Key physician behaviors that predict prudent, preference concordant decisions at the end of life

Abstract

Background:

This study introduces an empirical approach for studying the role of prudence in physician treatment of end-of-life (EOL) decision making.

Methods:

A mixed-methods analysis of transcripts from 88 simulated patient encounters in a multicenter study on EOL decision making. Physicians in internal medicine, emergency medicine, and critical care medicine were asked to evaluate a decompensating, end-stage cancer patient. Transcripts of the encounters were coded for actor, action, and content to capture the concept of Aristotelian prudence, and then quantitatively and qualitatively analyzed to identify actions associated with preference-concordant treatment.

Results:

Focusing on codes that describe characteristics of physician-patient interaction, the code for physicians restating patient preferences was associated with avoiding intubation. Multiple codes were associated with secondary measures of preference-concordant treatment.

Conclusions:

Prudent actions can be identified empirically, and research focused on the virtue of prudence may provide a new avenue for assessment and training in EOL care.

Introduction

Serious illness can compel patients and their families to face difficult questions about the values, goals, and preferences that should inform end-of-life (EOL) care trajectories. Whether clinicians hinder or help patients/families in these considerations depends on numerous factors, including clinical skills for guiding challenging conversations at various junctures along the continuum of care. Despite the decades-old legal and ethical consensus permitting patients and clinicians to forego life-saving treatments, persisting cultural and existential challenges experienced by patients and their families during end-of-life decision-making require that clinicians employ effective communication skills during the end-of-life process.

Landmark studies have shown that “high-quality end-of-life care results when health care professionals ensure desired physical comfort and emotional support, promote shared decision-making, treat the dying person with respect, provide information and emotional support to family members, and coordinate care across settings” (Teno et al. 2004). Related research has informed educational initiatives supporting EOL care through acquisition of discrete clinical skills focusing on communication (Anstey et al. 2016, Fineberg, Kawashima, and Asch 2011, Szmuilowicz et al. 2010). Optimally, these skills are understood within broader ethical frameworks that include shared decision-making and the relevant principles and virtues. Whereas training in technical skills (e.g. inserting a catheter) can improve proficiency in specific health care activities, studying humanistic skills (e.g. mindful communication in end-of-life care) may contribute to the formation of the kind of humane physicians required for high-quality EOL care. By understanding the current practices of physicians seeking to provide this humanistic care, we hope to facilitate the acquisition of these skills by current and future practicing physicians.

To date, the medical ethics literature regarding EOL decision-making has focused on questions of morally permissible action (e.g., withdrawing and withholding treatment) and the importance of respecting patients’ autonomy. Much less attention has been given to the practical benefits of virtue-based concepts for identifying relevant and specific behaviors that constitute clinical excellence in this area of health care. Barriers to operationalizing the virtues are many: “The extent to which a course of action is “virtuous” is a complicated matter that requires much detailed knowledge and careful judgment of people and the situations in which they act” (Kotzee, Paton, and Conroy 2016). Nonetheless, there may be theoretical and practical advantages to a virtue-based approach to EOL care. Given the persistent difficulties in effectively empowering patients to exercise autonomy in EOL care, a virtue-based approach may provide new approaches that could resolve what appears to be an impasse.

For example, imagine testing the following hypothesis: Clinicians exhibiting practical wisdom (e.g., prudence) in goals of care conversations are more likely to develop a medically appropriate EOL care plan that is consistent with the patient’s preferences. Successful demonstration of this hypothesis not only would bring a cardinal virtue into the realm of evidence-based medicine, but also, and more importantly, could transform how shared decision making and practical wisdom are understood and taught across health care specialties. For EOL care specifically, improvements might include increased attention to the dispositional habits and skills of facilitating a decisional process (e.g., means) that is consistent with the intended outcome (e.g., ends). Methodological skepticism is warranted not because the hypothesis above cannot be tested, but rather because the conceptual infrastructure to support this kind of ethics research needs development. How would a practically wise performance be recognized in discrete goals of care conversations? What observable behaviors [e.g. words, modes of communication] would constitute a practically wise interaction in terms of the clinician’s contribution? What relationships can be established between prudent clinician actions and decisional outcomes?

Questions like these motivated the study reported here. Specifically, we sought to test whether the theoretical construct of prudence could be effectively measured and analyzed in a sample of simulated patient encounters where physicians encountered a dying and decompensating patient-actor with clear EOL preferences that, when elicited and respected, would lead to a preference-concordant treatment outcome. If empirically prudent behaviors could be identified (e.g., ones associated with the desired outcome of avoiding intubation in favor of palliative care), then these behaviors could serve as the building blocks for constructing an evidence-based, prudent practice of EOL care. Establishing empirical ground for a virtue-based attribution of excellent performance could also promote wider understanding of how the virtues translate across various care settings as Aristotle predicted they should.

