The Science of Diagnostic Performance

In the practice of medicine, every decision pivots on a crucial element: information. Diagnostic tests are our primary tools for gathering this information, turning suspicion into certainty and guiding treatment. However, the value of a test is not inherent; it must be rigorously measured and understood. The challenge lies in navigating the complex statistics and potential biases that can either illuminate the path to a correct diagnosis or lead us astray. This article addresses this fundamental challenge by providing a clear framework for evaluating diagnostic performance. The journey begins in the first section, "Principles and Mechanisms," where we will dissect the core metrics of accuracy, including sensitivity, specificity, and the powerful ROC curve, while also confronting the common pitfalls and biases that can distort our results. Following this, the second section, "Applications and Interdisciplinary Connections," will demonstrate how these foundational concepts are applied in the real world, from the learning curve of a clinician and the synthesis of multiple clues to the strategic deployment of tests and the evaluation of cutting-edge AI technologies.

Imagine you are a doctor. A patient comes to you with a set of symptoms, and you have a suspicion—a hypothesis—about what might be wrong. How do you move from a suspicion to a confident diagnosis? You gather more information. You ask questions, you perform a physical exam, and often, you order a test. A diagnostic test is, at its heart, a tool for reducing uncertainty. It is an experiment you perform on a single person to help you decide between two or more competing stories about their health. But how do we know if the tool is any good? How do we measure its performance? This is not just an academic question; it is the foundation upon which every medical decision rests. To understand this, we must start from first principles.

Let's imagine we have a new test, perhaps for a specific type of cancer. We apply it to a group of people, and for each person, we also have a perfect, god-like way of knowing whether they truly have the cancer or not. This perfect knowledge is what we call the ​​reference standard​​ or ​​gold standard​​. In the real world, this "gold standard" is often the result of a biopsy and histopathology—a direct look at the tissue in question.

With our test result and the "truth" in hand for every person, we can draw a simple, powerful box with four compartments. This is the famous ​​2x2 contingency table​​, the fundamental ledger for all of diagnostic medicine.

	​​Truly Has Disease​​	​​Truly Disease-Free​​
​​Test Positive​​	True Positive (TP)	False Positive (FP)
​​Test Negative​​	False Negative (FN)	True Negative (TN)

• ​​True Positives (TP):​​ The test correctly identifies someone who has the disease. A success.
• ​​True Negatives (TN):​​ The test correctly clears someone who is healthy. Another success.
• ​​False Positives (FP):​​ The test raises a false alarm, flagging a healthy person. This is a Type I error.
• ​​False Negatives (FN):​​ The test misses the disease, giving a false reassurance. This is a Type II error, and often the most dangerous kind.

Every single measure of a test's performance flows directly from the numbers in these four boxes.

Two of the most important properties of a test are its ​​sensitivity​​ and ​​specificity​​. These are often called the test's intrinsic characteristics because, under ideal conditions, they tell us about the test itself, independent of how common or rare the disease is in the population.

​​Sensitivity​​ answers the question: If a person has the disease, what is the probability that the test will be positive? It is the test's ability to detect what it is looking for.

Imagine a study of a sentinel lymph node mapping test for endometrial cancer found 70 true positives and 2 false negatives. The total number of people with cancer was $70 + 2 = 72$. The sensitivity would be $\frac{70}{72} \approx 0.9722$. This means the test successfully "catches" about $97.2\%$ of the cancers it encounters. The $2.8\%$ it misses is called the ​​False Negative Rate (FNR)​​, which is simply $1 - \text{Sensitivity}$.

​​Specificity​​ answers the complementary question: If a person does not have the disease, what is the probability that the test will be negative? It is the test's ability to correctly ignore those who are healthy.

In that same cancer study, there were 123 true negatives and 5 false positives. The total number of people without cancer was $123 + 5 = 128$. The specificity would be $\frac{123}{128} \approx 0.9609$. This means the test correctly gives an "all-clear" to about $96.1\%$ of cancer-free individuals. The remaining $3.9\%$ who get a false alarm represent the ​​False Positive Rate (FPR)​​, which is $1 - \text{Specificity}$.

It's crucial to understand that sensitivity and specificity are defined as probabilities conditional on the true disease status. They don't depend on the disease prevalence. A test with a sensitivity of $97\%$ should, in principle, detect $97\%$ of diseased individuals whether it's used in a high-risk clinic or for mass screening.

