Differential diagnosis: how to train with clues
The preceptor asks for the differential of an adolescent who vomits, and the only guess that comes to mind is pyloric stenosis, a condition typical of infants. The scene is common at the end of the clinical rotation: the material was studied, but when it is time to list alternatives, memory delivers the first name that appears. Knowing many diseases is different from knowing how to compare them when the case arrives out of order.
NEXO is a daily medical diagnosis game: one clinical case a day, six progressive clues, and ICD-10 proximity feedback across five levels.
This text shows how to do differential diagnosis in a trainable way: generate guesses early, treat each new piece of data as a test that raises or lowers the probability of each one, compare two similar diseases with guideline data and recognize the biases that most undermine reasoning. At the end, we show how to set up a daily training with cases in clues.
Main points
- The differential starts with a few guesses, generated early, which are reviewed with each new piece of data.
- Each finding works as a test: the likelihood ratio says how much it increases or decreases the probability of a disease.
- Two similar diseases are separated by findings that weigh more toward one side, and rarely by a single finding.
- Closing the diagnosis too early was the most common cognitive cause in a series of diagnostic errors in internal medicine.
- Knowing the names of the biases helps little; organizing knowledge and reflecting on the findings of each guess has better results in studies.
Summary
- Why the differential gets stuck
- Early guess and review with each piece of data
- Each clue as a test
- Two similar diseases side by side
- Anchoring and premature closure
- How to train with clues
- Frequently asked questions
- Differential checklist
Why the differential gets stuck
Differential diagnosis is the list of diseases that can explain a patient's condition, ordered by probability and severity, and the process of eliminating or confirming each one. On the written exam, the list usually comes ready in the alternatives. In patient care, you need to generate it yourself.
Two ways of thinking that work together
Current theories describe clinical reasoning with two modes: a fast and intuitive one, which recognizes patterns, and a slow and analytical one, which compares guesses (Croskerry, 2009). A review for clinical teachers concludes that both modes have strengths and weaknesses, and that the student should learn to use both flexibly, according to the case (Eva, 2005).
The differential gets stuck when one of the modes is missing. Without pattern recognition, the list becomes long and unordered. Without analysis, the first guess that appears becomes the only one.
Diagnostic error is frequent
Getting the diagnosis wrong happens with trained physicians, and not only with students. An estimate made from three large observational studies in the United States reached a diagnostic error rate of 5.08% in adult outpatient care (Singh et al., 2014). In a series of 100 diagnostic errors involving internists, cognitive factors contributed in 74% of situations and health system factors in 65% (Graber et al., 2005). Training the differential is precisely training the cognitive part.
Where this text fits
The text on case-based learning explains how to study with cases and why structured reflection helps. To generate the initial list of causes with a mnemonic, see the guide on VINDICATE. Here the focus is the next step: how the list changes with each piece of data and how to train that.
Early guess and review with each piece of data
Generating guesses right at the start of the case gives direction to the rest of the investigation. The main complaint and the first pieces of data already suggest some diseases, and the rest of the investigation serves to separate them.
Few guesses, well organized
Working memory has limited capacity, which restricts analytical reasoning (Norman et al., 2017), and a list of twenty diseases is hard to hold at the same time. What helps is organizing knowledge into schemes: for a complaint such as abdominal pain or jaundice, a map of major groups of cause that you go through in order. In a study with think-aloud protocols in gastroenterology, the chance of a correct answer was higher when participants used pattern recognition or scheme-guided reasoning than with pure hypothetical-deductive reasoning (Coderre et al., 2003).
In practice, this means entering the case with three or four hypotheses from different groups, instead of ten from the same area. A mnemonic such as VINDICATE helps ensure that the major groups of cause have been remembered.
Serious causes enter the list early
The list is not ordered by probability alone. A less likely but serious and treatable disease deserves a place among the first hypotheses until it is ruled out, because the cost of letting it pass is high. In practice, the initial list usually has two or three common causes and one serious cause that you need to rule out. That serious cause changes the order of the investigation, even when it is not the favorite.
Separating the two questions helps: "what is most likely?" and "what can I not let pass?". The answers do not always coincide, and a well-made differential answers both.