Methods

This study is a mixed method analysis of qualitative codes applied to transcripts of a multicenter simulation study. All procedures were approved by the University of Pittsburgh Institutional Review Board. The simulation was conducted at three academic medical centers as described in detail elsewhere (Barnato et al. 2014, Mohan et al. 2010). In brief, the simulation asked physicians to evaluate a patient with impending respiratory collapse. The actors simulating the patient and family member were coached to share a preference against mechanical ventilation, favoring instead a palliative care plan that would ease symptoms and avoid intubation as the patient died from the last stages of disseminated cancer. However, the actors shared these preferences only when physicians actively solicited them. As the simulation proceeded, the patient’s physiological parameters deteriorated such that the physician was forced to decide whether or not to intubate the patient and transfer the patient to the intensive care unit (ICU). Participating physicians had primary training and clinical practice in internal medicine, emergency medicine or critical care medicine. Each simulation was recorded and transcribed for analysis.

Codebook Construction

Codebook construction was aimed at characterizing behaviors manifesting the Aristotelian virtue of “practical wisdom” (Greek: phronesis), defined as the capacity not only to discern the right and good outcome (e.g., avoiding intubation while easing the symptoms of dying), but also the capacity to practically achieve the good outcome through means consistent with the intended good (e.g., respect for persons at the end of life) (Aristotle 1987, Kaldjian 2010). Codebook construction began without reference to the transcribed simulations based on the authors’ (DH & JF) theoretical understanding of practical wisdom. The codebook was then refined through an iterative process referencing the transcripts and in collaboration with the lead coding analyst (KS). The codebook consisted of agent codes (e.g., patient, physician, nurse, etc.), action codes (e.g., asking a question, challenging, telling information, recommending), and content codes (e.g., prognostic information, diagnostic information, code status, goals of care).

Content codes were defined to capture aspects of the clinical interaction that would aid or thwart attempts to make practically wise decisions in this scenario (e.g., avoid intubation by respecting the patient’s preferences, goals and values). Two concepts deserve elaboration. First, “Healing” (K84, Table 1) was defined as any evidence that the agents were employing a definition of healing that was broader than “fixing” a merely physiological problem. It was intended to capture attempts to meet patients on a human level rather than merely providing technical expertise or expert advice. As such, the term might also be called “polite” or “compassionate” or “humane”. Examples include asking if there is anything that would make them more comfortable, or offering to contact a chaplain or family member to assist with bereavement, or asking what the clinician could do to help make the time remaining as rich and meaningful as possible. Also deserving elaboration is the concept of “comfort” (K22, K59) that entails not only a concern for patient comfort, but also an explicit preference for eschewing curative treatments. As such it is often used as an alternative (and even antonym) to typical, high-intensity treatments that may be conducted with attention to patient comfort, but are primarily aimed at curing the underlying disease.

Table 1:

Absolute and mean counts of codes, combination codes and supercodes across all simulations.