But where do these numbers come from? Many tests, especially those based on biomarkers, don't just say "yes" or "no." They return a continuous value—a concentration, a score, a measurement. The physician must then choose a ​​threshold​​ or ​​cut-off​​ to decide what counts as "positive". And this is where things get interesting.

Imagine a security guard trying to tell authorized personnel from intruders based on how fast they walk. If the guard sets a very low speed limit as the "intruder" threshold (a low threshold), they will catch nearly every actual intruder (high sensitivity), but they will also flag many fast-walking employees (low specificity). If they set the speed limit very high, they will almost never bother an employee (high specificity), but they will miss all but the fastest-running intruders (low sensitivity).

This is a fundamental trade-off. You can't change one without affecting the other. By sliding the threshold up and down, you can generate an entire family of sensitivity and specificity pairs. If we plot these pairs on a graph—with Sensitivity (True Positive Rate) on the y-axis and $1 - \text{Specificity}$ (False Positive Rate) on the x-axis—we trace out a beautiful arc called the ​​Receiver Operating Characteristic (ROC) curve​​.

The ROC curve is the test's complete performance resume. It shows you every possible trade-off it can offer. A test that is no better than a coin flip will have an ROC curve that is just a diagonal line from $(0,0)$ to $(1,1)$. A perfect test would shoot straight up from $(0,0)$ to $(0,1)$ and then across to $(1,1)$, hugging the top-left corner.

The total area under this curve, the ​​Area Under the Curve (AUC)​​, gives us a single, elegant summary of the test's overall discriminative power. The AUC has a wonderfully intuitive meaning: it is the probability that a randomly chosen diseased individual will have a higher test result than a randomly chosen non-diseased individual. An AUC of $0.5$ is a coin flip; an AUC of $1.0$ is perfection. A typical good diagnostic biomarker might have an AUC around $0.85$ or higher, like the one calculated in a hypothetical example which yielded an AUC of $0.8647$.

Sensitivity and specificity are the test's properties, but they don't directly answer the question a patient (or their doctor) is asking. A patient doesn't ask, "Given that I have cancer, what's the chance my test is positive?" They ask, "Given that my test is positive, what's the chance I have cancer?"

This question is answered by the ​​Positive Predictive Value (PPV)​​.

Similarly, the ​​Negative Predictive Value (NPV)​​ answers the question for a negative test result: "Given that my test is negative, what's the chance I am disease-free?"

Using our cancer test example, the PPV is $\frac{70}{70+5} \approx 0.9333$, and the NPV is $\frac{123}{123+2} \approx 0.9840$. These look great! But there's a hidden catch. Unlike sensitivity and specificity, ​​predictive values are critically dependent on the prevalence of the disease​​—how common it is in the population being tested.

Let's do a thought experiment. Take that same excellent test (Sens $0.9722$, Spec $0.9609$) and use it to screen $100,000$ people in the general population where the cancer is rare, say $1$ in $1,000$ (prevalence = $0.001$).

• We expect $100$ people to have cancer. The test will find about $97$ of them (TP) and miss $3$ (FN).
• We expect $99,900$ people to be cancer-free. The test will falsely flag about $3.9\%$ of them, which is $3,896$ people (FP).
• The total number of positive tests is $97 + 3896 = 3993$.
• Now, what is the PPV? It's $\frac{TP}{\text{all positives}} = \frac{97}{3993} \approx 0.024$.

Think about that. For a person who gets a positive result, the chance they actually have cancer is only $2.4\%$. The vast majority of positive results are false alarms. This is not a flaw in the test; it is an inexorable consequence of applying a good test to a low-prevalence population. This single concept explains why mass screening for rare diseases is so fraught with challenges.

So far, we have assumed we have a perfect "gold standard" to judge our test against. But in the real world, establishing the truth is often the hardest part. The pursuit of diagnostic accuracy is filled with subtle traps and biases that can lead us to believe a test is much better than it really is.

The Problem with the Gold Standard

What is the "truth" for acute appendicitis? A surgeon's visual inspection? A pathologist's report on the removed appendix? A panel of experts reviewing all the data? As described in guidelines for evaluating diagnostic AI, defining the reference standard is a monumental task. To be valid, it must be defined with pre-specified criteria, and the people applying it must be ​​blinded​​—they cannot know the result of the test being evaluated. If they know the test was positive, they might look harder for the disease, a phenomenon called ​​incorporation bias​​, which creates a vicious cycle that artificially inflates the test's apparent accuracy.

