Learning the Art and Science of Diagnosis

“The diagnostic process begins with gathering data. Key elements involve ascertaining the person’s current concerns; reviewing the medical history; performing a physical examination; evaluating findings from laboratory, imaging, and pathology studies; and exploring the inferences and plans of previous clinicians. In the modern era, much of this data gathering takes place through a review of the electronic medical record. While that is a valuable and efficient tool, physicians must continue to learn the value of listening to a person’s descriptions and accounts of their symptoms and concerns firsthand. [..]

After gathering data, the next step is to determine which pieces of information (findings) are salient and which are not. This is an enormous challenge for clinicians, as individuals may report their concerns or clinical “stories” in a variety of ways that can potentially lead a physician to form very different hypotheses. Identifying the correct “stem” of the story—the key finding or combination of findings on which to build a list of possible diagnoses—is crucial and requires experience and practice over time. [..]

What clinicians really want to know is the opposite direction of conditional probability—the chance that someone with lymphadenopathy has lupus, ie, the positive predictive value. The Bayes theorem teaches that the prevalence, or prior probability, and the specificity of the finding connect these 2 conditional probabilities (sensitivity and positive predictive value). For diagnosticians, this problem-solving activity is iterative; comparing the list of the person’s findings with the disease profiles will help guide acquisition of new information and revise the list of potential diagnoses (eg, a negative antinuclear antibody test result might rule out lupus). The exercise requires that physicians learn how to access information about the accuracy of tests derived from scientific knowledge, such as sensitivity and specificity or likelihood ratios.

But there is much more to this process than clinical knowledge. Physicians must understand the variety of ways an illness can present to recognize it. For example, individuals with pulmonary embolism can present with several distinct clusters of findings. Pulmonary infarction will often cause pleuritic chest pain, hemoptysis, fever, and a wedge-shaped defect on chest imaging. Showers of emboli could lead to dyspnea, tachycardia, and a feeling of doom. Massive pulmonary emboli could result in syncope, chest pain, and signs of right heart strain on physical examination and electrocardiogram. Additionally, study-derived estimates of likelihood ratios for these findings may be based on individuals who participate in research studies in settings that differ greatly from people who seek care in other settings. Appreciating these variations requires seeing many people over time with follow-up and feedback as continuous learning.

A key lesson in the diagnostic process is to appreciate when clinicians need synchronous communication (when all involved participants in the discussion are present at the same time) and when they should use asynchronous communication (when the participants hear, read, or access the information at different times). One determinant is urgency; for instance, a radiologist who identifies a dissecting aortic aneurysm on imaging cannot simply send an email about the finding to the clinician who ordered the test. A more subtle requirement for synchronous communication is the extent to which a clinical decision depends on the nuanced interpretation of an image, pathology specimen, or consultant’s opinion. In those cases, real-time synchronous exchange of ideas and information is invaluable. The advantages of synchronous communication can be difficult to teach. Modern cancer management requires “tumor boards” with synchronous interdisciplinary meetings to derive clinical plans. Many other disciplines in medicine lag behind this approach. [..]

Physicians also need to learn to differentiate circumstances in which making a diagnosis is urgent from when it is not. To do so, they need to be able to answer 3 questions. Does the person have a condition or acute illness that is potentially life-threatening? Does the person potentially have a problem that can be treated successfully? Must this problem be treated immediately? Conversely, for a person without acute illness, there are times when it is best to order tests serially over time with close follow-up by the same physician. The ability to answer these questions correctly ultimately requires good judgment. [..]

Artificial intelligence applications might augment a physician’s knowledge in directing the search for and derivation of a correct diagnosis. Artificial intelligence (and other related machine learning technologies) has been successfully applied in reading images like radiographs, skin lesions, or retinal scans.

However, attempts involving diagnoses that require integration of clinical findings have not achieved the same success (such as IBM Watson Health). Failure to date is likely the result of an inability to perfect the choice of key data (ie, the “stem” of a person’s story) that provide the inputs for machine-based diagnostic algorithms. Another recent difference is the reality that much of the same information physicians access online is available to the people who seek help, so “self-diagnosis” is more common.

[..] diagnosis involves both the art and the science of medicine. At times, diagnosis involves fast thinking via pattern recognition (for people who have findings that are highly specific for a certain disease), whereas at other times, it involves slower thinking with iterative analyses. Putting it all together to achieve diagnostic excellence requires caring, curiosity, practice, experience, and feedback, all components of lifelong learning that contribute to the joy and satisfaction derived from the practice of medicine.”

Full article, AS Detsky JAMA