Diagnostic Accuracy Studies: Sensitivity, Specificity, and Beyond

ABSTRACT

Diagnostic tests are central to medical care, and their evaluation remains methodologically challenging. This methodological article outlines the key measures used to quantify diagnostic accuracy, highlights common sources of bias and variability in diagnostic accuracy studies, and discusses how the clinical effectiveness of medical tests can be evaluated. Examples of diagnostic tests from various fields, including clinical, radiographic, biomarker, and emerging AI‐based approaches, are discussed. We argue that diagnostic accuracy represents only the first step in test evaluation and must be complemented by assessments of feasibility, clinical impact, and patient‐centered outcomes.

This methodological article covers key diagnostic accuracy measures, common pitfalls in diagnostic accuracy studies, and how evidence should progress from diagnostic accuracy to clinical effectiveness, with examples from various fields.

Abbreviations

AI

artificial intelligence

BOP

bleeding on probing

CBCT

cone beam computed tomography

GBOB

gingival bleeding on brushing

GCF

gingival crevicular fluid

LR

likelihood ratio

NPV

negative predictive value

PPV

positive and negative predictive value

1. Introduction

Medical tests are indispensable tools in healthcare, guiding diagnostic, prognostic, and therapeutic decisions. They range from clinical assessments and imaging techniques to biomarker assays, and rigorous evaluation is essential before clinical implementation. A new diagnostic test must be assessed to determine its accuracy in distinguishing patients with or without a target condition. Diagnostic accuracy is usually quantified through measures such as sensitivity and specificity. Yet, diagnostic accuracy alone does not capture the full benefit‐harm balance of a test. The consequences of a test on clinical decision‐making, patient outcomes, and healthcare resource allocation must also be considered.

While diagnostic tests aim to determine the current presence or absence of disease, prognostic tests aim to predict the occurrence of a future event or disease state. Although they differ in time horizon, the evaluation of diagnostic and prognostic tests relies on the same core methodological principles and often uses similar performance measures (e.g., sensitivity, specificity, predictive values).

In this methodological article, we will focus on diagnostic tests and then present and discuss examples from the literature. Some examples are from periodontal health, but the methodological principles and considerations apply more broadly to other medical fields as well. This methodological article is structured as follows: (i) core definitions; (ii) measures of diagnostic accuracy; (iii) sources of bias and variability; (iv) advanced diagnostic accuracy designs; and (v) beyond diagnostic accuracy: the effectiveness of medical tests.

2. Core Definitions

A diagnostic accuracy study aims to quantify how well an index test (i.e., the diagnostic test under evaluation) can distinguish patients with the target condition from those without it. The target condition is the specific health condition of interest for the index test. Typically, diagnostic accuracy studies use a cross‐sectional design, enrolling a group of patients suspected of having the target condition, and all participants undergo both the index test and a reference standard, which represents the best available clinical method for determining disease status (Table 1) [1, 2].

TABLE 1.

Glossary of terms used in the methodological article.

Term	Definition
Index test	Diagnostic test under evaluation.
Reference standard	Best available clinical method for determining disease status.
Target condition	Particular health condition of interest.
Contingency table	2 × 2 table in which the results of the index test are cross‐tabulated with those of the reference standard.

In these studies, the results of the index test are cross‐classified against the reference standard into one of four categories: true positives (patients correctly identified as having the target condition), false positives (patients incorrectly identified as having the target condition), true negatives (patients correctly identified as not having the target condition), and false negatives (patients incorrectly identified as not having the target condition). These numbers can be presented in a contingency table and form the basis for computing diagnostic accuracy measures (Table 2).

TABLE 2.

Illustrative contingency table and associated diagnostic accuracy measures.

		Reference standard
		Positive	Negative	Total
Index test	Positive	32	134	166
	Negative	8	226	234
	Total	40	360	400

Note: diagnostic accuracy measures: sensitivity = 32/40 = 80%, specificity = 226/360 = 63%, positive predictive value = 32/166 = 19%, negative predictive value = 226/234 = 97%, positive likelihood ratio = 2.1, negative likelihood ratio = 0.32.

Diagnostic accuracy studies can vary significantly in design. In a single‐gate design, all participants are recruited using a single set of eligibility criteria, whereas in multiple‐gate designs, different participant groups are recruited using different eligibility criteria. Single‐gate designs are more likely to yield a study population that reflects the intended clinical use of the test and the full spectrum of disease encountered in practice [3]. Two‐gate designs may be relevant in the early phases of test evaluation following biomarker discovery, whereas single‐gate designs are more appropriate in later stages aimed at clinical implementation.