The Trap of Verification Bias

Another trap is ​​verification bias​​. Imagine a new, non-invasive imaging test (like a Cardiac MRI) for myocarditis, a condition where the gold standard is an invasive heart biopsy (Endomyocardial Biopsy). It's only natural that doctors are more likely to send patients with a "positive" and concerning MRI result for the risky biopsy than patients with a clean, "negative" MRI.

This creates a biased sample of "verified" cases. The test-positive group is heavily scrutinized, while the test-negative group is largely ignored. When you calculate sensitivity from only the verified patients, you've over-sampled the true positives and under-sampled the false negatives, making the test's sensitivity appear much higher than it really is. Fortunately, if we know the rates of verification, we can use statistical methods like inverse probability weighting to correct for this, re-balancing the scales to estimate the test's true performance.

The Spectrum of Disease

Furthermore, a test's performance can change depending on the ​​patient spectrum​​. A test validated on patients with advanced, florid disease versus perfectly healthy volunteers might show spectacular results. But when that same test is deployed in a primary care clinic to detect early, subtle disease among a crowd of patients with other confounding conditions, its performance can drop dramatically. A good diagnostic study must enroll a representative spectrum of patients that reflects the test's intended clinical use.

Science progresses by accumulating evidence. A single study is never the final word. To get a comprehensive picture, researchers perform a ​​meta-analysis​​, a statistical method for combining the results of many studies. However, this is also full of peril.

When we see a meta-analysis that reports a high "pooled" sensitivity, we must be skeptical. If there is high ​​heterogeneity​​ (measured by statistics like $I^2$), it means the individual studies are telling very different stories. This might be because they used different thresholds (the threshold effect), studied different patient populations (the spectrum effect), or had different biases. A single average number can be profoundly misleading in this situation; it's like reporting the average weather for the entire planet. Advanced methods like modeling a ​​Summary ROC (SROC) curve​​ can help, as they try to summarize the trade-off across studies rather than just one number.

Even more worrying is ​​publication bias​​. Studies that find exciting, positive results for a new test are much more likely to be published than boring studies that find the test doesn't work well. The published literature can therefore present an overly rosy view of a test's performance. Statistical tools can look for this asymmetry, but the bias is hard to eliminate.

We've journeyed through a complex landscape, from the simple 2x2 table to the minefield of meta-analysis. But we must take one final step. Even a perfectly accurate test can be useless or even harmful if it doesn't lead to better outcomes. This brings us to the crucial distinction between clinical validity and clinical utility.

• ​​Analytical Validity:​​ Does the test measure what it claims to measure reliably and accurately in the lab?
• ​​Clinical Validity:​​ Does the test correlate well with the presence or absence of the disease? (This is where Sensitivity, Specificity, and AUC live).
• ​​Clinical Utility:​​ Does using the test in practice actually help patients? Does it lead to better treatment decisions, improved survival, lower costs, or better quality of life?

Proving clinical utility is the highest bar. It often requires large, expensive randomized controlled trials that compare a strategy with the new test against a strategy without it. A test might be accurate but lead to the over-diagnosis of harmless conditions, causing patient anxiety and unnecessary treatments. Or it might be accurate but have no effective treatment to offer.

The journey to understanding a diagnostic test's performance is a microcosm of the scientific method itself. It begins with simple classification, delves into the mathematics of probability and trade-offs, confronts the messy reality of bias and human factors, and ultimately must answer the most human question of all: Does this actually make a difference?

Having journeyed through the foundational principles of diagnostic performance, we might be tempted to view concepts like sensitivity, specificity, and predictive value as mere abstract calculations—tools for passing an exam and then promptly forgetting. But nothing could be further from the truth. These ideas are not sterile academic constructs; they are the very grammar of medical reasoning, the universal language that allows us to navigate the fog of clinical uncertainty. They form the bridge between a subtle shadow on an X-ray and a life-saving intervention, between a faint signal in a blood sample and a family's future.

