Thinking diagnostically means moving from a patient’s symptoms, test results, and history toward a specific explanation of what is wrong and what to do about it. It sounds straightforward, but the process is riddled with cognitive shortcuts, technological limitations, and systemic gaps that affect millions of people every year. In the United States alone, emergency departments may produce roughly 7.4 million diagnostic errors annually, with hundreds of thousands resulting in serious harm. Understanding how diagnoses are made, where they fail, and how new tools are reshaping the process matters for clinicians and patients alike.
Two Systems Running at Once
When a physician sees a patient, two distinct mental processes fire almost simultaneously. The first is fast, automatic, and pattern-based: a clinician with years of experience sees a constellation of symptoms and a likely diagnosis surfaces almost instantly. The second is slower and deliberate, involving conscious reasoning through evidence, guidelines, and differential diagnoses. These two modes of thinking, sometimes called System 1 and System 2, form the backbone of what researchers call dual-process theory in clinical reasoning.1PubMed. Dual process models of clinical reasoning: The central role of knowledge in diagnostic expertise
The trouble is that these systems do not always agree, and their interaction can distort decisions. Research on treatment thresholds has shown that the fast, intuitive system can push a clinician’s willingness to treat well below what a careful analytical calculation would justify, helping explain patterns of overtreatment in everyday practice.2PubMed Central. Dual processing model of medical decision-making A busy emergency physician who saw three heart attacks this week might, without realizing it, weight cardiac diagnoses more heavily for the next patient with chest pain. That kind of availability bias is just one of many cognitive traps documented in clinical settings.
Cognitive Biases That Derail Diagnoses
A systematic review of cognitive biases in medical decision-making found that overconfidence, anchoring (locking onto an early impression), low tolerance for risk, and availability bias were linked to diagnostic inaccuracies in anywhere from about 37% to 77% of studied case scenarios.3PubMed Central. Cognitive biases associated with medical decisions: a systematic review Those numbers are striking, because they do not describe rare lapses. They describe how frequently normal human thinking bends diagnosis off course when the conditions are right for error.
Emergency rooms are a particularly fertile ground for these biases. A study of ER physicians found that the most commonly observed biases were overconfidence (about 23%), confirmation bias (21%), availability bias (12%), and anchoring (11%).4BMC Emergency Medicine. Cognitive biases encountered by physicians in the emergency room Confirmation bias is especially insidious: once a clinician forms an initial hypothesis, they tend to seek evidence that supports it and downplay findings that contradict it. In a setting where decisions happen under time pressure and incomplete information, these tendencies compound.
The Scale of Diagnostic Error
Diagnostic errors are not a niche problem. A systematic review commissioned by the Agency for Healthcare Research and Quality estimated that roughly 5.7% of all emergency department visits involve at least one diagnostic error. Scaled to the approximately 130 million annual ED visits in the United States, that translates to an estimated 7.4 million errors per year, with about 371,000 resulting in serious harm, including more than 100,000 cases of permanent high-severity disability and an estimated 250,000 deaths.5PubMed. Diagnostic Errors in the Emergency Department: A Systematic Review
A 2025 study looking at emergency hospitalizations found that about 3.2% of patients admitted for emergency conditions had been discharged from an ED within the previous nine days, suggesting a missed or delayed diagnosis. Those patients had higher 30-day mortality (about 15.7% versus 14.9% for similar patients without a prior discharge) and spent fewer days at home in the month following admission.6JAMA Network Open. Potential Diagnostic Error for Emergency Conditions, Mortality, and Healthy Days at Home The error rates varied widely by condition: spinal abscess was missed at the highest rate (about 16%), while spontaneous brain hemorrhage was missed at roughly 2%. This variation matters, because it means the risk of a missed diagnosis depends heavily on what the patient actually has.
Trigger-based strategies for identifying diagnostic errors in emergency settings have also been explored. One study found that among cases flagged by error-detection triggers, sepsis was the most common missed diagnosis, accounting for about a third of the identified errors. Patients presenting with vague complaints like altered mental status or shortness of breath were at higher risk.7BMJ Open Quality. Identifying diagnostic errors in the emergency department using trigger-based strategies Vague symptoms are, almost by definition, the hardest to diagnose and the easiest to dismiss.