3. Measures of Diagnostic Accuracy

Different accuracy measures can be computed from a contingency table (Table 3). Here we present the most common measures of diagnostic accuracy and discuss their interpretation.

TABLE 3.

Diagnostic accuracy measures.

Diagnostic accuracy measure	Definition	Formula	Typical interpretation a
Population level
Sensitivity	Proportion of patients with positive index test results among those with the target condition (the “diseased”).	Sens = TP/(TP + FN)	High values (e.g., > 90%) are expected for screening or first‐line tests.
Specificity	Proportion of patients with negative index test results among those without the target condition (the “non‐diseased”).	Spec = TN/(TN + FP)	High values (e.g., > 90%) are expected for confirmatory or second‐line tests.
Patient level
Positive predictive value	Probability that a patient with a positive index test result truly has the target condition.	PPV = TP/(TP + FP)	High values (e.g., > 90%) are expected for rule‐in tests.
Negative predictive value	Probability that a patient with a negative index test result does not have the target condition.	NPV = TN/(TN + FN)	High values (e.g., > 90%) are expected for rule‐out tests.
Decision‐making level
Positive likelihood ratio	Ratio between the proportion of positive index tests among the diseased and the proportion of positive index tests among the non‐diseased.	LR + = [TP/(TP + FN)]/[FP/(TN + FP)] = Sens/(1‐Spec)	It shows how much more likely a positive test result is in the diseased than in the non‐diseased. High values (e.g., > 10) are expected for rule‐in tests.
Negative likelihood ratio	Ratio between the proportion of negative index tests among the diseased and the proportion of negative index tests among the non‐diseased.	LR‐ = [FN/(TP + FN)]/[TN/(TN + FP)] = (1‐Sens)/Spec	It shows how much less likely a negative test result is in the diseased than in the non‐diseased. Low values (e.g., < 0.1) are expected for rule‐out tests.
Overall diagnostic accuracy
Area under the ROC curve	The ROC curve plots Sensitivity against (1‐Specificity) across all possible thresholds of a test. The area under the curve (AUC) summarizes overall discrimination between diseased and non‐diseased patients.	AUC = ∫ Sensitivity d(1 − Specificity) The AUC is estimated by integrating the ROC curve. In practice, when the ROC curve is drawn from observed data, the AUC is often estimated by summing trapezoidal areas under the curve.	0.50: no discrimination 0.50–0.70: low accuracy 0.70–0.90: accuracy that may be useful for some purposes > 0.90: high accuracy

^a The values presented here are given for illustrative purposes and may serve as rules of thumb, but interpretation depends on the clinical context.

3.1. Sensitivity and Specificity

Sensitivity quantifies the proportion of patients with the target condition who receive a positive result on the index test. It reflects the test's ability to correctly identify those with the target condition. Conversely, specificity measures the proportion of patients without the condition who receive a negative index test result, indicating the test's ability to correctly identify those without the target condition. Both measures are expressed as percentages, calculated from the frequencies observed in the contingency table. The uncertainty around point estimates is usually conveyed by 95% confidence intervals.

High sensitivity is particularly valuable in conditions where missing a diagnosis could have serious consequences and where early detection and intervention are essential. High specificity is desirable when false‐positive results may lead to unnecessary or invasive diagnostic procedures and potentially harmful treatments. The desirable properties of diagnostic accuracy measures may vary depending on the context in which a test is used. For example, screening tests typically aim for very high sensitivity to minimize false negatives, whereas higher specificity is generally more important for confirmatory diagnostic tests.

For example, self‐reported gingival bleeding on tooth brushing (GBOB) has been evaluated as a diagnostic test for the identification of periodontitis and gingivitis. Deng et al. reported a sensitivity of 37%–61% and a specificity of 84%–85% for discriminating periodontitis or gingivitis from periodontal health [4]. These figures illustrate how a test can have limited sensitivity and thus miss a substantial proportion of true cases while still maintaining high specificity, meaning that a positive result is rarely a false alarm. This example also illustrates that sensitivity and specificity are not fixed properties of a test: they vary depending on the target condition (gingivitis vs. periodontitis) and the clinical context.