In this chapter, we will see how these fundamental principles blossom into a rich tapestry of applications, connecting the quiet bedside observation to the bustling laboratory, the surgeon's calculated risk to the AI developer's algorithm, and the pathologist's slide to the epidemiologist's population-wide policies. We will discover that understanding diagnostic performance is not just about knowing the odds; it is about learning to think wisely.

Every medical professional begins as a novice, overwhelmed by a torrent of information. With time and practice, a remarkable transformation occurs: they develop a finely-tuned intuition, an almost subconscious ability to weigh evidence and spot patterns. What is this "intuition"? In large part, it is an internalized, experience-driven model of diagnostic performance. A seasoned clinician has, through exposure to thousands of cases, developed a gut feeling for the sensitivity and specificity of different signs and symptoms.

We can even model this process mathematically. Imagine a learning curve where diagnostic accuracy improves with both years of experience ($E$) and the volume of cases seen ($V$). We could propose that sensitivity, $Se$, doesn't just jump to a fixed value, but grows from a minimum baseline ($Se_{\min}$) towards a maximum potential ($Se_{\max}$) as a clinician learns. A simple and elegant model for this growth is the negative exponential curve:

$Se(E, V) = Se_{\min} + (Se_{\max} - Se_{\min}) \cdot (1 - \exp(-(\alpha E + \beta V)))$

Here, $\alpha$ and $\beta$ are learning rates that quantify how quickly experience and case volume translate into better performance. A similar equation can be written for specificity. This model shows us that becoming an expert diagnostician is an asymptotic journey towards a peak of performance, a journey that can be quantified and understood.

Of course, we don't have to leave this learning to chance. Formal training programs are designed to accelerate this journey. Consider a clinic where dermatologists are being trained to recognize a specific skin condition. Before the training, their clinical eye might have a sensitivity of $0.70$ and a specificity of $0.80$. After a structured program, these could improve to $0.85$ and $0.90$, respectively. By applying the accuracy formula, $Acc = (\pi \cdot Se) + ((1 - \pi) \cdot Sp)$, where $\pi$ is the disease prevalence, we can precisely calculate the absolute improvement in the proportion of patients correctly classified. In a plausible scenario, this seemingly modest boost in sensitivity and specificity could lead to an $11\%$ absolute improvement in overall accuracy—meaning 110 more patients out of every 1000 receive a correct diagnosis, all thanks to a targeted educational intervention.

A detective rarely solves a case with a single clue, and a clinician rarely makes a diagnosis with a single test. The art of diagnosis lies in synthesis—in weaving together multiple, often imperfect, pieces of evidence into a coherent and compelling conclusion. This process, which feels like an intuitive leap, has a beautiful mathematical foundation in Bayesian reasoning.

When multiple diagnostic features are conditionally independent (meaning the presence of one doesn't affect the probability of another, given the disease is present or absent), their diagnostic power doesn't simply add up; it multiplies. We update our belief not by adding probabilities, but by multiplying likelihood ratios.

Imagine a patient with a suspected autoimmune disease like dermatomyositis. They present with a constellation of signs: a characteristic violet-hued rash in a "shawl" distribution, subtle changes in the tiny blood vessels of their nailfolds, and a specific pattern of inflammation on a skin biopsy. Individually, none of these clues is definitive. The rash might have a positive likelihood ratio ($LR(+)$) of about $4.7$, the nailfold findings an $LR(+)$ of $6.0$, and the biopsy an $LR(+)$ of $3.75$. If our initial suspicion (pre-test probability) was $0.25$, the pre-test odds are $0.25/0.75 = 1/3$. Now, watch the magic. The combined likelihood ratio is the product: $4.7 \times 6.0 \times 3.75 \approx 105$. Our post-test odds become the pre-test odds multiplied by this powerful factor: $(1/3) \times 105 = 35$. The odds are now 35-to-1 in favor of the diagnosis. This translates to a post-test probability of $35/(35+1) \approx 0.97$. Our confidence has skyrocketed from a 25% suspicion to a 97% certainty. This is the mathematical soul of clinical reasoning—a formal description of how a "classic presentation" emerges from the synergy of multiple clues.

We often talk about sensitivity and specificity as if they are fixed, immutable properties of a test. But a diagnostic test does not operate in a vacuum. It operates within the complex, dynamic, and sometimes confounding ecosystem of the human body. The patient's own biological state can fundamentally alter the performance of a test, a crucial insight that separates the novice from the expert.

