Hippocrates in the ICU: Predicting the End of Life in Critical Care
Special Feature
Hippocrates in the ICU: Predicting the End of Life in Critical Care
By Gordon D. Rubenfeld, MD
It is in my opinion a most excellent thing for the physician to practice forecasting. [. . .] For it is impossible to make all sick people well—that would indeed be better than foretelling the course of their illness. But men do, as a matter of fact, die; some through the severity of their disease, before the doctor is called in—some immediately, some living on for a day, others a little longer. [. . .] Hence we must know the nature of these diseases, how far they are superior to the bodily powers. In this way one will justly gain a reputation and will be a good physician; one will indeed be better able to save those capable of cure if he has made up his mind long beforehand in each case, and no blame will attach to him if he has already forseen and announced who is to die and who is to be preserved. —Hippocrates, Prognostica, 5th century BC
Interest in predicting outcome for critically ill patients dates back to the beginning of critical care itself. In 1966 the prognostic significance of coagulation abnormalities in patients with shock was noted.1 In 1971, an index was published in the New England Journal of Medicine to predict the outcome of patients admitted to ICUs who had taken an overdose of sedatives.2 As Hippocrates noted, separating those who are doomed to die from those with a chance at benefiting from therapy allows clinicians to devote resources to salvageable patients and inform decisions about care for all patients. Many patients with severe underlying diseases spend time in the ICU prior to their death. The SUPPORT study followed patients with acute respiratory failure, multiple organ system failure with sepsis, multiple organ system failure with malignancy, coma, chronic obstructive lung disease, congestive heart failure, cirrhosis, metastatic colon cancer, and non-small-cell lung cancer. Prior to death, 38% of these patients spent at least 10 days in an ICU and 46% received mechanical ventilation within three days of death.3 In Washington state, at least one-third of all patients who die in a hospital receive some care in an ICU during the hospitalization in which they die. Currently, the majority of patients who die in ICUs do so after a decision has been made to limit life-sustaining therapies.4 Therefore, the decision to limit life-sustaining therapy in the ICU is a common and important one. An understanding of the outcomes of intensive care should play a significant role in this decision.
Subjective vs. Objective Prognosis
Most clinicians use no formal risk assessment when they make prognoses for critically ill patients. It is the rare clinician indeed who sits down to calculate an APACHE (Acute Physiology and Chronic Health Evaluation) or MPM (Mortality Prediction Model) score before discussing end-of-life issues with a family. Should clinicians routinely use some objective method to estimate the probability of survival? Does this improve care? Does it improve discussions with families? Before advocating the widespread use of objective probability estimates, we should consider what might be wrong with current practice. Most clinicians base ICU prognoses on personal experience, consultation with colleagues, and informal assessment of the literature. Unfortunately, there is a large body of literature suggesting that these subjective probability assessments by clinicians are fraught with problems.5 Cook and colleagues surveyed critical care clinicians with vignettes of cases and provided them with five management choices: discontinue inotropes and mechanical ventilation but continue comfort measures; discontinue inotropes and other maintenance therapy but continue mechanical ventilation and comfort measures; continue with current management but add no new therapeutic intervention; continue with current management, add further inotropes, change antibiotics, and the like as needed, but do not start dialysis; or continue with full aggressive management and plan for dialysis if necessary.6 In many of the scenarios, equal numbers of responding clinicians selected choices at opposite ends of the therapy spectrum. It is concerning to think that the same patient would receive full aggressive support from one physician while another would provide comfort measures only. What is particularly interesting is that this variability occurred even though the clinicians were provided with an APACHE prediction of survival in the vignette.