Codes†		Total Code Count across all simulations	Mean Code Count per Simulation	Standard Deviation per Simulation	Minimum Code Count per Simulation	Maximum Code Count per Simulation	Number of Simulations with ≥1 occurence
	Action Codes
K2	Asking follow-up question	373	4.2	3.9	0	16	77 (88%)
K3	Asking question	2311	26.3	10.8	3	62	88 (100%)
K5	Challenging	39	0.4	1.1	0	6	20 (23%)
K9	Justifying	227	2.6	3.4	0	20	57 (65%)
K12	Recommending	687	7.8	6.1	0	30	86 (98%)
K13	Telling of information	5557	63.1	26.8	8	127	88 (100%)
	Agent Codes
K16	Patient	988	11.2	6.7	1	33	88 (100%)
K17	Physician	9768	111.0	38.8	24	218	88 (100%)
K18	Surrogate	2902	33.0	17.6	3	83	88 (100%)
	Content Codes
	Code Status Preferences
K20	Patient’s preferences	943	10.7	7.6	0	31	79 (90%)
K21	Physician’s preferences	311	3.5	3.1	0	11	74 (84%)
	Comfort
K22	Comfort as the goal of care	488	5.5	5.4	0	25	68 (77%)
K23	Comfort as some other meaning	91	1.0	1.7	0	7	38 (43%)
	Discussion of Death
K40	Direct discussion of death	47	0.5	1.0	0	5	25 (28%)
K41	Oblique discussion of death	209	2.4	3.1	0	12	49 (56%)
	Future Therapeutic Preferences
K42	Patient’s Therapeutic Preferences	235	2.7	3.6	0	20	58 (66%)
K43	Physician’s Therapeutic Preferences	641	7.3	6.0	0	28	80 (91%)
	Goals of Care
K56	Patient’s Goals	383	4.4	6.4	0	49	67 (76%)
K57	Physician’s Goals	433	4.9	4.4	0	26	79 (90%)
K58	Attempt cure within limits	177	2.0	2.9	0	12	45 (51%)
K59	Comfort/Non-Curative/Allow-to-die	248	2.8	4.1	0	25	55 (63%)
K60	Attempt cure at all costs	28	0.3	0.8	0	4	16 (18%)
K61	No goal beyond diagnosis of problem	36	0.4	1.1	0	7	18 (20%)
	Information about:
K70	Prognosis	493	5.6	5.2	0	24	76 (86%)
	Feeling about Uncertainty
K77	Calm/comfortable with uncertainty	107	1.2	1.9	0	9	41 (47%)
K78	Anxious/uncomfortable with uncertainty	17	0.2	0.7	0	5	9 (10%)
	Values
K79	Patient’s values	75	0.9	1.3	0	8	42 (48%)
K80	Physician’s values	60	0.7	1.5	0	7	23 (26%)
	Single/Uncategorized Codes
K84	Healing	402	4.6	5.0	0	20	65 (74%)
	Combination Codes ‡
K24	Comfort & Direct Death	8	0.1	0.4	0	3	5 (6%)
	[K22 & K40]
K25	Comfort & Non-Curative	170	1.9	3.0	0	18	46 (52%)
	[K22 & K59]
K26	Comfort & Oblique Death	52	0.6	1.2	0	7	23 (26%)
	[K22 & K41]
K27	Comfort/Non & Oblique Death	21	0.2	0.6	0	3	16 (18%)
	[K22 & K59 & K41]
K48	Patient’s Goals: Comfort/Non	84	1.0	2.1	0	11	30 (34%)
	[K56 & K59]
K50	Physician’s Goal: Direct Death	6	0.1	0.4	0	3	4 (5%)
	[K57 & K40]
K51	Physician’s Goal: Limited	114	1.3	2.0	0	11	40 (67%)
	[K57 & K58]
K52	Physician’s Goal: Comfort/Non	147	1.7	2.2	0	10	48 (55%)
	[K57 & K59]
K53	Physician’s Goal: Oblique Death	32	0.4	0.9	0	4	16 (18%)
	[K57 & K41]
K50_53	Physician’s Goal: Oblique or Direct	38	0.4	0.9	0	4	20 (23%)
	[K57 & (K40 or K41)]
K54	Physician’s Goal: Unlimited	18	0.2	0.6	0	3	13 (15%)
	[K57 & K60]
	Supercodes ‡
SC9	Physician names death	219	2.5	2.8	0	12	60 (68%)
	[K17&(K40|K41)]
SC12	Physician recommends preferences, values or goals	339	3.9	4.0	0	23	78 (89%)
	K17&K12&(K20|K21|K56|K57|K79|K80)
SC13	Physician tells patient preferences	412	4.7	3.9	0	15	71 (81%)
	[K17&K13& (K20|K42|K56|K79)]
SC14	Physician asks about healing	142	1.6	2.3	0	13	47 (53%)
	[K17&K3&K84]
SC16	Physician justifying own preferences	91	1.0	1.8	0	9	34 (39%)
	[K17&K9&(K21|K57|K80)]
SC17	Physician recommends healing	40	0.5	1.0	0	5	20 (23%)
	[K17&K12&K84]
SC19	Physician comfortable with uncertainty	107	1.2	1.9	0	9	41 (47%)
	[K17&K77]
SC23	Patient/Surrogate telling preferences, values or goals	443	5.0	4.5	0	30	77 (89%)
	[(K16|K18)&K13&(K20|K42|K56|K79)]
SC27	Physician soliciting preferences, values or goals	346	3.9	3.9	0	22	79 (90%)
	[K17&(K2|K3)&(K20|K42|K56|K79)]
SC28	Physician testing preferences, values or goals	549	6.2	5.1	0	32	49 (56%)
	[K17&(K2|K5)&(K20|K42|K56|K79)]
SC30	Physician tells death prognosis	405	4.6	4.2	0	19	75 (85%)
	[K17&K13&(K40|K70)]
SC31	Physician tells/recommends comfort as the goal of care	355	4.0	4.3	0	23	67 (76%)
	[K17&(K12|K13)&K22]

^‡In combination codes and supercodes “&” denotes the Boolean operator “AND” and “|” denotes the logical operator “OR”.

^†This table includes only those codes used in this analysis. This includes all codes, combination codes and supercodes with significant univariate associations with outcomes of interest (Table 2) along with the underlying codes used to build the significant combination codes and supercodes. Codes K28–29, K44–49, K55, K62–67, K75–76, K81, K85, K88, AND K91 were omitted because they were never/rarely applied to text in this dataset. Codes K1, K4, K6–7, K10–11, K14–15, K19, K30–39, K68–69, K71–74, K82–83, K85–95 were part of the codebook and were applied to the text, but did not factor in this analysis because they did not demonstrate univariate associations with the outcomes of interest.

