Diagnostic Accuracy

The proportion of patients with the target condition who receive a positive test result. A highly sensitive test is a reliable 'rule-out' tool.

• Formula: True Positives / (True Positives + False Negatives)
• Clinical Utility: A negative result on a 99% sensitive test confidently excludes disease.
• Risk: Over-prioritizing sensitivity without specificity leads to over-diagnosis and unnecessary invasive follow-ups.

The proportion of patients without the target condition who receive a negative test result. A highly specific test is a reliable 'rule-in' tool.

• Formula: True Negatives / (True Negatives + False Positives)
• Clinical Utility: A positive result on a 99% specific test confidently confirms disease presence.
• Trade-off: Increasing specificity often reduces sensitivity. The optimal balance depends on the clinical context and disease prevalence.

The Receiver Operating Characteristic (ROC) curve plots the true positive rate against the false positive rate at every possible classification threshold. The Area Under the Curve (AUC) provides a single scalar value representing overall diagnostic performance.

• AUC = 1.0: Perfect discrimination.
• AUC = 0.5: No discrimination (equivalent to random guessing).
• Regulatory Relevance: The FDA often requires ROC analysis to demonstrate that a SaMD algorithm performs consistently across diverse operating points.

The probability that a patient with a positive test result actually has the target condition. Unlike sensitivity and specificity, PPV is heavily influenced by disease prevalence.

• Formula: True Positives / (True Positives + False Positives)
• Prevalence Dependence: A test with 99% sensitivity and 99% specificity can have a low PPV if the disease is extremely rare.
• Clinical Impact: PPV directly answers the clinician's question: 'Given this positive result, how likely is my patient to be truly sick?'

The probability that a patient with a negative test result is truly free of the target condition. Like PPV, NPV shifts with disease prevalence in the tested population.

• Formula: True Negatives / (True Negatives + False Negatives)
• High NPV: Essential for screening tests to confidently discharge patients.
• Regulatory Context: For SaMD intended as a screening tool, demonstrating a high NPV across diverse demographic subgroups is a critical component of analytical validation.

Likelihood ratios combine sensitivity and specificity into a single metric that quantifies how much a test result will change the odds of having a disease.

• Positive Likelihood Ratio (LR+): Sensitivity / (1 - Specificity). Values > 10 generate large shifts in post-test probability.
• Negative Likelihood Ratio (LR-): (1 - Sensitivity) / Specificity. Values < 0.1 effectively rule out disease.
• Advantage: Unlike PPV and NPV, likelihood ratios are independent of disease prevalence, making them portable across different clinical settings.

The proportion of patients with a target condition who are correctly identified by the test. A highly sensitive test rarely misses the disease, making it ideal for screening where the cost of a false negative is catastrophic.

• Formula: TP / (TP + FN)
• Clinical Goal: Rule-out disease (SnNout).
• Example: A cancer screening AI with 99% sensitivity will miss only 1 in 100 malignant cases.

The proportion of patients without a target condition who are correctly identified as disease-free. A highly specific test rarely triggers a false alarm, making it crucial for confirming a diagnosis before invasive procedures.

• Formula: TN / (TN + FP)
• Clinical Goal: Rule-in disease (SpPin).
• Example: A biopsy triage AI with 95% specificity correctly spares 95% of benign patients from unnecessary interventions.

A graphical plot illustrating the diagnostic ability of a binary classifier as its discrimination threshold is varied. The curve plots the True Positive Rate (Sensitivity) against the False Positive Rate (1 - Specificity) at every possible threshold.

• AUC (Area Under the Curve): A single scalar value summarizing overall performance. An AUC of 1.0 represents perfect discrimination; 0.5 represents random chance.
• Use Case: Comparing two different AI models to determine which has superior overall diagnostic power regardless of the chosen operating point.

The probability that a patient with a positive test result actually has the disease. Unlike sensitivity, PPV is highly dependent on disease prevalence in the tested population.

• Formula: TP / (TP + FP)
• Clinical Impact: A high PPV ensures that a positive AI alert is actionable and does not overwhelm clinicians with false alarms.
• Example: An AI with 99% sensitivity may still have a low PPV if screening a very rare disease in a general population.

The probability that a patient with a negative test result truly does not have the disease. NPV provides confidence to safely discharge a patient or avoid further workup.

• Formula: TN / (TN + FN)
• Clinical Impact: Critical for emergency room triage where a high NPV allows clinicians to confidently rule out life-threatening conditions like pulmonary embolism or aortic dissection.
• Prevalence Dependency: NPV increases as disease prevalence decreases in the tested population.

The foundational NxN table that visualizes the performance of a classification algorithm by tabulating True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN).

• Structure: Actual class vs. Predicted class.
• Utility: Every other accuracy metric—sensitivity, specificity, precision, and F1-score—is derived directly from the four cells of this matrix.
• Medical Context: A critical tool for analyzing the specific types of errors an AI makes, distinguishing between dangerous misses (FN) and costly over-calls (FP).