As an Emergency physician, my job is to make predictions. Throughout the day (and night), I make predictions related to treatment outcomes, management plans, and patient dispositions. These predictions are often fraught with risk (but mostly uncertainty) and I am constantly aware that every decision I make is inherently probabilistic. The framework that most physicians lean on to make decisions is the hypothetico-deductive model. It is the most widely studied and accepted model of clinical decision making. It is defined as an iterative process of data collection and analysis designed to solve diagnostic problems and generate hypotheses. The central idea is one of sequential, staged data collection, followed by data interpretation and the generation of hypothesis. As data is collected at each stage, it is added to the growing database of observations and used to reformulate or refine the active hypothesis. Let’s take the example of a patient that presents to the emergency department with the chief complaint of chest pain. I instinctively glance at his age, gender, vital signs…take a look at his EKG… all the while creating a mental list of hypotheses. I subsequently parse through the EHR to look for clinically relevant past medical data (prior tests, labs, studies) and finally add the history and physical exam to create a differential diagnosis. Consideration of prevalence (or priors) is important during hypothesis generation not only because common things are common, but also because it will direct the history and physical exam (H&P) and ultimately drive the decision-making process.
This type of reasoning (unknowingly to most physicians) is based on the Bayes theorem – named after Thomas Bayes – who was an English philosopher and statistician. Bayes’s theorem is primarily concerned with conditional probability. That is, probability of a hypothesis in light of new evidence is proportional to its prior probability (“the prior) multiplied by the strength of the new evidence. The prior in medicine is subjective, as it is sometimes an unknown quantity based on limited data and the biases of the person assigning the value. It is a measure of belief that the claim is true and it requires us to think probabilistically about our predictions. Furthermore, the prior of one test can be the post of the next test. The ability to combine test results in series achieves greater importance once we accept that each question and physical examination during a clinical encounter constitutes a diagnostic test with an attached likelihood ratio. The two most important components of a differential diagnosis is the past medical history (known diagnoses, relevant procedures, imaging studies, and lab values) and the current presentation (vital signs, history and physical). For example, if the patient with chest pain has had prior abnormal cardiac testing (high prior), heart-related causes of chest pain will be higher on my hypothesis list and if my history and physical is concordant with this “high prior,” then I will even more preferentially test for this hypothesis. Conversely, if the patient with chest pain recently had a normal cardiac catheterization (low prior) and the H&P is consistent with this low prior, then cardiac-related causes of chest pain will be much lower on my differential diagnosis list and I will test preferentially for competing diagnoses. Overall, it is Bayesian reasoning that directs focused testing. Physicians don’t literally apply Bayes’s theorem every time we make decisions, but this practice of decision-making is intrinsically Bayesian.
However, it has been shown that humans (including physicians) are notoriously poor at calculating probabilities. We have a natural tendency to misperceive and miscalculate probabilities and to think anecdotally instead of statistically. Our decision-making is fraught with pitfalls such as confirmation biases and overconfidence. We are exposed to more inputs than our brains can consciously process, and therefore, we simplify and approximate. With experience, the simplifications and approximations are usually a useful guide and constitute our working knowledge but these biases are dangerous and can lead to faulty predictions. The miscalculation of priors because of our biases is what often leads to faulty predictions. Utilizing Bayesian reasoning, the electronic health record can be the perfect tool to decrease these biases. Imagine a scenario where a presentation of a patient to the Emergency Department with a specific chief complaint leads to a calculated set of priors based on the patient’s “risk factors.” The EHR integrates the stored past medical data with encoded medical literature to give a refined and updated prior. The provider then undertakes a history and physical and orders tests based on this calculated prior. This type of calculated pre-test probabilities has the potential to not only drive physician decision-making but also to improve care processes, the quality of care, and overall health system productivity. A Bayesian EHR will allow providers to focus on their strengths such as asking the right questions and interpreting results with what software is good at: computations, analysis, and statistics using large datasets.