Qualitative Coding

Primary coding was completed by trained analysts contracted through the University of Pittsburgh Center for Social and Urban Research. Codes were applied at the level of the sentence. Each sentence was assigned one agent code, one action code, and as many content codes as applicable to the content of the sentence. Coding was done in pairs with adjudication by the lead coder when necessary. Coding was completed using Atlas.ti (Berlin, Scientific Software Development).

Once primary coding was complete, the authors (DH, KS & JF) used Atlas.ti to build a limited set of “combination codes” and “supercodes”. Combination codes denote the co-occurrence of two or more content codes in the same sentence of coded text. For example, because we thought practical wisdom would require framing goals of care or the discussion of death within a wider context of “comfort care”, we created 4 unique content codes by combining the “comfort” code with codes for the goals of care or the referencing of death (Table 1, K24–27). Supercodes were comprised of specific agent, action and content codes that correspond to behaviors relevant to the exercise of practical wisdom. For example, “Physician names death” is a supercode consisting of the “physician” agent code, the “telling of information” action code, and either of the two content codes pertaining to a “direct” or “oblique” reference to death. An initial list of 31 proposed supercodes was then edited by DH to eliminate redundant or rare codes (e.g., codes applied less than 100 times), yielding a final set of 12 supercodes (Table 1, SC 9–31).

Quantitative Analysis of Qualitative Coding

Quantitative analysis of qualitative coding began by exporting from Atlas.ti the final, adjudicated database in a format suitable for statistical analysis. For each simulation, we tabulated the total number of quotations containing each code or supercode. We also created a binary indicator variable for each code and supercode to indicate if it was ever applied to a sentence from each simulation. For example, if a simulation never mentioned “comfort as a goal of care (K22)”, this indicator variable was assigned the value “0”, and regardless of the number of times the physician mentioned “comfort as a goal of care”, the indicator variable was assigned the value “1”. We then constructed a series of logistic regression models to explore the association between these predictor variables and the desired outcome: no intubation (DNI), no attempted cardiopulmonary resuscitation at the time of arrest (DNR), and a plan of care focused on managing the patient’s symptoms and keeping him comfortable (CMO). AEB had previously coded each simulation according to the “desired,” “preference concordant” outcomes (Barnato et al. 2014, Haliko et al. 2018), and we imported these outcomes assessments into our analytic dataset. Using the simulated encounter as the unit of analysis, we first built univariate models using each code to predict a specific outcome, controlling for physician type (e.g., internist, intensivist or emergency medicine physician). We then built multiple regression models retaining codes/supercodes associated with the outcomes of interest at p≤0.1, controlling for physician type. All statistical analyses were completed with SAS (Cary, North Carolina).

Thematic Analysis of Quantitative Findings

In order to further understand and characterize the findings of the quantitative analysis, we undertook a thematic analysis of the behaviors identified through the quantitative analysis as predictive of preference-concordant decisions. The goal of this thematic analysis was to further explore if and how these behaviors demonstrated practical wisdom. Working in Atlas.ti, the authors (AM & JF) applied a deductive, latent approach consistent with the method described by Braun, Clarke and Terry (Braun, Clarke, and Terry 2014) to understand the use and meaning of the identified codes. After reading the transcripts for familiarization, AM then reread the text tagged with the specific codes and supercodes associated with preference-concordant decisions to identify themes that were illustrative of practical wisdom. They refined these themes and rechecked them against the coded transcripts to ensure that they accurately encompassed the body of data. The coded quotations were read both individually and in their embedded context within the simulation so that understanding of setting of use could also inform the analysis.

Results

Codebook and Coding

The final dataset included transcripts from 88 simulations comprised of over 19,000 individual sentences, each coded according to the scheme described above. The final codebook included 6 agent codes, 13 action codes, 61 content codes, 16 combination codes and 31 supercodes. Because some codes were never (or only rarely) applied to the text, we excluded them from the analysis. Other codes did not factor in this analysis because they did not demonstrate univariate associations with the outcomes of interest, and are thus not described in detail here. Table 1 includes all codes used in this analysis with descriptive statistics of their application to the dataset. The full codebook is available from DH on request.

Quantitative Analysis

Across all 88 simulations, 66 were made DNR/DNI, 46 were made CMO, 39 were transferred to the ICU and 12 were intubated. Significant univariate associations between codes and outcomes of interest are summarized in Table 2. Univariate associations were observed for only combination codes and supercodes (e.g., no associations were observed for individual agent, action, or content codes). As expected, preference-concordant outcomes were more likely when the patient or surrogate told their preferences, values or goals (SC23), or when the physician solicited (SC27) or tested (SC28) those preferences. Outcomes of interest were more likely when the physician’s goal of care was comfort-oriented (K52), as well as when the physician named death (SC9), told something about patient preferences (SC13), or asked about (SC14) or recommended (SC17) “healing” (e.g., non-technical opportunities that could achieve some sense of healing or compassion, despite the setting of impending death). Other positive associations were found when physicians demonstrated comfort with uncertainty (SC19), told death as a prognosis (SC30), or recommended “comfort” as the goal of care (SC31).