Each new piece of data reorders the list
With each finding, ask: this piece of data is more expected in which of my hypotheses? If it fits all of them, it weighs little. If it fits one and is rare in the others, it changes the order. If none explains the finding, the list needs a new hypothesis.
This habit is the opposite of looking for data that confirm the first idea. It also makes clear when a finding is discordant, that is, when it contradicts the favorite hypothesis and calls for a pause.
Reflecting on the alternatives teaches more
In an experiment with 110 fourth-year students, practicing cases with structured reflection, listing findings for and against each hypothesis, yielded more correct answers one week later than giving a single diagnosis or listing a differential without justification (Mamede et al., 2014). The gain appeared both in cases of the trained diseases and in cases of new diseases, which were plausible alternatives to the trained ones (Mamede et al., 2014). Listing the differential, alone, had a result similar to giving a single diagnosis in the same study.
The lesson for training is direct: the list is worth what you do with it. For each hypothesis, write what supports it, what contradicts it and what you would expect to find if it were the right one.
Each clue as a test
In the language of evidence-based medicine, every finding from the history and physical examination works as a diagnostic test. It has a likelihood ratio, which says how much the presence or absence of the finding changes the probability of a disease.
The rule of 15, 30 and 45
Calculating the probability after a finding requires converting probability into odds and back, which few people do in their head. A simplified method solves this with three numbers: a likelihood ratio of 2 increases the probability of the disease by about 15 percentage points, one of 5 by about 30 and one of 10 by about 45 (McGee, 2002). The inverses work the other way: 0.5 decreases about 15 points, 0.2 about 30 and 0.1 about 45 (McGee, 2002). The approximation holds for initial probabilities between 10% and 90% (McGee, 2002).
You do not need to memorize the ratios of each finding to use the idea. Just ask, with each clue, whether it weighs little, a little or a lot, and for which hypothesis.
Findings that weigh little and findings that weigh a lot
Findings that appear in almost all diseases on the list weigh little: fever in an infectious condition, for example, rarely separates one cause from another. Findings that are rare outside one disease weigh a lot. In the review on streptococcal pharyngitis, no isolated finding from the history or examination was precise enough to confirm or rule out the diagnosis alone (Ebell et al., 2000). That is why clinical rules combine several findings.
The absence of a finding also informs
Beginners tend to look only at what is present. The absence of an expected finding also changes the probability, and sometimes weighs more than the presence of another. In pharyngitis, the absence of painful anterior cervical lymph nodes had a negative likelihood ratio of 0.60 in the clinical exam review (Ebell et al., 2000). In a case with clues, also notice what the clue did not mention and whether that fits your hypothesis.
How information changes the hypothesis
Think of the case with clues as a sequence of tests applied in order. The first clue suggests groups of diseases; each following clue raises one hypothesis and lowers another. The training is in making this update consciously: after each clue, did your list change? If it did not change, did the clue confirm what you already thought or did you fail to look at it?
Two similar diseases side by side
A good differential exercise is to compare two diseases that are confused in practice. Group A streptococcal pharyngitis and viral pharyngitis are a classic example: both cause sore throat, and management changes according to the cause.
What the guideline and the review show
According to the guideline from the Infectious Diseases Society of America, group A streptococcus accounts for 5% to 15% of visits for sore throat in adults and 20% to 30% in children (Shulman et al., 2012). The same guideline lists the findings that suggest each cause, summarized in the table below (Shulman et al., 2012).
| Finding | Suggests group A streptococcus | Suggests viral cause |
|---|---|---|
| Onset and age | Sore throat of sudden onset; 5 to 15 years | – |
| Season and context | Winter and early spring (in the northern hemisphere); exposure to streptococcal pharyngitis | – |
| General symptoms | Fever, headache; nausea, vomiting and abdominal pain | Absence of fever; diarrhea |
| Oropharynx | Tonsillopharyngeal inflammation; patchy exudate; petechiae on the palate | Stomatitis with discrete ulcers |
| Neck | Anterior cervical adenitis, with painful lymph nodes | – |
| Airways and eyes | – | Cough, runny nose, hoarseness, conjunctivitis |
| Skin | Scarletiniform rash | Viral rash |
Table source: Table 4 and text of the guideline by Shulman et al., 2012, translated and reorganized by group of finding; the dash marks the rows in which the guideline lists findings on only one side. Clinical content for medical review.
