Clinical decision-making, cognitive biases and the importance of time

Regardless of the process used to reach a decision in the setting of an emergency, the final decision is essentially binary: ‘yes’ versus ‘no’; ‘do’ versus ‘do not’; ‘proceed’ versus ‘cancel’. A third option, ‘undetermined’ or ‘wait for more information’ could be the same as a ‘no’ decision, given the critical importance of time in decision-making.4 Unfortunately, much of the difficulty involved with making decisions regarding treatments for patients with acute stroke is that ‘time is brain’. In fact, it is estimated that 1.9 million neurons die per minute during an untreated stroke.5 Further evidence for the importance of time is provided by the lack of benefit from intravenous tissue plasminogen activator (tPA) beyond the 4.5 h window, with earlier treatment affording greater benefit.6 ,7 Therefore, when a patient presents with symptoms of acute ischemic stroke, delay in diagnosis and initiation of treatment must be avoided. The diagnostic benefit gained from obtaining an additional imaging procedure must be weighed against the time (and brain) sacrificed by waiting.8 For this reason, in patients with acute stroke a delayed decision may end up being the same as a ‘no’ decision from the perspective of patient outcome.

In an acute stroke scenario, when a physician encounters a decision-making step, he/she deals with probabilities. A physician is able to estimate these probabilities using his/her prior experiences and knowledge gained from peers, books and publications. This estimation is, however, prone to cognitive biases, the most striking of which is an ‘outcome’ bias.9 ,10

Information becomes available sequentially, not simultaneously

Consider the patient who presents to the emergency room at a tertiary care hospital after suffering sudden-onset aphasia and right hemiparesis at home. Upon presentation, several key factors required in the decision to proceed with thrombolysis or thrombectomy are apparent. Much of the history and an initial examination have already been obtained by emergency responders such that the emergency room is aware of these factors prior to arrival. The patient's age is usually known or can be estimated. The time of onset or time last seen normal is usually known. The patient's medical and social history may be available through family members or medical records, detailing comorbidities, medications, type of residence and living will. This information will include known contraindications for thrombolysis. The physical examination provides the vital signs, clinical condition, suspected vascular territory involved and the stroke severity. Furthermore, intravenous lines have been placed and laboratory studies are sent at this time.

From this point forward, further information becomes available sequentially rather than simultaneously, thus consuming more time before a decision is made. Most centers use imaging studies with non-contrast CT versus CT angiography, with or without perfusion studies. Non-contrast imaging is obtained in minutes and can usually be reviewed immediately for the presence or absence of intracranial hemorrhage (seconds–minutes). A more thorough review will also provide further evidence of stroke, such as the presence of a dense middle cerebral artery (MCA) sign and determination of the Alberta Stroke Program Early CT (ASPECTS) score (<5 min). Next, if CT angiographic imaging is obtained, the imaging can be reviewed in detail after completion (<10 min). Perfusion imaging, if obtained, requires longer for acquisition, transfer, post-processing and interpretation (15–25 min). Stat laboratory studies may be available in a series of minutes, such as the International Normalized Ratio (INR), but others may take longer (eg, platelets). MRI, if obtained, will provide further delays in acquisition and interpretation and also in ruling out exclusions such as pacemakers (>30 min). Waiting at each stage to make a more informed decision results in further delays in decision-making. The cognitive biases mentioned above that have prompted physicians to use these tests as ‘gold standards’ for decision-making may be preventing physicians from applying knowledge that is considered by everyone as true: the importance of saving time while achieving reperfusion. While very few clinicians would proceed emergently with thrombectomy without a CT scan (optimize time while sacrificing data), very few would also wait 2 h until both CT angiography and MRI had been completed to evaluate whether the patient is a candidate for thrombectomy (optimizing data while sacrificing time). It would therefore seem obvious that the pragmatic solution lies in the middle. The focus on efficiency and process must happen in parallel, and a clinical decision should not await tools that are still being evaluated. More importantly, physicians are very prone to information bias; each bit of information they gain during the process of managing patients has the ability to color their decision-making in ways that are not reflective of the true worth of that information.11 The pragmatic approach is to get by with the bare minimum information that is relevant to decision-making and to put the rest of the information accrued to scientific use later. Information accrued does not bias decision-making and can be used in a truly scientific manner to create decision algorithms for future validation and clinical use. How can one do this? The answer lies with a mathematician from the 18th century.

Tests give us test probabilities, not the actual probabilities

Thomas Bayes (1701–1761) was an English mathematician and Presbyterian minister known for having formulated a specific case of the theorem that bears his name—Bayes’ theorem.12 ‘Bayesian probability’ is the name given to several related interpretations of probability, wherein probability represents a belief rather than a frequency.13 ,14 Considering probabilities in this manner allows the application of probability to any number of propositions. Knowledge of the various cognitive biases highlighted above, including recency and outcome bias, confirmation bias, action bias, association bias, overconfidence effect and information bias along with a focus on saving time when making clinical decisions in the acute stroke milieu, could be put to use within a Bayesian framework for decision-making (a simple and elegant explanation is given at http://betterexplained.com/articles/an-intuitive-and-short-explanation-of-bayes-theorem/). There are several key principles within the Bayesian framework:

1. Diagnostic tests do not represent the index event: an imaging test to identify stroke may lead to the diagnosis, but the stroke occurred regardless of whether a test was ordered.

2. Diagnostic tests are inherently flawed: imaging tests identify events that do not exist (false positive) while missing events that do exist (false negative).