3.2. Predictive Values

Positive and negative predictive values (PPV and NPV) describe test performance from the perspective of a patient's result. PPV is the probability that a patient with a positive index test result truly has the target condition, while NPV is the probability that a patient with a negative test result does not have the target condition. Predictive values are highly influenced by disease prevalence in the tested population. The higher the prevalence, the higher the PPV and the lower the NPV [4].

When the clinical objective is to rule in disease, a high PPV is required. Confirmatory HIV assays illustrate this principle: the World Health Organization recommends that HIV testing algorithms achieve a PPV of at least 99%, ensuring that a positive result can trigger initiation of antiretroviral therapy with a high level of confidence [5]. When the clinical objective is to rule out disease, a high NPV is required. For example, in periodontology, bleeding on probing (BOP) has a PPV of 6% and an NPV of 98% in predicting future attachment loss [6].

Self‐reported periodontal indicators deserve specific attention, particularly in epidemiological studies where full‐mouth clinical examinations are not feasible. Variables such as self‐reported GBOB have been evaluated against clinical reference standards and show high NPV for periodontal health (93%–98%), despite limited sensitivity (37%–61%). These results support GBOB as a relevant triage test to safely rule out gingivitis [4].

3.3. Likelihood Ratios

Likelihood ratios provide a summary of diagnostic accuracy by combining all four cells of the contingency table into a single measure. The positive likelihood ratio (LR+) indicates how much more likely a positive test result is in patients with the disease than in those without. The higher the LR+, the better the test is at confirming the target condition. The negative likelihood ratio (LR−) indicates how much less likely a negative result is in patients with the disease than in those without. The closer to 0 the LR−, the better the test is at excluding the target condition. As a rule of thumb, likelihood ratios above 10 and below 0.1 are typically considered strong evidence for ruling in or ruling out disease, respectively [7].

While likelihood ratios may sound useful in clinical practice, they also have limitations. First, while sensitivity and specificity are easily understandable because they are proportions between 0 and 100%, likelihood ratios do not easily translate into numbers and proportions of patient misclassifications. A second issue is that LR+ and LR− are not on the same scale. While LR+ should be as high as possible, LR− should be as close to zero as possible [8].

In the field of periodontology, likelihood ratios have been applied to salivary biomarkers such as MMP‐8 or IL‐1β [9]. MMP‐8‐based assays primarily reflect periodontal inflammation and therefore provide information on periodontal inflammation or risk rather than definitive diagnosis. A 2025 meta‐analysis (22 studies; 2492 participants) reported a summary sensitivity of 0.83 at a specificity of 0.77, and an AUC of 0.87 for detecting periodontitis. These values translate into an LR+ of 3.6 and an LR− of 0.22, indicating that the test only moderately provides confidence in ruling disease in or out among at‐risk patients [10].

Recent evidence suggests moderate‐to‐good accuracy for MMP‐8‐based assays; however, the target condition assessed by these tests requires clarification. Most studies evaluate the ability of MMP‐8 to discriminate between periodontal health and inflammatory periodontal conditions, rather than to diagnose periodontitis as defined by clinical attachment loss and radiographic bone destruction. Of note, in the present methodological article, we use the term “peridontitis” to refer to both gingivitis and periodontitis, whereas we use “periodontitis” when referring specifically to the established disease entity.

3.4. Receiver Operating Characteristic Curves (ROC) and Area Under the Curve (AUC)

While contingency tables and the above‐mentioned derived measures of accuracy imply binary index test results (i.e., positive or negative), tests and biomarkers that are continuous in nature may be evaluated using receiver operating characteristic (ROC) curves (Figure 1). The ROC curve plots Sensitivity against 1‐Specificity across different test thresholds, and the area under the curve (AUC) quantifies overall discrimination between diseased and non‐diseased patients [11]. According to Swets' classification, an AUC between about 0.50 and 0.70 indicates low accuracy, values between 0.70 and 0.90 indicate accuracy that may be useful for some purposes, and higher values reflect high diagnostic accuracy [12].

FIGURE 1.

Example receiver operating characteristic (ROC) curve. Four illustrative positivity thresholds are shown, with corresponding sensitivity, and specificity estimates: A, 150 (sensitivity 8%, specificity 99%); B, 100 (sensitivity 48%, specificity 89%); C, 50 (sensitivity 80%, specificity 63%); D, 20 (sensitivity 93%, specificity 39%). The curve was generated using simulated data with a target area under the ROC curve of 0.80 under a binormal model (N = 400; prevalence 10%), assuming biomarker values ranging from 5 to 350.