Consider the diagnosis of tuberculous pericarditis, a serious infection around the heart. A useful biomarker is Adenosine Deaminase (ADA), an enzyme released by activated T-lymphocytes, the soldiers of our cell-mediated immune system. In an otherwise healthy person, a tuberculosis infection triggers a robust T-cell response, flooding the pericardial fluid with ADA. A high ADA level is therefore a sensitive marker for the disease.

But what happens if the patient is also co-infected with HIV, especially in its advanced stages? HIV decimates T-lymphocytes. The patient's immune system can no longer mount a strong response to the tuberculosis bacteria. Even with an active infection, there are fewer T-cells to produce ADA. The result? The sensitivity of the ADA test plummets. A level that would be reassuringly low in an immunocompetent patient might be a dangerous false negative in a patient with HIV. In this context, a low ADA value cannot be used to rule out the disease. This powerful example from immunology and infectious disease teaches us that diagnostic metrics are not absolute truths; they are conditional probabilities that depend critically on the host's underlying pathophysiology.

A wise diagnostician knows that the goal is not merely to be correct, but to be useful. The diagnostic process is a series of strategic decisions aimed at maximizing benefit and minimizing harm for the patient. This involves choosing not only the right test to interpret, but the right test to order and the right way to obtain a sample in the first place.

Imagine two patients with a suspected sigmoid volvulus, a life-threatening twisting of the colon. One patient is stable, with mild pain. The other is unstable—tachycardic, hypotensive, and showing signs of peritonitis, suggesting the bowel may be gangrenous. We have two imaging options: a CT scan or a contrast enema. Which is better? The answer depends entirely on the context. For the stable patient, a CT scan is superior. It not only confirms the diagnosis with high accuracy but, crucially, can assess for signs of ischemia (lack of blood flow), which dictates the next steps. For the unstable patient, however, the "best" test is no test at all. The clinical signs already scream "surgical emergency!" Taking the time to perform a CT scan would be a dangerous, potentially fatal delay. Furthermore, a contrast enema would be absolutely contraindicated due to the high risk of perforating the compromised bowel. The guiding principle here is not diagnostic accuracy in isolation, but clinical utility in a dynamic, high-stakes environment.

This strategic thinking extends all the way to the initial step of obtaining a tissue sample. Consider a child with suspected Langerhans cell histiocytosis (LCH), a complex disease that can affect multiple organ systems. Imaging reveals suspicious lesions in the skin, bone, liver, and brain (pituitary stalk). Where should we perform the biopsy to confirm the diagnosis? We must weigh the probability of getting a diagnostic sample against the procedural risk. A biopsy of the pituitary stalk, while likely to be diagnostic, is an incredibly high-risk neurosurgical procedure. A liver biopsy is also risky, especially in a child with a bleeding tendency. A bone marrow biopsy is safer, but LCH involvement is often patchy, so the diagnostic yield is low. The clear winner is a simple punch biopsy of a skin lesion. It is highly accessible, carries very low risk even with a bleeding disorder, and has a high probability of containing the diagnostic cells. The optimal diagnostic strategy is not about aiming for the most "interesting" lesion, but about maximizing a conceptual ratio of $\frac{\text{Yield} \times \text{Accessibility}}{\text{Risk}}$.

Where do diagnostic tests come from? They are the end product of a long and rigorous journey that spans laboratory science, biomarker discovery, clinical validation, and regulatory oversight. Our core concepts of performance are the guiding stars at every stage of this lifecycle.

The journey begins in the laboratory, where the quality of a test is forged. Consider a modern molecular test like RT-qPCR, used to detect viral RNA. Its ultimate performance—its Limit of Detection (LOD), or the smallest amount of virus it can reliably find—doesn't just depend on the final chemical reaction. It is built on a foundation of pre-analytical quality. The integrity of the RNA extracted from a patient's blood is paramount. If the RNA is degraded (a low RNA Integrity Number, or RIN), the test will fail. A rigorous validation plan doesn't just test the final assay under ideal conditions; it "stresses" the system by intentionally using samples with varying quality (e.g., a range of RIN values) and models how performance metrics like LOD degrade as sample quality declines. This allows the lab to set rational quality control-based acceptance criteria, ensuring that a reported result is trustworthy.

