostication or the use of inappropriate subjective assessments can lead to imbalanced deaths between the two trial arms, thereby biasing the estimate of the treatment effect [11]. This bias persists despite randomization, as it occurs as a postrandomization intervention. Future studies should determine whether current prognostication practices bias TBI clinical trial results.
Our research study has several strengths. By using a qualitative methodology with open-ended interviews, we were able to explore clinicians’ clinical reasoning, allowing them to reflect on their experience in their own words and natural framing of the clinical situation. This method is less likely to result in them telling us what they think they want us to hear; instead, they are more likely to describe their true approach to prognostication. A second strength is asking clinicians to recall their decision-making in an actual clinical case, rather than asking them to summarize over an unspecified sample of cases [15]. A third strength is working with clinicians who practice in multiple hospitals across North America, supporting the generalizability of our findings [19]. A fourth is reaching thematic saturation in both studies, giving us confidence that we have heard any theme occurring with any frequency in these populations.
One limitation of the study is that we conducted interviews after the cases had ended, rather than in real time, exposing clinicians’ reports to hindsight, outcome, and recall bias [41]. This bias may have been worse in the TBI interviews compared with the CA ones because of the longer time from the initial event to when the interview was conducted. A second limitation is not being able to evaluate the decision-making processes described in interviews in light of the case outcomes. A third limitation is not asking clinicians how they responded to patients’ preferences or how they communicated their prognostication to patients’ loved ones. Prior research has shown substantial variability in the way clinicians communicate their prognostication [42], and future qualitative studies should explore how this variability influences prognostication and treatment decisions. A fourth limitation is not knowing how well clinicians’ reports corresponded to their actual decision-making processes. A fifth is our sample, restricted to 18 clinicians currently practicing in major academic hospitals in North America, mainly the USA, leaving it unclear how these results generalize to other places. We are missing clinicians who are fellows or advanced practice providers, so we cannot generalize our results to all clinicians. The small sample, common to qualitative research, is designed to hear individuals’ thoughts about complex decision-making, allowing for limited generalization and laying the foundations for larger samples and more structured studies. The small sample size also results in a sixth limitation, which is that we were not able to assess for associations between subjective assessments and physician-, hospital-, or patient-level characteristics. Future research should explore whether these potential confounders play a role in prognostication.
Analyzing qualitative interviews with clinicians participating in a multi-hospital study of TBI cases, we found that clinicians were less likely to report relying on initial clinical gestalt impressions than with patients who had experienced CA and were more likely to report relying on later subjective assessments to refine uncertain initial prognostic judgments, which often led to WLST. Rigorously characterizing the phenomena, we identified that using prospective qualitative methodologies could reduce bias introduced by WLST in future clinical trials.
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1007/s12028-026-02452-z.