Cognitive psychologists have studied the way humans make predictions for years. Physicians and nurses, like weathermen, judges, and plumbers, are subject to a variety of psychological biases in the way they make predictions.7 A number of these biases are well studied and have specific names. (See Table.) Generally, clinicians tend to be overconfident in their personal prognoses, are unduly influenced by isolated cases, and are reluctant to apply population data to individuals.8 There are good reasons to suspect that ICU clinicians will apply subjective prognostic information inconsistently and with a variety of unconscious biases. It is possible that having objective outcome data may improve the consistency and quality of these decisions.
| Table | |
| Biases in Subjective Prediction | |
| Overconfidence | Tendency to be overconfident in the accuracy of one’s own predictions |
| Availability | Tendency to exaggerate the frequency of recent or easily remembered cases |
| Individuals vs. populations | Reluctance to apply group statistics to individual cases |
| ______________________________________________________________________ | |
Mathematical Models and Assessing the Risk of Death
What kind of information would clinicians find useful in making decisions about limiting life support in the ICU? Their options in the literature take two forms. One is the continuous probability prognosis generated by a mathematical model. These models are sometimes referred to as "severity of illness models," but they generate probabilities of death. Perhaps the most well known of these is the APACHE (Acute Physiology and Chronic Health Evaluation) score, but there is also the MPM (Mortality Prediction Model), SAPS (Simplified Acute Physiology Score), LODS (Logistic Organ Dysfunction Score), and the SUPPORT model. These models can be quite complex; for example, the SUPPORT model has more than 20 variables built into it and generates prognoses on days 1, 3, 7, 14, and 25 after admission to the hospital.
Complicating the interpretation of these predictions is their applicability to your patient population. If, for example, your hospital provides either better or worse care than the patient received in the data set that generated the model, then the patient’s actual prognosis may be better or worse than the prediction. All studies of outcome of intensive care are based on the outcomes of patients who have been selected for admission to the ICU. A recent study that noted unusually high survival after cardiopulmonary resuscitation hypothesized that their results were due to the selection of patients who received CPR.9 Another hospital that used CPR less selectively could not make similar prognoses for its patients. Similarly, prognostic data from an ICU with selective admission criteria may be more optimistic than outcomes in an ICU that frequently admits moribund patients.
The primary use for these prognostic models is to evaluate populations of patients for research and quality assurance purposes. These models are designed to make accurate predictions across the entire range of probabilities. For example, the models are designed to be equally accurate for patients with a 50% chance of dying as they are for patients with a 90% chance of dying. However, accuracy in the middle and low ranges of prognosis are unlikely to affect a decision to continue intensive care, particularly after a decision has been made to initiate it.
Clinical Prediction Rules
Clinicians have a lot of experience using clinical prediction rules and categorical prognostic information in clinical medicine, even though they may not realize it. Cancer staging is a useful example. Based on a combination of a few variables, usually the type and size of the tumor, patients with cancer are placed into one of several stages with distinct prognoses. In critical care there are a number of examples of clinical prediction rules. The clinical criteria for brain death are accepted to predict an irreversible coma.10 Other clinical prediction rules have been developed in critical care. No mechanically ventilated bone marrow transplant recipients who had a combination of lung injury and either hepatic and renal failure or required vasopressors survived out of an estimated 398 patients. Only 2% of patients with respiratory failure, cirrhosis, and a creatinine more than 1.5 mg/dL survive.11 Prognostic information about CPR is well known. Few patients with metastatic cancer, who receive CPR in a nursing home, or who require ongoing CPR after initial attempts in the field fail have been reported to hospital discharge.12 No patients who were comatose after a cardiac arrest with a cerebrospinal fluid level of creatine kinase isoenzyme BB greater than 204 U/L regained consciousness.13 Of course, clinical prediction rules must be evaluated critically for their validity.14
Outcomes Other Than Mortality
Predicting hospital mortality is not the only outcome of interest. For many patients, survival in a severely impaired state may be a fate worse than death. Unfortunately, data on the quality of life and functional status after intensive care are only slowly becoming available. One simple observation that can be made in the absence of data is that, unless admission to the ICU follows a procedure intended to improve the quality of life (for example, coronary bypass surgery or aggressive cancer therapy), patients who survive critical illness will not have an improved quality of life compared to their pre-ICU status. Therefore, if a patient is debilitated with an unacceptable quality of life before entering the ICU, that patient will certainly be no better after admission for septic shock and pneumonia. Declines in quality of life after critical illness, particularly the effects of septic encephalopathy, prolonged sedation, lung injury, and neurologic catastrophes, are only now being studied.