Table 2:

Predictors of preference concordant outcomes in univariate models.

Code			Outcome
			CMO	DNR/DNI	ICU	Intubate
K52†	Physician’s Goal: Comfort/Non	Count	+	+	+	+
	[K57 & K59]	Binary	+	+	+
SC9†	Physician names death	Count		+	+	+
	[K17&(K40|K41)]	Binary		+	+	+
SC13†	Physician tells patient preferences	Count	+	+	+	+
	[K17&K13& (K20|K42|K56|K79)]	Binary	+	+	+	+
SC14†	Physician asks about healing	Count		+	+
	[K17&K3&K84]	Binary	+	+	+
SC17†	Physician recommends healing	Count	+		+
	[K17&K12&K84]	Binary	+		+
SC19‡	Physician comfortable with uncertainty	Count
	[K17&K77]	Binary		+
SC23†	Patient/Surrogate telling preferences, values or goals	Count	+	+		+
	[(K16|K18)&K13&(K20|K42|K56|K79)]	Binary			+	+
SC27†	Physician soliciting preferences, values or goals	Count	+	+		+
	[K17&(K2|K3)&(K20|K42|K56|K79)]	Binary			+	+
SC28‡	Physician testing preferences, values or goals	Count
	[K17&(K2|K5)&(K20|K42|K56|K79)]	Binary		+		+
SC30†	Physician tells death prognosis	Count		+		+
	[K17&K13&(K40|K70)]	Binary		+	+	+
SC31†	Physician tells/recommends comfort as the goal of care	Count	+	+	+	+
	[K17&(K12|K13)&K22]	Binary		+	+	+

+ indicates p<.05 for supercode predicting outcome, controlling for physician type.

^†Indicates codes used in multivariate modeling of both count and binary data.

^‡Indicates codes used in multivariate modeling of only the binary data.

Abbreviations: Comfort Measures Only (CMO); Do not resuscitate/Do not intubate (DNR/DNI); Intensive Care Unit (ICU).

Logistic regression models using binary indicator variables indicating the presence/absence of the code of interest are reported as “Binary” whereas models using integer counts of the occurrence of each code are reported as “Count”.

For the multivariable models, our primary analysis focused on the binary indicator variables indicating whether or not a particular behavior ever occurred during the simulation. Models were built in 3 steps. The first step included all codes identified by univariate analyses as listed in Table 2, and as expected, the primary outcome of intubation was much less likely to occur when the patient or surrogate told their preferences, values or goals (SC23) [OR 0.006, 95% CI <0.001–0.066, Table 3]. However, because this finding essentially demonstrates that physicians listened to patients when they expressed preferences against intubation, we eliminated SC23 in the second step of the analysis to focus on the behaviors that might effectively elicit patient preferences. In this second model, physician solicitation of preferences, values and goals (SC27) emerged as the most significant predictor of intubation (OR 0.038, 95% CI 0.003–0.487), but this finding essentially demonstrates that the patient-actors were well-trained to articulate their preference against intubation if requested. We therefore eliminated SC27 from the final model (Step 3) to explore not only if, but how the physicians solicited preferences, values, and goals.

Table 3:

Predictors of preference concordant outcomes that achieved significance in multivariable models controlling for physician type.

Model	Predictor	Primary Outcome			Secondary Outcomes
		Intubate			ICU		DNR/DNI		CMO
Step	Code	OR	95% CI	c-stat	OR	95% CI	OR	95% CI	OR	95% CI
1				0.904
	SC23	0.006	<0.001, 0.066		ns		ns		ns
2				0.903
	SC27	0.038	0.003, 0.487		ns		ns		ns
	SC13	0.092	0.014, 0.595		ns		ns		ns
3				0.871
	SC13	0.036	0.007, 0.177		ns		13.98	2.892, 67.569	ns
	SC14	ns			0.245	0.082, 0.737	11.647	2.232, 60.787	ns
	SC30	ns			0.069	0.006, 0.756	ns		ns
	SC31	ns			0.128	0.024, 0.683	ns		ns

Note: The Step 1 model includes SC9, SC13, SC14, SC17, SC19, SC23, SC27, SC28, SC30, SC31, and K52 with Physician Type as a covariate. The step 2 model is identical except that SC23 is eliminated. Step 3 eliminates both SC23 and SC27. Predictor variables are modeled as binary indicator variables indicating whether or not a particular code was ever applied to text from a specific simulation. Coefficients are presented for only those variables achieving statsistical significance; model coefficients for all variables are available on request to the senior author.

Abbreviations: Odds Ratio (OR); Confidence Interval (CI); c-statistic (c-stat); Intensive Care Unit (ICU); Do not resuscitate/Do not intubate (DNR/DNI); Do Comfort Measures Only (CMO); not statistically significant (ns).