The Diagnostic Odyssey in Rare Disease
For people with rare diseases, the diagnostic challenge is a different beast entirely. Rare conditions are inherently difficult to identify because individual clinicians may encounter them once or twice in a career, if ever. A large retrospective survey across Europe examined the factors that drive delays in rare disease diagnosis. Age at symptom onset was the strongest predictor of a prolonged diagnostic journey: adolescents at the time their symptoms began were nearly five times more likely to experience a significant delay compared to patients whose symptoms started later in life.8European Journal of Human Genetics. Time to diagnosis and determinants of diagnostic delays of people living with a rare disease: results of a Rare Barometer retrospective patient survey Women were also somewhat more likely to experience delays, particularly in reaching a specialist who could recognize the condition.
The reasons for these delays are layered: low disease prevalence, high symptom variability from patient to patient, and limited awareness among general practitioners all contribute.9PubMed Central. Repercussions of Diagnostic Delay in Rare Diseases Geography plays a role too. The European survey found that patients in Western and Northern Europe actually had higher odds of experiencing diagnostic delays compared to those in Eastern, Central, and Southern Europe, a counterintuitive finding that may reflect differences in referral pathways and healthcare system structure rather than clinical quality per se.
Tools for Measuring How Well a Test Works
Every diagnostic test sits somewhere on a trade-off between catching true cases and avoiding false alarms. Sensitivity measures how well a test identifies people who actually have a condition, while specificity measures how well it correctly clears people who do not. A test can be great at one and poor at the other. A pregnancy test with extremely high sensitivity will catch virtually every pregnancy but might also produce some false positives; a test tuned for maximum specificity will rarely falsely alarm but might miss some real cases.
When a diagnostic test produces results on a continuous scale rather than a simple yes-or-no, researchers use something called an ROC curve to visualize how sensitivity and specificity shift as you change the threshold for what counts as a positive result.10PubMed Central. Sensitivity, specificity, receiver-operating characteristic (ROC) curves and likelihood ratios: communicating the performance of diagnostic tests The area under that curve gives a single number summarizing overall test performance, with 1.0 being perfect and 0.5 being no better than flipping a coin. When that value exceeds about 0.80, it becomes practical to identify an optimal cutoff point, often the one that maximizes both sensitivity and specificity together. But the “optimal” threshold is not always the one that scores best on paper. In screening for a deadly cancer, for instance, clinicians might accept more false positives (lower specificity) to avoid missing any true cases (higher sensitivity).11PubMed Central. Receiver operating characteristic curve analysis in diagnostic accuracy studies: A guide to interpreting the area under the curve value
The pre-test probability of disease also matters enormously. A positive result on a highly accurate test means something very different depending on whether the patient had a 2% or a 60% chance of having the condition before the test was run. Bayesian reasoning uses the test’s likelihood ratios together with the clinical picture to update how probable a diagnosis is after the result comes in.12PubMed Central. Propagation of Uncertainty in Bayesian Diagnostic Test Interpretation When clinicians skip this step and treat a positive test as confirmation of disease without considering how likely the disease was in the first place, the result is unnecessary follow-up, patient anxiety, and sometimes harmful treatment.
Artificial Intelligence as a Diagnostic Partner
AI systems are increasingly being tested as tools to flag findings that human readers might overlook or deprioritize. One system trained on chest X-rays was able to detect normal radiographs with about 71% sensitivity and 95% specificity, enabling automated triage that routes urgent-appearing films to a radiologist faster.13PubMed Central. Automated Triaging of Adult Chest Radiographs with Deep Artificial Neural Networks The goal is not to replace the radiologist but to rearrange the queue so the most concerning images get human eyes sooner.
A more dramatic demonstration involved a deep learning tool deployed to detect incidental pulmonary embolism in cancer patients undergoing routine CT scans. After the AI tool was implemented, the detection rate of these blood clots tripled, from about 0.8% to 2.5% of studies. The median time from scan to report dropped from nearly 25 hours to under one hour, and the median time to treatment fell from about 28 hours to roughly one hour.14PubMed Central. Use of a Deep Learning Algorithm for Detection and Triage of Cancer-associated Incidental Pulmonary Embolism For cancer patients at elevated clotting risk, getting treatment in an hour instead of a day can be the difference between a manageable complication and a fatal one.