For continuous markers, defining test positivity requires selecting a decision threshold. The choice of a positivity threshold depends primarily on the clinical context and the consequences of testing errors. In practice, selecting a threshold requires weighing the relative clinical costs of false‐negative and false‐positive results and deciding whether higher sensitivity should be prioritized over higher specificity, or vice versa. The Youden index is frequently used to identify the threshold that maximizes the sum of sensitivity and specificity. However, its use should be interpreted with caution, as it implicitly assigns equal weight to false‐negative and false‐positive errors and may therefore be poorly aligned with clinical decision‐making.

Although ROC analyses of salivary biomarkers such as MMP‐8 or IL‐1β have shown promising discriminative ability (i.e., high AUC), particularly for identifying inflammatory periodontal conditions [13], the 2025 EFP Workshop emphasized that these biomarkers should not be interpreted as indicators of ongoing tissue destruction. Rather, they may contribute to risk assessment and screening strategies when combined with conventional clinical examination [14]. At present, no biomarker has demonstrated a combination of sufficient sensitivity and specificity, adequate predictive values at real‐world disease prevalence, and longitudinal validation required to replace clinical probing or to reliably predict future disease progression at the individual patient level.

3.5. Choosing the Most Relevant Diagnostic Accuracy Measure

The different pairs of measures that can be computed from a contingency table—sensitivity and specificity, predictive values, and likelihood ratios—are not equivalent, and each provides a distinct perspective on test performance (Table 3). Sensitivity and specificity reflect the test's performance in a group perspective, while predictive values are closer to individual decision‐making. Likelihood ratios offer an approach for integrating test results with pretest probabilities (i.e., disease probability before performing the test) to estimate posttest probabilities (i.e., disease probability according to the results of the test), using Bayes' theorem [7]. In practice, Bayes' theorem means that the clinical interpretation of a test result depends on the probability of disease before testing. Likelihood ratios quantify how much a positive or negative result changes that probability, from pretest to posttest probability. This can be illustrated using Fagan's nomogram (Figure 2) [15].

FIGURE 2.

A Fagan nomogram illustrating the shift from pretest to posttest probabilities according to likelihood ratios. This situation corresponds to the numbers presented in Table 2, with associated diagnostic accuracy measures.

The choice of which measure(s) to highlight in a study should depend on the study question, perspective, and hypotheses (Figure 3). Understanding the nuances of these measures ensures appropriate interpretation. Ideally, a study should focus on a limited set of predefined accuracy outcomes to avoid post hoc selection of the measure that appears more favorable to the test.

FIGURE 3.

Relationships among the main diagnostic accuracy measures computed from a 2 × 2 contingency table. Sensitivity and specificity describe test performance at the population level by comparing index test results with the reference standard in a group of patients with and without the target condition. Predictive values support patient‐level interpretation by estimating the probability of disease given a patient's test result. Likelihood ratios support clinical decision‐making by quantifying how much a test result changes the probability of disease.

When designing a diagnostic accuracy study, researchers should prespecify the most clinically relevant accuracy measure, and this choice should form the basis of the sample size calculation. For example, in a prospective study evaluating a diagnostic strategy combining the Wells rule and D‐dimer testing for pulmonary embolism in primary care, Geersing et al. based their sample size calculation on their primary objective of demonstrating that the false‐negative rate (defined as 1—NPV) of the strategy would be below 2.0%, reflecting a clinically acceptable rule‐out threshold [16].

In periodontology, clinicians often prioritize sensitivity to avoid missing the presence of periodontitis, accepting lower specificity. Yet from a patient perspective, false positives may lead to overtreatment (e.g., subgingival instrumentation such as unnecessary scaling and root planing), underscoring the need for balanced interpretation.

4. Sources of Bias and Variability in Diagnostic Accuracy Studies

To ensure transparency and reproducibility, diagnostic accuracy studies should follow the STARD (Standards for the Reporting of Diagnostic accuracy studies) reporting guidelines [2, 17]. Reporting quality remains a key determinant of study interpretability. Dental diagnostic accuracy studies often suffer from incomplete reporting with regard to essential STARD items, complicating risk‐of‐bias appraisal and meta‐analysis. Adherence to STARD (and discipline‐specific checklists) should be encouraged to enhance transparency and reproducibility [17, 18].