In this final model, we found that the single predictor of intubation was whether or not physicians told something about patient preferences (SC13) during the simulation (OR 0.036, 95% CI 0.007–0.177). This same behavior is also associated with a substantial increase in the odds of making the patient DNR/DNI (OR 13.980, 95% CI 2.892–67.569). Although no other codes were significantly associated with intubation, the odds of transfer to the ICU were reduced when physicians asked about “healing” (SC14), told death as a prognosis (SC30) or recommended “comfort” as the goal of care (SC31) (Table 3).

To test the robustness of these findings in a sensitivity analysis, we constructed similar, 3-step models using the code counts rather than the binary indicator variables (Table 4). We again found that the strongest predictor of intubation was the patient or surrogate telling their preferences, values or goals (SC23), but when this code was eliminated in Step 2, instead of physician solicitation of patient preferences (SC27), the significant predictors of intubation were telling death as a prognosis (SC30) and recommending “comfort” as the goal of care (SC31). In the final (Step 3) sensitivity analysis, recommending “comfort” as the goal of care (SC31) was associated with lower odds of intubation and ICU transfer and increased odds of being made CMO (Table 4). In addition, physician inquiry about “healing” (SC14) or recommendation of “healing” (SC17) were associated with lower odds of ICU admission and increased odds of being made DNR/DNI or CMO (Table 4).

Table 4:

Predictors of preference concordant outcomes that achieved significance in multivariable models controlling for physician type (Sensitivity Analysis).

Model	Predictor	Primary Outcome			Secondary Outcomes
		Intubate			ICU		DNR/DNI		CMO
Step	Code	OR	95% CI	c-stat	OR	95% CI	OR	95% CI	OR	95% CI
1				0.976
	SC23	0.253	0.096, 0.668		0.816	0.689, 0.967	1.764	1.269, 2.451	ns
	SC30	0.327	0.103, 1.042		ns		ns		ns
2				0.975
	SC27	ns			ns		1.408	1.134, 1.749	ns
	SC30	0.358	0.128, 0.999		ns		ns		ns
	SC31	0.201	0.054, 0.748		0.8	0.643, 0.994	ns		1.232	1.031, 1.473
3				0.975
	SC13	ns			ns		1.425	1.129, 1.798	ns
	SC14	ns			0.421	0.256, 0.694	3.043	1.300, 7.126	ns
	SC17	ns			ns		ns		3.75	1.216, 11.564
	SC30	0.358	0.128, 0.999		ns		ns		ns
	SC31	0.201	0.054, 0.748		0.8	0.643, 0.994	ns		1.232	1.031, 1.473

Note: The Step 1 model includes SC9, SC13, SC14, SC17, SC23, SC27, SC30, SC31, and K52 with Physician Type as a covariate. The step 2 model is identical except that SC23 is eliminated. Step 3 eliminates both SC23 and SC27. Predictor variables are modeled as integer counts indicating the total number of times a particular code was applied to text from a specific simulation. Because SC19 and SC28 were not significant univariate predictors of any of these outcomes, they were omitted from these models. Coefficients are presented for only those variables achieving statsistical significance; model coefficients for all variables are available on request to the senior author.

Abbreviations: Odds Ratio (OR); Confidence Interval (CI); c-statistic (c-stat); Intensive Care Unit (ICU); Do not resuscitate/Do not intubate (DNR/DNI); Do Comfort Measures Only (CMO); not statistically significant (ns).

Qualitative Analysis

The primary finding of the quantitative analysis was that intubation was much less likely when physicians told something about patient preferences (SC13). Secondary findings suggest that other preference-concordant outcomes are associated with physicians telling death as a prognosis (SC30) or recommending “comfort” as the goal of care (SC31). However, the statistical analysis is not by itself sufficient to characterize whether these behaviors were prudent or paternalistic because we were unable to statistically establish the chronological relationship between the elicitation of patient preferences, values, and goals (SC27) and the physicians’ telling (SC13, SC30) or recommending (SC31). If the physician told or recommended before first soliciting patient values, the interaction could conceivably reflect paternalism (Ubel, Scherr, and Fagerlin 2017), whereas telling or recommending after values elicitation would likely be better aligned with those values and thus more prudent. We therefore focused our qualitative analysis on the coded behaviors associated with the desired outcomes, paying particular attention to this chronological relationship. Figure 1 demonstrates the proposed model of prudent goal of care conversations as abstracted through thematic analysis of the transcripts.

Figure 1: Conversational paradigm for prudent discussion of noncurative goals of care.

Desires: SC23; Distillation: SC13; Assimilation: K52; Confirmation: SC31; Affirmation: K1+K17

Typically, the goals of care conversation began when the physician inquired about the family’s wishes regarding intubation or aggressive management, and the patient or surrogate responded with “no tube” or “I want to keep him comfortable” (K22: Comfort as the Goal of Care). These statements did not in themselves commit the patient or physician to a comfort-oriented plan of care (K59: Goal of care: Comfort/Noncurative/Allow-to-die), but did offer physicians the chance to discuss treatment goals in greater detail. Many physicians took the opportunity to synthesize expressed desires and to explicitly reframe the family’s goals as focused primarily on the patient’s comfort (SC13).