Large language models are also being tested as diagnostic aids. In a comparative analysis, advanced models achieved over 90% accuracy on common clinical scenarios, with one model reaching perfect accuracy on certain conditions. For complex cases, top performance was around 83% accuracy at the final diagnostic stage, though smaller models lagged well behind.15PubMed Central. Comparative analysis of large language models in clinical diagnosis: performance evaluation across common and complex medical cases Performance drops further in the territory of rare diseases, where even the best-performing model placed the correct diagnosis on its shortlist only about 22% of the time. Combining that model with a bioinformatics tool pushed the combined accuracy to 30%, better than either alone but still reflecting how far technology has to go for the hardest cases.16PubMed Central. Accuracy of Large Language Models in Generating Rare Disease Differential Diagnosis Using Key Clinical Features
Radiomics and the Image Beyond What Eyes Can See
Traditional radiology depends on a trained human recognizing visual patterns. Radiomics takes the same images and extracts hundreds or thousands of quantitative features, many invisible to the naked eye, that can be fed into predictive models. Research in oncology has shown that radiomic models built from standard imaging can improve the accuracy of tumor classification and treatment response assessment.17PubMed Central. The Applications of Radiomics in Precision Diagnosis and Treatment of Oncology: Opportunities and Challenges In breast cancer, for example, researchers have combined radiomic features extracted from imaging with deep learning features from transfer learning models, fusing them into a single classification framework designed to outperform either approach alone.18Diagnostics. Enhanced Breast Cancer Diagnosis Using Multimodal Feature Fusion with Radiomics and Transfer Learning
Liquid biopsies represent another frontier. Fragments of cell-free DNA circulating in the bloodstream carry genetic and epigenetic signatures of tumors, and detecting those signatures through a simple blood draw could eventually catch cancers before symptoms appear.19PubMed Central. Circulating cell-free DNA for cancer early detection Clinical trials are already exploring cell-free DNA for real-time monitoring of treatment resistance and disease progression, and early results suggest it can provide useful surveillance information while sparing patients repeated biopsies.20British Journal of Cancer. Cell-free DNA analysis in current cancer clinical trials: a review This is still an evolving field with questions about sensitivity for early-stage disease, but the trajectory is clear: the diagnostic act is moving from waiting for a lump to finding a molecular fingerprint in a tube of blood.
The Overdiagnosis Problem
More diagnostic power does not automatically translate into better outcomes. Overdiagnosis occurs when a condition is detected that would never have caused symptoms or death during the patient’s lifetime. Cancer screening is the most studied context. Estimates from randomized trials suggest that roughly a quarter of breast cancers found by mammography, about half of lung cancers found by chest X-ray or sputum analysis, and around 60% of prostate cancers detected through PSA testing represent overdiagnosis.21JNCI: Journal of the National Cancer Institute. Overdiagnosis in Cancer Each of those diagnoses can lead to surgery, radiation, or chemotherapy for a disease that was never going to be a threat.
Incidental findings compound the issue. A systematic review of imaging studies found that about 24% of diagnostic imaging tests turned up unexpected findings unrelated to the reason the scan was ordered. CT scans had a higher rate, averaging around 31%. About two-thirds of incidental findings prompted clinical follow-up, but fewer than half were ultimately confirmed as clinically meaningful.22Oxford Academic (British Journal of Radiology). Incidental findings in imaging diagnostic tests: a systematic review This cascade of follow-up creates anxiety, cost, and occasionally physical harm from additional procedures, all for a finding that turned up by accident and may have been better left undiscovered.
Diagnostic Stewardship and Smarter Ordering
The recognition that more testing is not always better has fueled interest in diagnostic stewardship, essentially applying the same restraint to diagnostic tests that antibiotic stewardship applies to prescribing. One common target is unnecessary urine cultures, which often detect bacteria that are not causing infection but prompt antibiotic prescriptions anyway. A clinical decision support tool implemented across one healthcare system cut reflex-to-culture urine testing by about 63 per 1,000 patient-days, though the effect was not fully sustained over time, and antibiotics were still frequently prescribed without documented symptoms of infection.23PubMed Central. Indication-based clinical decision support for inpatient urine testing: impacts on utilization, appropriateness, and antibiotic prescribing
A separate study in a healthcare network that already used conditional urine testing found that adding an electronic health record-based stewardship tool further reduced total urine culture orders by about 5 cultures per 1,000 patient-days per month.24Antimicrobial Stewardship & Healthcare Epidemiology. Moving Beyond the Reflex: Effect of a Clinical Decision Support Tool on Urine Culture Ordering Practices The lesson from both studies is that decision support can nudge ordering behavior in the right direction, but technology alone does not fix the problem. Ordering habits, clinical culture, and education all have to move together.