Diagnostic accuracy reflects the ability of a test to correctly classify patients as having or not having a given disease of interest (the target condition). It is a common misconception that diagnostic accuracy is a fixed and inherent property of a medical test. There is empirical evidence that, in reality, diagnostic accuracy is also affected by the context in which it is deployed, the population in which it is used, and the methods by which it is evaluated [19]. Several factors can introduce bias or variability into diagnostic accuracy studies, influencing the trustworthiness and generalizability of study findings (Table 4) [20]. These sources of bias and variation can lead to either overestimation or underestimation of test performance, sometimes with substantial implications for clinical practice.

TABLE 4.

Potential sources of bias in diagnostic accuracy studies.

Bias	Definition
Selection bias	When the study population is not representative of the intended use population.
Spectrum bias	When the spectrum of disease severity (and non‐diseased states) in the study group does not reflect that of the intended clinical setting.
Review bias	When the result of the index test is interpreted with knowledge of the reference standard result (and vice versa).
Incorporation bias	When the index test is included as part of a composite reference standard.
Partial verification bias	When not all participants receive the reference standard.
Differential verification bias	When not all participants receive the same reference standard.

Selection bias arises when the study population is not representative of the intended use population. For example, if a study evaluating a diagnostic test includes only patients with advanced disease and compares them to healthy controls, the resulting sensitivity and specificity are likely to be overestimated. This design, commonly referred to as a two‐gate or “case–control” design, does not capture the actual mix of cases across the spectrum of disease that is typically encountered in practice [3]. A well‐designed study should use a single‐gate “cohort” approach, enrolling patients who represent the intended use population.

Test performance can vary across patient subgroups defined by disease severity or other clinical characteristics, which is captured by the concept of “spectrum effect” [21]. In contrast, spectrum bias occurs when, in addition to expected variability, estimates of sensitivity and specificity are distorted due to systematic deviations from their true values because of inadequate patient selection. A common example of spectrum bias is seen in case–control studies that enroll only healthy participants. In periodontology, spectrum bias is an important consideration: studies limited to advanced periodontitis cases tend to overestimate sensitivity compared with studies that include early or moderate cases [22].

The way tests under evaluation are performed and interpreted in a study may also introduce bias. The index test(s) procedures should align well with routine clinical practice, ensuring that findings are relevant and reproducible in the field. Also, the index test should be interpreted without knowledge of the reference standard result; otherwise, it could lead to review bias.

Similarly, the reference standard used to establish the presence or absence of the target condition should be chosen and conducted appropriately. Incorporation bias can occur if the index test is included as part of the reference standard itself, leading to circular reasoning.

The reference standard should also be applied consistently across study participants; otherwise, it may lead to verification issues [23]. If not all participants receive the reference standard, this may lead to partial verification bias. Differential verification bias may occur if not all participants receive the same reference standard. Delayed verification bias may arise if there are long delays between the index test and the reference standard, allowing disease progression or treatment to affect results.

Even a well‐conducted study may have limited applicability. In such cases, accuracy estimates may be unbiased, but may not apply well to other populations and settings. For example, the results of a study conducted in a high‐risk population (e.g., tertiary care) may not be applicable to a low‐risk population (e.g., primary care), even if study methods are sound and unbiased.

One of the most widely used frameworks for assessing the quality of diagnostic accuracy studies is the QUADAS‐2 tool (Quality Assessment of Diagnostic Accuracy Studies version 2) [24]. This tool provides a structured approach to evaluating potential sources of bias and applicability concerns in a study, notably within the context of systematic reviews. QUADAS‐2 covers four domains: patient selection, index test, reference standard, and flow and timing. Each domain is evaluated for risk of bias, and the first three domains are also assessed for concerns about applicability.

Reproducibility refers to the consistency of test results when measurements are repeated under similar conditions and is commonly assessed using inter‐ and intra‐observer reliability metrics such as Cohen's kappa and the intraclass correlation coefficient; agreement between repeated measurements can be evaluated using Bland–Altman limits of agreement. Examiner calibration and reproducibility are conceptually distinct from diagnostic accuracy and have been covered in depth elsewhere [25]. They are nonetheless relevant to accuracy studies: without adequate calibration, systematic errors in index test measurements, such as probing depth or BOP, can bias diagnostic accuracy estimates and lead to overoptimistic or misleading conclusions. In addition, changes in disease case definitions over time (e.g., between the 1999 and the 2018 classifications of periodontitis [26]) may introduce heterogeneity that complicates evidence synthesis and meta‐analysis.

