In his book Surfing Uncertainty: Prediction, Action, and the Embodied Mind, Andy Clark used the metaphor of driving on a foggy but familiar road versus driving on a similarly foggy but unfamiliar road, to the tension between the top-down predictions and bottom-up sensory data. When navigating the familiar but foggy road, memory of that road and experiences of driving in fog carry weight, whereas sensory information is less helpful. In contrast, driving on a similarly foggy but unfamiliar road, your prior knowledge is less helpful and your attention is directed towards improving the quality of sensory data and looking for areas in the environment that yield salient clues. Now, imagine driving on a familiar road on a clear day. Guided by prior knowledge, your attention to the “outside world” is not needed and thus, directed inwards; you arrive at the destination with hardly any effort or awareness of how (future essay).
It is through the mechanism of action-oriented predictive attention that humans seek relevant information, assess the quality of the information, find data for their explanations, and also update our beliefs. We anticipatorily move our eyes to where highly reliable and salient information resides and we actively change the environment, so it is more aligned to our explanations. Once causes and effects are identified for a task, we selectively sample from this source of data and in effect ignore (or at least minimize) the influence of less relevant sources. Returning to the analogy, when driving on a foggy and unfamiliar road, we attend to a map to anticipate turns and drive slower to increase the quality of sensory data. Once causes and effects are identified for a task, we selectively sample from this source of data and in effect ignore (or at least minimize) the influence of less relevant sources. It is the balance of top-down explanations seeking corroborative data and bottom up information seeking plausible explanations that help us navigate the noisy and ambiguous world.
Many patient encounters in the emergency department are shrouded in fog and the road often has the illusion of familiarity and the mirage of clarity. Information presented to providers is incomplete, imprecise, and inconsistent. Patients ambiguously describe their symptoms, are unable to provide a history, or experience their symptoms “atypically.” Patients with complex socio-medical histories often do not know their medications nor their medical history. Historically, emergency physicians leaned on their history, physical, and observational skills to generate differential diagnoses. However over the past 15 years, electronic health records (EHR) have become the locus of activity in the emergency department. EHRs are not only the repository of patient data but also where the patient encounter is documented. It has been estimated that physicians spend two-thirds of their time interacting with the electronic health record and 15% of their time in “direct” patient care. Even real-time data such as vital signs are interpreted in the ecosystem of the electronic health record and away from the patient. In effect, attentional resources have been diverted away from an encounter with the patient to an encounter with the electronic health record.
In the ecology of the emergency department (future essay), this attention away from the patient, and to the EHR can have potentially deleterious consequences. Hurried, reflexive, and overly reductive (mis)classifications of complex and ambiguous presentations can not only cascade within an encounter from triage to disposition but also reverberate across encounters. This is especially problematic for the emergency physician because we are not only tasked with generating a differential diagnosis that includes the most prevalent diagnoses but also the most life threatening. Is this headache a migraine or a ruptured aneurysm? Does the prior change if she was here last week for a headache? Or if headaches are her problem list? Or if she can’t provide a coherent history because of dementia or intoxication? In the face of this escalating complexity and the dual mandate to respect and ignore population level base rates, inferences based predominantly on EHR data can be precarious. Although patient generated data from a history and physical exam are generally less reliable than “objective” measures such as labs tests and radiological imaging, it is sometimes a subtle but salient clue from an incisive history and physical exam that leads to the intuitive, informed, and targeted search for a life-threatening but less common diagnoses.
It has been said that “not all Bayesian inference is optimal but all optimal inference is Bayesian.” In a complex and foggy world, the features of Bayesian inference – subjective priors, attentional search for data, precision weighting of sensory data, knowing when and how much to update beliefs (next essay) – are possibly the best tool we have to model and infer features of the world. However, attention is a finite resource and itself is a prediction of salience and reliability. Once patterns are learned and reliability of data is established those same sources of information are selectively sampled at the expense of alternative sources. This can be problematic when you are sampling data that is not fully relevant to the task. In the emergency department, the task is to diagnose the common and identify the dangerous. The attentional-shift away from the direct observation of the patient and to EHR is just such a selective sampling that is not fully relevant to the emergency physician’s task.