The next step in the paradigm was for the physician to state his or her own comfort-oriented goals of care (K52: Physician’s Goal: Comfort/Non). Nearly all of the physicians who explicitly named their own goals (K52) mentioned the patient’s expressed desires in shaping their treatment decision, but most of the physicians who had distilled patient desires into patient goals (SC13) also made statements to show that their physician-goals were in alignment. After naming their goals, some physicians proceeded to confirm their treatment goals with the family (SC31: Physician tells/recommends comfort as the goal of care); thereby signaling that they had processed the family’s goals and were setting up a plan of action in line with those goals. Finally, a small subset of physicians chose to affirm the family in their decision (K1&K17) by making statements such as “I would do the same” or “I understand why you are making this decision”, statements which normalized the family’s goals and again decreased the relational distance between the physician and the family. All of these findings suggest that the physician telling and recommending was, in fact, prudent.

It was harder to determine the prudence of simulations in which both parties did not state their comfort-oriented goal of care (e.g., K52 or K48 was absent). In interactions that contained the code for physicians stating comfort-oriented goals (K52) but not patients stating those goals (K48), the physician typically elicited the family’s desires for comfort (K22) and referred to those desires when naming their own goals for treatment, but did not create an opportunity to confirm the family’s noncurative treatment goal (K48) and demonstrate that these goals aligned. In these cases, the physicians effectively discerned the right outcome of avoiding resuscitation. However, they failed to show that they were reaching this right end through the means consistent with respect for persons at the end of life. Conversely, the few cases in which interactions included patient noncurative goals (K48) but not physician goals (K52) were characterized by the physician summarizing patient wishes, relating them to a comfort-oriented goal of care that must have generated these wishes, and then treating accordingly without advancing the physician’s goals and showing that they align.

Discussion

This mixed method analysis of simulated end of life decision-making has three primary findings. First, the preference concordant outcome of comfort-directed palliative care that avoids intubation is most often achieved when the patient or surrogate actually express their preferences, values and goals (SC23). Second, patients’ preferences, values, and goals are more often expressed when they are explicitly solicited by physicians (SC27). Third, once physicians have solicited and listened to patients’ values and goals, the preferred outcomes are associated with physicians’ (1) verifying their understanding of the patient’s wishes by “telling” or “repeating back” what they have heard the patient say (Dias et al. 2003, D’Agostino et al. 2017) (SC13), and then (2) making a recommendation that aligns the care plan with the patient’s preferences for comfort or a more holistic concept of healing (SC31,17). Although the first two findings are unsurprising in the setting of a simulation designed around eliciting and enacting patient preferences, doing so plainly is necessary to achieving the intended good of value-concordant end of life care. They also serve as important tests of simulation fidelity that internally validate the analytic approach. The third finding goes further by specifying two techniques that may help physicians elicit patients’ preferences and fashion a plan of care consistent with those preferences.

Because (as this study supports) physicians’ elicitation of patient preferences is basic to realizing those preferences, physicians must be proficient in communicating with patients and their surrogates. A review of the literature reveals several barriers to effective communication at the end of life. Sometimes patients’ incapacity precludes meaningful communication between physician and patient, and past documentation of goals-of-care conversations may not be adequate to guide physicians (Fassier et al. 2016). Even when the patient is alert and oriented, effective communication may be undermined by barriers such as patient anxiety or denial, challenges inherent in prognostication, and physicians’ discomfort or lack of time (Bernacki, Block, and American College of Physicians High Value Care Task 2014). In this study we found that physician comfort with uncertainty (SC19) was associated with more favorable outcomes, and although we did test for physician’s perception of time, this was not found to be a significant predictor of outcomes (data not shown). These findings corroborate those of one qualitative British study that explored physician’s reactions to a legal requirement to discuss resuscitation decisions with their patients which highlighted physicians’ discomfort with incorporating patients’ families in the medical decision-making process—especially when the family members may not have the cognitive ability or emotional distance to develop an adequate understanding of the relevant medical complexities (MacCormick et al. 2018). Another French ICU study of end-of-life decision-making found that the majority of decisions were initiated by medical personnel (Quenot et al. 2012), further underscoring the importance of practitioner initiative in framing these conversations.

By reinforcing the connection between prudent behaviors and right outcomes, this study contributes to the evidence base for training clinicians with innovative and scalable resources like the Ariadne Serious Illness Conversation Guide (2017). Several core elements of this guide run parallel with this study’s key findings. After clinicians open a space for the conversation, the Ariadne guide prompts them to assess understanding and preferences (SC27), share the prognosis (SC30) and explore key topics including sources of strength, which is consistent with physicians asking about healing (SC14). In closing the conversation, clinicians are prompted to summarize (SC13) and make recommendation (SC17 and SC31). Additionally, the findings reported here emphasize the need to formulate a recommendation that is compatible with the patient’s goals including personalized conceptions of comfort and healing.