Periodontal diagnostic studies are particularly prone to several methodological issues. Inter‐examiner variability in probing depth measurement is a major concern, with possible differences of 1–2 mm between trained clinicians [27]. Radiographic assessment introduces further variability, as conventional two‐dimensional imaging underestimates bone loss compared with cone beam computed tomography (CBCT). Disease case definitions are heterogeneous across studies, with thresholds for clinical attachment loss or probing depth varying between classifications [28]. Moreover, incorporation bias may occur when the index test is included, wholly or in part, in the reference standard used to define the target condition. This situation is common in periodontal research, where periodontitis is frequently defined using composite criteria that include probing depth, clinical attachment loss, and radiographic bone loss. When probing depth is evaluated as an index test while simultaneously contributing to the reference standard, diagnostic accuracy may be artificially inflated due to circular reasoning [29]. In this context, the reference standard does not represent an independent measure of disease, but rather partially overlaps with the test under evaluation, which compromises the validity of accuracy estimates.

5. Advanced Diagnostic Accuracy Designs

5.1. Comparative Designs

Comparative diagnostic accuracy studies evaluate two or more index tests within the same population and against a common reference standard [30]. This allows direct comparison of test accuracy. When several index tests are performed in a study, researchers may compute various measures of comparative accuracy, including absolute and relative differences in accuracy. For example, when comparing Test A and Test B, researchers may estimate the absolute difference in sensitivity (i.e., Sensitivity_A—Sensitivity_B) or the ratio of sensitivities (i.e., Sensitivity_A/Sensitivity_B) between tests. In paired study designs, where multiple index tests are performed in the same individuals, statistical methods for estimating comparative accuracy measures should account for the nonindependence of test results [31].

Comparative designs are relevant in periodontology. For instance, Mohammed et al. directly compared a series of salivary biomarkers (MMP‐8, MMP‐9, and TIMP‐1) for the diagnosis of periodontitis against a consistent reference standard (alveolar bone level assessed by radiography) in the same 63 individuals [32]. Such head‐to‐head comparisons allow more robust conclusions than indirect comparisons of tests across studies that enrolled different patient groups. Direct comparisons may still be prone to bias. Moreover, direct head‐to‐head comparisons are not always ideal, as applying multiple tests to the same patient may be impractical or invasive. In such cases, randomized designs may appear justified.

5.2. Randomized Trials With Diagnostic Accuracy Outcomes

Randomized accuracy studies, where participants are randomly allocated to different index tests, minimize bias and may be valuable for assessing the added value of a new diagnostic test. These designs reduce the potential for bias because allocation is determined by chance. Randomization enhances comparability and strengthens causal inference that differences in accuracy arise from the test rather than from patient characteristics. However, randomized trials of tests where diagnostic accuracy is the primary outcome are often considered inefficient or unjustified [33]. Indeed, in most cases, multiple tests can be applied to the same patients, allowing for direct comparison of test performance within the same individuals (see above). When this is possible, applying both tests avoids the need for randomization. Randomized designs may still be warranted when tests are invasive or when performing both index tests in all patients is not feasible.

6. Beyond Diagnostic Accuracy: Evaluating the Effectiveness of Medical Tests

Figure 4 provides a conceptual framework summarizing the different levels at which medical tests can be evaluated, from clinical performance to clinical effectiveness. Indeed, diagnostic accuracy, while essential, does not fully capture the effectiveness of testing on patient‐centered outcomes, such as morbidity, mortality, quality of life, and healthcare resource use. Most medical tests do not improve patient outcomes directly; instead, tests may indirectly improve patient outcomes through their influence on patient management.

FIGURE 4.

A conceptual framework for the evaluation of medical tests: From clinical performance to clinical effectiveness.

Diagnostic accuracy studies do not inform about the effectiveness of tests to improve patient‐centered outcomes. Randomized trials comparing diagnostic pathways with and without the test are considered an appropriate approach to assess clinical effectiveness. In such trials, patients are randomized to be managed according to different test‐based strategies, and primary outcomes are patient‐centered (e.g., mortality, morbidity, quality of life) or system‐centered (e.g., costs) [34, 35].