Many tests rely on biomarkers—molecules in the blood or tissue whose levels correlate with disease. The story of CA-125 and HE4 for ovarian cancer provides a masterclass in biomarker utility. CA-125 was an early hope, but it suffers from poor specificity; many benign conditions like endometriosis can cause it to be elevated, leading to false positives. A newer marker, HE4, has better specificity but has its own blind spots (e.g., its levels can be falsely elevated in kidney disease, and it is not sensitive for all subtypes of ovarian cancer). The solution? Don't rely on one marker. Algorithms like ROMA combine CA-125, HE4, and the patient's menopausal status to achieve better diagnostic discrimination than any single marker. Yet, even this sophisticated tool is not used for general population screening. Why? Because ovarian cancer is rare in the general population (low prevalence). As we know, when prevalence is very low, even a highly specific test will have a low Positive Predictive Value (PPV), leading to an unacceptably high number of false positives who would undergo unnecessary, anxious, and invasive follow-up procedures. The markers are therefore reserved for triaging patients who are already at high risk (e.g., those with a pelvic mass on ultrasound), where the pre-test probability is much higher.

The newest and most exciting class of diagnostic tools comes from the world of Artificial Intelligence. These complex algorithms promise to revolutionize medicine, but they must be evaluated with the same, if not greater, rigor as any traditional test.

A crucial distinction, enshrined in regulatory standards like ISO 14971, is the difference between performance and benefit. An AI tool for detecting pneumothorax on chest radiographs may boast stunning analytical performance on a curated dataset (e.g., an Area Under the Curve of 0.94) and excellent clinical performance in a trial (e.g., sensitivity of 0.96 and specificity of 0.85). But these are means to an end. The true measure of the tool is its benefit to patients. Does using the tool actually lead to better health outcomes? In one hypothetical scenario, deploying such a tool led to a decrease in the median time-to-treatment from 75 to 55 minutes and a reduction in the rate of serious complications from 4.0% to 3.2%. This is the benefit: eight fewer patients out of every thousand suffering a major complication. This focus on patient outcomes is the ultimate arbiter of a new technology's value.

The sophistication of these tools demands an equal sophistication in how we study them. The scientific questions we ask about AI in medicine are diverse, and each requires a distinct study design and reporting standard.

1. If we are developing a model to predict a future event (e.g., a risk score for sepsis), our goal is to estimate a conditional probability, $P(Y \mid X)$. The key challenges are avoiding overfitting and validating the model's performance in new patients. The TRIPOD reporting guideline is designed for this.
2. If we are evaluating the impact of using an AI tool (e.g., a triage system), we are asking a causal question. We want to know the effect of the intervention, $E[Y(1) - Y(0)]$. The gold standard is a randomized controlled trial, and the CONSORT-AI guideline ensures we report the details needed to make a valid causal claim.
3. If we are assessing the raw accuracy of an AI classifier against a reference standard (e.g., a tuberculosis detector), we are conducting a classic diagnostic accuracy study. Our goal is to measure sensitivity and specificity, and the STARD-AI guideline helps us report the study in a way that mitigates common biases. The fact that we need these distinct frameworks highlights the intellectual depth of modern clinical evidence generation. One size does not fit all, because prediction, causation, and classification are fundamentally different scientific endeavors.

From the learning curve of a single physician to the regulatory framework for global AI, we see the same fundamental concepts at play. The simple, elegant definitions of sensitivity and specificity are like the foundational notes of a grand symphony. They provide the language for quantifying improvement, the logic for synthesizing evidence, and the framework for navigating the complexities of human biology. They remind us that the utility of a test is inseparable from the clinical context and the prevalence of disease. They guide our strategy, ensuring we balance the quest for information against the mandate to "first, do no harm."

Ultimately, all these applications point to a single, profound truth. Improving our diagnostic performance, whether through training, technology, or better science, is not an academic exercise. It translates directly into human terms: a reduction in the risk of misdiagnosis. In one study of a protocol to better identify psychogenic seizures, improving diagnostic accuracy from 75% to 90% resulted in a 15% absolute risk reduction in misdiagnosis. This is the bottom line. Behind every ROC curve and every likelihood ratio lies the potential to reduce harm, alleviate suffering, and guide a patient safely through their moment of uncertainty. That is the inherent beauty and the ultimate purpose of understanding the science of diagnosis.