Calculation of Sensitivity and Specificity from Partial Data for Meta-Analyses: Introducing Some Practical Methods.

Area of Science:

• Biostatistics
• Medical Informatics

Background:

• Meta-analyses require detailed diagnostic accuracy metrics (true positives, true negatives, false positives, false negatives).
• Primary studies often report incomplete data, hindering accurate meta-analysis.
• Reconstructing sensitivity and specificity from minimal data is crucial for comprehensive systematic reviews.

Purpose of the Study:

• To consolidate practical methods for reconstructing sensitivity and specificity from incomplete data.
• To evaluate the accuracy of different reconstruction techniques.
• To provide guidance for meta-analysis of diagnostic studies.

Main Methods:

• Utilized algebraic rearrangements to compute specificity from partial metrics.
• Employed receiver operating characteristic (ROC) curve digitization to extract threshold-specific sensitivity and specificity.
• Applied the binormal model using Area Under the Curve (AUC) and prevalence data.
• Tested methods on a myocardial infarction mortality prediction dataset using machine learning models.

Main Results:

• Algebraic formulas and ROC digitization provided reliable sensitivity and specificity estimates.
• The binormal model demonstrated inaccuracies, particularly for sensitivity, when assuming equal variances.
• Higher prevalence and AUC were associated with reduced estimation errors in linear regression analyses.

Conclusions:

• Practical methods exist to reconstruct diagnostic accuracy measures from incomplete data.
• AUC-based estimations can introduce significant bias, especially in low-prevalence settings.
• Primary studies should report threshold-specific sensitivity and specificity to improve meta-analytic accuracy.