How much each finding weighs
In the systematic review of the clinical exam, the most useful findings to increase suspicion of streptococcus were tonsillar exudate, pharyngeal exudate and exposure to streptococcal pharyngitis in the previous two weeks, with positive likelihood ratios of 3.4, 2.1 and 1.9 (Ebell et al., 2000). To decrease suspicion, the most useful findings were the absence of painful anterior cervical lymph nodes, of tonsillar enlargement or of exudate, with negative ratios of 0.60, 0.63 and 0.74 (Ebell et al., 2000).
By the rule of 15, 30 and 45, tonsillar exudate raises the probability by about 20 points, and the absence of painful lymph nodes lowers it by about 10 to 15 (McGee, 2002). Neither of the two, alone, closes the case.
What the comparison teaches
The guideline recommends rapid test or oropharyngeal culture because clinical findings, alone, separate the two causes poorly, except when there are evident viral signs, such as runny nose, cough, oral ulcers or hoarseness (Shulman et al., 2012). That is the point of the exercise: the clinical differential orders the hypotheses and shows how much each finding weighs, and the next step of the investigation comes from that order.
Repeat the exercise with other pairs that you confuse: write the findings in three columns, in favor of one, in favor of the other and common to both. The middle column is the one that helps least to decide.
Anchoring and premature closure
Cognitive biases are shortcuts of thought that, in certain situations, lead to error. Two of them appear frequently in the differential: anchoring, which is getting stuck on the first impression, and premature closure, which is stopping considering alternatives after reaching a plausible hypothesis.
The catalog of biases
A classic article catalogued the main biases in diagnosis, called cognitive dispositions to respond, and proposed strategies to reduce them, including metacognition: taking a step back and examining your own reasoning (Croskerry, 2003). Anchoring and premature closure are on that list.
What the data show
In the series of diagnostic errors with internists, premature closure, defined as failing to consider reasonable alternatives after reaching an initial diagnosis, was the most common single cause (Graber et al., 2005). In a systematic review with 20 studies, overconfidence, anchoring and information and availability biases were associated with diagnostic inaccuracy in 36.5% to 77% of case scenarios, with the caveat that 60% of the studies were classified as low quality (Saposnik et al., 2016).
Availability bias, a close relative of anchoring, has also been reproduced in an experiment. Second-year internal medicine residents who had recently seen similar cases made more mistakes on new cases with similar findings, and reflective reasoning corrected part of those mistakes (Mamede et al., 2010). It is the mechanism of the opening scene: the disease you studied last is the first one to come to mind.
Knowing the name of the bias helps little
A review on the causes of error in reasoning concluded that strategies aimed at recognizing biases were ineffective at reducing errors, while strategies to reorganize knowledge had small but consistent benefits (Norman et al., 2017). In a trial with students on an emergency rotation, training in strategies to force cognition did not reduce diagnostic error (Sherbino et al., 2014). And when residents decided on their own to review the diagnoses, the review brought minimal gain compared with knowing the right answer (Monteiro et al., 2015).
How anchoring shows up in a case with clues
In a case with clues, anchoring has a format that is easy to recognize. The first clue suggests a disease, and the following clues start to be read as confirmation, even when they bring a piece of data that would weigh more toward another hypothesis. The warning sign is realizing, at the end, that the decisive clue was there since the middle of the case and you read it as a detail. Writing down the clue that does not fit, at the moment it appears, is the simplest antidote to this pattern.
What has given the best results
Guided reflection was the intervention with the most consistent results in a systematic review on interventions for diagnostic reasoning, in five studies (Lambe et al., 2016). Checklists for diagnosis, including differential lists by complaint, have been proposed to reduce reliance on memory, but they are still a proposal under evaluation (Ely et al., 2011).
For the student, the sum of these findings points to a practical path: know the diseases and the differences between them well, practice many cases and, in difficult cases, stop to check the findings for and against before closing.
How to train with clues
The differential improves with deliberate practice: many cases, with feedback on whether you were right and on how far off the error was. Cases with clues give this kind of practice in a short time.
A script for each case
- On the first clue, write three hypotheses from different groups.
- With each clue, ask whether it weighs little, a little or a lot, and for which hypothesis.
- When a clue does not match your favorite, write it down. It is what protects against premature closure.