For example, one could randomize patients to initial assessment with salivary biomarker testing versus conventional probing (usual care), followed by standardized management pathways. Such a test‐treatment trial would help clarify whether the addition of biomarkers modifies clinical decisions and leads to better outcomes. Of note, test‐treatment trials evaluate an entire diagnostic‐therapeutic strategy, and observed improvements in patient outcomes may reflect protocolized care rather than improved discrimination attributable to the new test.

Randomized trials of medical tests, however, face specific design challenges. One important and frequently overlooked issue is their statistical efficiency, that is, the ability of the trial to detect differences in patient‐relevant outcomes between diagnostic strategies for a given sample size. In randomized trials comparing outcomes in patients managed according to the results of Test A versus Test B, many participants will receive identical test results (i.e., A−/B− and A+/B+) and, consequently, identical downstream management regardless of the assigned test. These patients do not contribute information to the comparison of clinical outcomes between strategies. Only participants with discrepant test results (i.e., A+/B− or A−/B+), for whom test choice leads to different clinical decisions, can generate differences in treatment and outcomes. When discordant results are infrequent, very large sample sizes may therefore be required to detect modest differences in outcomes, limiting the feasibility of such trials [33]. This feature distinguishes diagnostic trials from therapeutic trials, where treatment assignment directly affects all randomized participants.

One approach to enhance feasibility and reduce costs is to randomize only patients with discordant test results. An example is a trial comparing hysterosalpingo‐foam sonography (HyFoSy) with hysterosalpingography (HSG) as the first‐line tubal test in infertile couples. Among the included women, 747/1026 (73%) had concordant test results. Only the 143/1026 (14%) women with discordant results were randomized to clinical management based on HyFoSy or HSG findings [36].

Diagnostic tests enabled by artificial intelligence (AI) are emerging as an important prospect. AI is a broad term describing computational approaches that perform tasks traditionally requiring human intelligence. AI encompasses a range of methods, including machine learning, deep learning, and generative models (such as large language models). In most medical applications, input data of various types (e.g., numerical variables, text, images, or biomarkers) are converted into numerical representations and processed by an algorithm to produce a classification or prediction. Although AI introduces specific terminology and technical considerations, the principles underlying the clinical evaluation of AI‐enabled diagnostic tests are not fundamentally different from those applied to other medical tests. For example, AI‐based methods applied to radiographs offer scalability but require evaluation across diverse populations and often remain a “black box.” Integrating these tools within a patient‐centered framework highlights the trade‐offs between diagnostic accuracy, feasibility, and long‐term consequences in terms of patient outcomes. One potential strength of AI is its ability to integrate multimodal data, such as clinical findings, radiographic information, and biomarker profiles, within a single predictive framework. This may support more refined risk stratification and possibly earlier screening or detection strategies [37].

7. Conclusion

Diagnostic accuracy is a crucial step in evaluating the performance of periodontal tests, but it is not sufficient to determine their clinical value. Sensitivity, specificity, predictive values, likelihood ratios, and ROC analyses provide complementary perspectives on test performance. Their estimation and interpretation may be subject to important methodological challenges. Beyond diagnostic accuracy, it is essential to address sources of bias, to adopt comparative or innovative study designs, and to evaluate whether new diagnostic approaches translate into better patient care and outcomes. The ultimate value of a diagnostic test lies in its ability to guide effective treatment decisions, to improve long‐term oral and general health outcomes. Periodontology illustrates the complexity and importance of this broader framework. As the field evolves to incorporate biomarkers, advanced imaging, and AI, rigorous methodological standards and patient‐centered evaluations will be crucial. Only by moving beyond diagnostic accuracy can we ensure that new diagnostic tests contribute meaningfully to precision periodontology and improved patient outcomes.

Author Contributions

Nancy Vegas: writing – first draft and writing – important intellectual content. Jérémie F. Cohen: writing – first draft and writing – important intellectual content. Lisa Friedlander: writing – important intellectual content. All authors approved the final version.

Funding

The authors have nothing to report.

Disclosure

The content of this manuscript was generated by the authors. The authors used ChatGPT‐5.5 for proofreading. All authors reviewed and approved the text after proofreading. No other use of AI was made.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

Data sharing not applicable to this article as no datasets were generated or analysed during the current study.