What Diagnostic Uncertainty Does to Patients
The diagnostic process does not happen in a vacuum. For patients, the period between “something might be wrong” and “here is what it is” can be psychologically brutal. Research on women undergoing evaluation for suspected breast cancer found that both uncertainty and anxiety were significantly higher before a diagnosis was reached compared to after, regardless of whether the final diagnosis was benign or malignant.25Cancer Nursing. Uncertainty and Anxiety During the Diagnostic Period for Women With Suspected Breast Cancer In some sense, knowing you have cancer can be less distressing than not knowing whether you do.
Patients living with unresolved diagnoses experience what researchers describe as intolerance of uncertainty, a pattern of worry, fear, and sometimes despair that compounds over time.26PubMed Central. How Patients with an Uncertain Diagnosis Experience Intolerance of Uncertainty: A Grounded Theory Study The question of how much diagnostic uncertainty to communicate to patients is itself an active ethical discussion. Some worry is adaptive; it can motivate follow-up visits and healthy behavior changes. But worry about outcomes that a patient cannot control and that may never materialize can do real psychological harm. Growing evidence suggests most patients prefer transparent communication about diagnostic uncertainty, even when it is unsettling, over being kept in the dark.27PubMed Central. In Defence of Causing Patients to Worry: Ethical Issues in the Communication of Diagnostic Uncertainty
Portable Diagnostics and the Global Gap
Much of the diagnostic technology discussed above requires expensive laboratory infrastructure, trained personnel, and stable supply chains. In low-resource settings, those prerequisites are often absent. Point-of-care testing aims to close that gap by miniaturizing diagnostic processes onto portable devices that can be used at a patient’s bedside or in a field clinic. Microfluidic chip technologies have shrunk PCR-based molecular testing into handheld formats with potential gains in speed, cost, and portability.28PubMed Central. Advances in microfluidic PCR for point-of-care infectious disease diagnostics During outbreaks, these “frugal innovations” can reduce the wait for results and enable containment measures that depend on rapid identification of infected individuals.29PubMed. Frugal Innovation for Point-of-Care Diagnostics Controlling Outbreaks and Epidemics
The challenge is not just building these devices but validating them against laboratory standards and integrating them into healthcare systems that may lack the infrastructure to act on their results. A rapid test that identifies drug-resistant tuberculosis is only useful if the patient can access the appropriate second-line treatment afterward. The diagnostic act, in other words, is never just about the test. It is about the entire chain from suspicion through confirmation to effective action.
Ethical Tensions in Pediatric Genetic Testing
Nowhere are the ethical dimensions of diagnostic capability more visible than in pediatric and neonatal genetics. Rapid genomic sequencing in neonatal intensive care units can identify the cause of a critically ill newborn’s condition in days rather than months. But the introduction of such a powerful tool into an emotionally charged environment raises difficult questions. Getting meaningful informed consent from parents who are frightened and overwhelmed is not straightforward. A rapid diagnosis can alter how parents bond with their child, particularly if the finding suggests a severe prognosis, and that shift in bonding can ripple through decisions about whether to continue intensive treatment.30Pediatrics. Rapid Challenges: Ethics and Genomic Neonatal Intensive Care
Beyond the NICU, broader genetic testing in children raises questions about what to do with incidental findings, results that reveal predispositions or conditions the family was not looking for. Should parents be told that their child carries a gene variant associated with adult-onset disease? The child cannot consent to knowing this information, and parents may make decisions based on it that the child, as an adult, might not have wanted.31PubMed Central. Ethical issues in pediatric genetic testing and screening These are not hypothetical dilemmas. As next-generation sequencing becomes more accessible and clinicians grapple with how to interpret and communicate its results, the gap between what can be diagnosed and what can be meaningfully acted upon continues to widen.32PubMed Central. Challenges in the clinical understanding of genetic testing in birth defects and pediatric diseases