- Before the guess, list one finding for and one against the chosen hypothesis.
- After the result, compare the right disease with the one you guessed and write the finding that separates the two.
Train on NEXO
On NEXO, the Daily case arrives in up to six clues, and each guess shows the proximity in ICD-10 across five levels. The proximity tells you whether you missed within the right group or far from it, which separates an error of comparison between similar diseases from an error of direction. On the third clue the area of the case appears and, on the fifth, the ICD range; you can skip a clue whenever you want to move forward. This design matches the script above: hypotheses early, review with each clue and feedback on how far off the error was.
The Archive gathers previous daily cases, good for training an area before the exam. Multimedia brings cases with images, audio and other exams, and Continents, cases linked to Africa, the Americas, Asia, Europe and Oceania, which changes the epidemiology of the differential. Adrenaline, with the dose chosen beforehand and one case per day, trains reasoning against the clock. To understand the limits of each mode, see the NEXO economy page.
Keep what separated the two diseases
The finding that separates the right disease from the one you guessed is what is worth keeping. The Anki button creates the card straight from NEXO and opens it in the Anki app, searches the cards you already have and, when the case has an image, takes the image to the card. The text on how to fit games into your study routine shows how to link the case to the book, the cards and the questions.
Discuss in a group
The ranking has the Global, Universities and Group tabs, the latter with an invite code. With your class or league playing the same Daily case, it is worth comparing at which clue each person changed their guess and why. If you coordinate a medical school, league or residency and want to use cases like these in teaching, check out NEXO Closer. Medical education researchers who want to study clue-based training can check out NEXO Researcher, and doctors who want to review cases, NEXO Curator.
Frequently asked questions
How do you do a step-by-step differential diagnosis?
Start with three or four guesses from different groups right from the first pieces of data. With each new finding, ask which guess it is most expected in and reorder the list. Before closing, list what supports and what contradicts the favorite guess. Structured reflection on the findings of each guess improved student learning in an experiment (Mamede et al., 2014).
How many guesses should I list?
Few, and from different groups. Working memory has limited capacity (Norman et al., 2017), and what helps is having schemes organized by complaint (Coderre et al., 2003). Three or four well-chosen guesses, covering common causes and serious causes, yield more than a long, unordered list.
What is a likelihood ratio and how do you use it in practice?
It is the number that says how much a finding changes the probability of a disease. By the simplified rule, ratios of 2, 5 and 10 increase the probability by about 15, 30 and 45 points, and 0.5, 0.2 and 0.1 decrease it by the same amount (McGee, 2002).
What is the difference between a diagnostic guess and a differential diagnosis?
The diagnostic guess is the disease you consider most likely at a given moment in the case. The differential diagnosis is the list of diseases competing with it, ordered by probability and severity. The two change together: each new finding can swap the main guess's place with an alternative on the list.
What is premature closure?
It is stopping considering reasonable alternatives after reaching an initial diagnosis. It was the most common single cause in a series of diagnostic errors with internists (Graber et al., 2005).
How do you train differential diagnosis on your own?
Use short cases and follow the same script in all of them: guesses early, review with each piece of data, findings for and against before closing, and comparison with the right answer at the end. Reflection on the findings of each guess was what helped students most in getting new cases right a week later (Mamede et al., 2014). Keep the finding that separated the right disease from your guess, because that is what you will need in the next similar case.
Does studying cognitive biases reduce error?
Studies indicate that recognizing biases, by itself, had little effect, and that organizing knowledge helped more (Norman et al., 2017). Guided reflection on the findings was the most consistent intervention in a systematic review (Lambe et al., 2016).
How does VINDICATE fit into the differential?
It helps generate the initial list, reminding you of the major groups of cause. The VINDICATE guide explains each letter. Once the list is generated, the work is that of this text: weighing each finding and reordering the guesses.
How does NEXO help train the differential?
Each case arrives with up to six clues, and the guess shows the proximity in ICD-10, which indicates whether the error was between neighboring diseases or in direction. With one case per day in the Daily, you practice the hypothesis script early and review at each clue. To learn about the game, read what NEXO is; for the evidence on games in teaching, the text on gamification in medical education.
Differential checklist
- Write three or four hypotheses from different groups right at the start.
- Include a serious cause that cannot be missed, even if less likely.