3. Diagnostic tests provide test probabilities, not actual probabilities: each test carries a probability of correct diagnosis and therefore the test result must be considered in terms of the error of the test. Since virtually no tests have 100% sensitivity and specificity, no test can provide 100% certainty of an accurate diagnosis.

4. Bayes’ theorem provides the actual probability of an event given the measured test probabilities.

Decision for thrombectomy as a Bayesian probability

Now let us consider the process of decision-making for endovascular treatment. While the criteria for candidacy for intra-arterial thrombectomy (IAT) are currently somewhat imprecise, a number of important factors play a role in the evolution of our decision for any given patient:

1. Persistent moderate or severe stroke: most centers agree that candidacy requires an NIH Stroke Scale (NIHSS) score of ≥8.

2. Age <80 years: while controversial, multiple studies have indicated that advanced age may be a poor prognostic factor.

3. Potential for independent living: the majority of centers would forego thrombectomy on those patients who, even with the best possible procedural outcome, would not be living independently. This includes patients with moderate to severe dementia, multiple debilitating comorbidities, those with certain pre-existing disabilities, those in nursing homes or patients already requiring assistance with activities of daily living.

4. Large vessel occlusion: this occlusion can be confirmed by imaging, such as with CT angiography, or suggested by clinical examination or non-contrast CT findings. Very few clinicians would proceed to the angiography suite with CT angiography imaging demonstrating no large vessel occlusion.

5. Time from symptom onset <8 h: the principle of time since stroke onset is currently highly controversial, given the increasing role of physiologic imaging in determining the status of the collaterals and/or penumbra. However, in general, most centers (and current trials) assess candidacy based on time, with many using 8 h as a threshold.

6. Presence of other imaging findings: while controversial, many centers use perfusion imaging findings to identify whether the ischemic territory is salvageable. Other considerations, such as vessel tortuosity, can be considered here.

7. Other factors: a number of contraindications exist, such as inability to obtain access due to vascular disease.

Each of these factors plays an important role in our decision to proceed with thrombectomy. While treatment centers and individual practitioners approach each patient differently, it is important to note that this accrual of information takes time, some of the accrual being serial rather than parallel. Therefore, the decision to pursue thrombectomy is an evolving decision with some factors pushing us towards intervening and some arguing against. Each bit of information processed through the physicians’ cognitive biases results in a Bayesian probability of benefit with IAT at each step, the estimated probability varying along a spectrum from 100% (absolutely will proceed) to 0% (absolutely will not). While not mathematically valid, considering the decision in this numeric context allows us to understand how our decisions evolve over time and how they are influenced by each subsequent bit of information. It is also worth noting that, even if all the factors mentioned above favor going towards IAT, it does not guarantee a good outcome due to the complex interplay of variables that come into play after the decision has been made. The key ones are: time elapsed between decision and recanalization, procedure complications, quality of recanalization, brain eloquence and patient's expectation/definition of good outcome. In addition, similar to mammography, we also have to think about false positives and negatives of the tests or factors we are considering. For example, if we use age of 80 years as being the threshold for decision-making, we may miss out on the bigger picture—for example, consider an 82-year-old who has run a marathon last week versus a 68-year-old with obesity, hypertension, diabetes, old infarcts and mild cognitive impairment.

We will consider how this plays out in a well-organized tertiary stroke center through two examples.

An algorithm for IAT decision-making

The principles of dual process theory, the nature of cognitive biases that physicians deal with, the sequential nature of the data that is presented and the Bayesian framework allows for the creation of an algorithm to assist clinicians with emergency decision-making for IAT. In most patients, pursuing thrombectomy is a careful and important decision process. We have found that rapid analysis with only the most relevant and necessary information makes the decision process faster and clearer. By using an algorithm to shorten the necessary explicit analysis, we reduce implicit bias while determining when we have reached enough confidence to proceed without need for further studies (‘no further test threshold’). Advanced imaging and other research studies go on in parallel at a rapid rate, interfering minimally with efficiency. This process helps us to recognize the importance of additional information while never being biased by an ‘information heuristic’. A suggested format of how information can be categorized and evaluated in a Bayesian framework for IAT decision-making is shown in figure 1. Whenever sufficient information is available that crosses the ‘no treatment’ or ‘no test’ threshold, a clinical decision is made. Note that, as one moves down the algorithm, the chance of a poor outcome with thrombectomy increases because time passes with each additional test obtained.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Basic algorithm depicting the decision-making process regarding thrombectomy in patients presenting with acute ischemic stroke. ASPECTS, Alberta Stroke Program Early CT Score; MCA, middle cerebral artery; NIHSS, NIH Stroke Scale.

This is a simplistic version of what this way of thinking can achieve. It is clear that the various factors mentioned above are not on an evenly spread scale—for example, an NIHSS score of 4 versus 9 can have a significant impact on decision-making while the same gap in the NIHSS score of 15 versus 20 is unlikely to influence decision-making. Similarly, an ASPECTS score of 3 versus 6 is much more important for decision-making than an ASPECTS score of 8 versus 10. Also, given the complexity and variability of stroke and the factors that influence patient outcome after the decision has been taken, our ability to calculate pretest probabilities and apply Bayesian mathematics precisely are limited at the present. However, we believe that with large well-documented datasets, shortening times to recanalization and reduction in procedural complication rates with newer devices, such calculations should be possible in the future. We feel that, even in the absence of precision, this model and the suggested way of decision-making adds value and speed. Hopefully, we will be able to add a greater degree of precision and sophistication to the model as more homogenous, prospectively collected large datasets become available.