Critical care consensus statements incorporate the concept of palliative care for the dying into their recommendations, reflecting a recognition that intensive care is not always centered on cure (Myburgh et al. 2016). Most of these consensus statements recommend that critical care physicians be facile in providing comfort for their patients, making palliative consults or arrangements when feasible, and endorsing comfort-oriented treatment as an important aspect of treating the dying. Similarly, an American College of Emergency Physicians ethics statement affirms the practices of correctly identifying patient wishes, supporting family presence at end of life, and developing the ability to communicate sensitive information (2016). A Society of Hospital Medicine practice guideline encourages every hospitalist to have a grasp of “primary palliative care” to better serve hospitalized patients (Anderson et al. 2017). These recommendations reflect a respect for end of life care across medical organizations and a call for physicians to develop the skills necessary to effectively assist patients at the end of life.

There is opportunity for physicians to more frequently practice the behaviors that this study suggests are effective in realizing patient preferences at the end of life: (1) verifying understanding of patient wishes by repeating back what they have heard the patient say and (2) making a recommendation consistent with the patient’s goals. Simulations, such as the one that forms the basis of this study, have been recommended as a means by which physicians can learn to practice new skills, including elicitation of patient preferences at the end of life specifically (Alexander et al. 2006, Back et al. 2007, Yedidia et al. 2003). However, implementation of intensive training by simulations is both time-consuming and expensive (Tulsky 2005), and thus alternate strategies have been developed to improve physician communication. For example, written patient-care vignettes have been shown to be adequate to evaluate the quality of physicians’ processes of providing care (Peabody et al. 2000, Peabody et al. 2004), though they are comparatively inferior to simulation for purposes of assessing the quality of physicians’ processes of caring for patients. Utilization of vignettes could afford healthcare organizations the opportunity to assess with some reliability whether physicians are engaged in the practices this study suggests as more likely to result in value-concordant end-of-life care. In addition, physicians responsible for the formal and informal training of junior colleagues could readily model the specific practices this study associates with value-concordant end-of-life care. While physician role modeling is a complex phenomenon, it is widely accepted that it significantly contributes to the formation of newer physicians, for both good and ill, and that physician role-modeling is informed by conscious effort of the model physicians (Kenny, Mann, and MacLeod 2003, Passi and Johnson 2016). Retrospective, structured feedback on actual clinical experiences communicating with patients improves clinician skills in so doing (Maguire and Pitceathly 2002), albeit in a higher-stakes setting than would be the case in a simulation. This study’s findings have a key advantage in replication through formal and informal education: regardless of which educational strategy is most effective for communication skills in general, these associated behaviors are straightforward and concrete enough that promotion through simulations, role modeling, or retrospective review of physician—patient interactions could all serve to increase the likelihood that physicians involved in goals-of-care discussions of (1) verify patients’ understanding by repeating what the patient has said and (2) make recommendations that are explicitly connected to, and consistent with, patients’ goals.

This study has several limitations. First, although standardized patients have been validated for assessing the quality of physicians’ care of actual patients (Peabody et al. 2000, Peabody et al. 2004), it is impossible to wholly exclude the possibility that physicians behave somewhat differently during simulations than they would have with actual patients. Second, because the simulated patient and his surrogate were instructed to offer the patient’s comparatively clear and settled preferences upon solicitation, this study’s recommended practices will not enjoy the same degree of reliability with actual patients and their families, who may be more hesitant to express themselves or uncertain or conflicted about their goals. Third, the threshold for statistical significance was not adjusted for multiple tests, and the relatively small sample size may have resulted in some overfitting bias, especially in models for rarer outcomes. However, the purpose of the analyses was to guide the qualitative findings and generate (rather than formally test) hypotheses. Fourth, the identification of behaviors consistent with prudence does not necessarily support the attribution of prudence per se because the latter requires much more exposure to an individual’s behaviors across a variety of care contexts. Attributions of prudence must also account for the plurality of moral frameworks surrounding end-of-life care including vitalistic and paternalistic approaches that would justify different conceptions of prudence.

Conclusion

This mixed-methods study offers empirical support for the common-sense idea that patients are more likely to receive end-of-life care consistent with their preferences when their physicians explicitly solicit patients’ preferences. More concretely, this study also supports the physician practices of verifying physicians’ understanding of patient wishes by repeating back what they have heard the patient say and making an explicit recommendation consistent with the patient’s goals, finding that these behaviors make preference concordant treatment more likely at the end of life. The patient-centered communication promoted by these practices is widely endorsed by physicians’ professional societies, and these practices could be promoted through simulations, vignettes, role modeling, and retrospective review of actual patient care encounters. Implementing these practices more widely should result in more patients receiving care consistent with their goals at the end of life.