The Challenge of the SUPPORT Study
Information about outcomes of care does affect patients’ decisions. How the information is presented is essential. Patients will make different choices about therapy, depending on whether the outcomes are presented as a 20% chance of survival or as an 80% chance of dying.15 When presented with prognostic information about outcomes of CPR in the form of a pie, elderly patients changed their minds about their preferences for resuscitation.16 Perhaps the best information we have for questions about the role of prognostic data in the care of critically ill patients is the SUPPORT study. This large, expensive, randomized, controlled trial provided half of the enrolled patients, their families, and physicians with extensive prognostic information, including the outcomes of cardiopulmonary resuscitation, six-month survival estimates, and probability of severe disability. These data were provided by trained nurses who engaged in detailed discussions with families. Despite this effort, there was no measurable effect of the SUPPORT intervention on the time until a do not resuscitate (DNR) order was written, agreement on whether the patient wanted a DNR order, the amount of time spent by the patient in undesirable states or pain, or the cost of care. Certainly there has been much commentary about the SUPPORT study and why the SUPPORT intervention failed. Some commentators felt it was wrong to have a nurse provide the prognostic data. Others felt that a more systemwide approach needed to be taken to change the culture of the providers to listen more carefully to patients. Others felt there was no indication of a problem and that patients and families basically got the care they wanted at the time decisions were made and that patients actually want more intensive care than we might expect prior to death. The SUPPORT study has posed many challenges to those who believe that more accurate and earlier prognostic information can affect decision making.
Recommendations
Decades of research into human decision making indicate that clinicians will generate biased prognoses for individual patients. Despite this fact, prognostic information is increasingly important in the ICU, where decisions to limit life-sustaining therapy occur on a daily basis. Simple clinical prediction rules that apply in diseases commonly seen in the ICU (for example, the acute respiratory distress syndrome or multiple organ failure) should be developed. These rules should take into account comorbidities and functional status prior to admission to the ICU as well as the patient’s response to initial therapy. Valid clinical prediction rules that identify patients with poor prognoses will be more helpful to clinicians and families than a model that yields a continuous risk assessment.
Even armed with a set of clinical prediction rules, decisions to limit life-sustaining therapy will never be easy. We need much more information about how families make these decisions and what sort of information they need to help them in their task. We need to understand how to better communicate prognostic information to patients and elicit informed consent for intensive care from patients and their families. Finally, critical care clinicians need to understand how to provide aggressive palliative care in the same way we know how to provide aggressive life-sustaining therapy. When we cannot cure, we need to know how to comfort.
References
1. Attar SM, et al. Surg Forum 1966;17:8-11.
2. Afifi AA, et al. N Engl J Med 1971;285:1497-1502.
3. The SUPPORT Principal Investigators. JAMA 1995; 274:1591-1598.
4. Prendergast TJ, Luce JM. Am J Respir Crit Care Med 1997;155:15-20.
5. Knaus WA, et al. Science 1991;254:389-394.
6. Cook DJ, et al. JAMA 1995;273:703-708.
7. Kahneman D, et al. Judgment Under Uncertainty: Heuristics and Biases. Cambridge, NY: Cambridge University Press; 1987.
8. Redelmeier DA, Tversky A. N Engl J Med 1990; 322:1162-1164.
9. Schwenzer KJ, et al. Anesth Analg 1993;76:478-484.
10. Report of the Ad Hoc Committee of the Harvard Medical School to Examine the Definition of Brain Death. JAMA 1968;205:337-340.
11. Shellman RG, et al. Crit Care Med 1988;16:671-678.
12. Faber-Langendoen K. Arch Intern Med 1991;151: 235-239.
13. Tirschwell DL, et al. Neurology 1997;48:352-357.
14. Wasson JH, et al. N Engl J Med 1985;313:793-799.
15. McNeil BJ, et al. N Engl J Med 1982;306: 1259-1262.
16. Murphy DJ, et al. N Engl J Med 1994;330: 545-549.
Subscribe Now for Access
You have reached your article limit for the month. We hope you found our articles both enjoyable and insightful. For information on new subscriptions, product trials, alternative billing arrangements or group and site discounts please call 800-688-2421. We look forward to having you as a long-term member of the Relias Media community.