- With each data point, reorder the list and note why.
- Ask how much the finding weighs: little, a little or a lot.
- Note the finding that does not fit the favorite hypothesis.
- List what supports and what contradicts before closing.
- Be wary of the disease you studied last.
- Compare the right disease with your guess and keep the finding that separates the two.
- Repeat with pairs of similar diseases in three columns.
- Practice one case per day to keep the habit.
Want to train today? Play the case of the day on NEXO.
References
- Eva KW. What every teacher needs to know about clinical reasoning. Med Educ. 2005. https://pubmed.ncbi.nlm.nih.gov/15612906/
- McGee S. Simplifying likelihood ratios. J Gen Intern Med. 2002. https://pubmed.ncbi.nlm.nih.gov/12213147/
- Croskerry P. The importance of cognitive errors in diagnosis and strategies to minimize them. Acad Med. 2003. https://pubmed.ncbi.nlm.nih.gov/12915363/
- Croskerry P. A universal model of diagnostic reasoning. Acad Med. 2009. https://pubmed.ncbi.nlm.nih.gov/19638766/
- Graber ML, Franklin N, Gordon R. Diagnostic error in internal medicine. Arch Intern Med. 2005. https://pubmed.ncbi.nlm.nih.gov/16009864/
- Saposnik G, Redelmeier D, Ruff CC, Tobler PN. Cognitive biases associated with medical decisions: a systematic review. BMC Med Inform Decis Mak. 2016. https://pubmed.ncbi.nlm.nih.gov/27809908/
- Mamede S, van Gog T, van den Berge K, et al. Effect of availability bias and reflective reasoning on diagnostic accuracy among internal medicine residents. JAMA. 2010. https://pubmed.ncbi.nlm.nih.gov/20841533/
- Norman GR, Monteiro SD, Sherbino J, et al. The Causes of Errors in Clinical Reasoning: Cognitive Biases, Knowledge Deficits, and Dual Process Thinking. Acad Med. 2017. https://pubmed.ncbi.nlm.nih.gov/27782919/
- Ely JW, Graber ML, Croskerry P. Checklists to reduce diagnostic errors. Acad Med. 2011. https://pubmed.ncbi.nlm.nih.gov/21248608/
- Ebell MH, Smith MA, Barry HC, Ives K, Carey M. The rational clinical examination. Does this patient have strep throat? JAMA. 2000. https://pubmed.ncbi.nlm.nih.gov/11147989/
- Shulman ST, Bisno AL, Clegg HW, et al. Clinical practice guideline for the diagnosis and management of group A streptococcal pharyngitis: 2012 update by the Infectious Diseases Society of America. Clin Infect Dis. 2012. https://pubmed.ncbi.nlm.nih.gov/22965026/
- Mamede S, van Gog T, Sampaio AM, et al. How can students' diagnostic competence benefit most from practice with clinical cases? The effects of structured reflection on future diagnosis of the same and novel diseases. Acad Med. 2014. https://pubmed.ncbi.nlm.nih.gov/24280855/
- Coderre S, Mandin H, Harasym PH, Fick GH. Diagnostic reasoning strategies and diagnostic success. Med Educ. 2003. https://pubmed.ncbi.nlm.nih.gov/12895249/
- Singh H, Meyer AN, Thomas EJ. The frequency of diagnostic errors in outpatient care: estimations from three large observational studies involving US adult populations. BMJ Qual Saf. 2014. https://pubmed.ncbi.nlm.nih.gov/24742777/
- Lambe KA, O'Reilly G, Kelly BD, Curristan S. Dual-process cognitive interventions to enhance diagnostic reasoning: a systematic review. BMJ Qual Saf. 2016. https://pubmed.ncbi.nlm.nih.gov/26873253/
- Sherbino J, Kulasegaram K, Howey E, Norman G. Ineffectiveness of cognitive forcing strategies to reduce biases in diagnostic reasoning: a controlled trial. CJEM. 2014. https://pubmed.ncbi.nlm.nih.gov/24423999/
- Monteiro SD, Sherbino J, Patel A, et al. Reflecting on Diagnostic Errors: Taking a Second Look is Not Enough. J Gen Intern Med. 2015. https://pubmed.ncbi.nlm.nih.gov/26173528/