Diagnostic Added-Value Explained: How NRI and IDI Improve Test Evaluation, Net Reclassification Improvement

Mayta
Jul 24
3 min read

Introduction

Adding a new test or biomarker to an existing diagnostic process might seem beneficial—but does it actually improve decision-making? Diagnostic added-value research is designed to answer this question. Unlike standard accuracy studies that focus on sensitivity and specificity, added-value research determines whether a new test meaningfully enhances clinical prediction or changes patient classification.

This is particularly important in modern medicine, where costly imaging, novel biomarkers, or genetic tests are often introduced without clear proof of benefit. This article outlines how to quantify added diagnostic value using statistical tools such as AUC difference, reclassification tables, Net Reclassification Improvement (NRI), and Integrated Discrimination Improvement (IDI).

What Is Diagnostic Added-Value?

Diagnostic added-value refers to the incremental benefit a new test (e.g., D-dimer, a biomarker, imaging, etc.) provides when added to an existing diagnostic method.

Illustrative Setup

Suppose:

Test A (existing model) has an AUC of 0.70.
Test B alone gives an AUC of 0.75.
Combining both (Test A + B) yields an AUC of 0.80.

The added diagnostic value of Test B is calculated as:

Added Value = AUC of combined model − AUC of original model = 0.80 − 0.70 = 0.10

This indicates that the new test improves overall discriminatory power by 10 percentage points.

Empirical Example: DVT Diagnosis

In a study of 2086 patients with suspected deep vein thrombosis (DVT), researchers evaluated whether adding a D-dimer biomarker to a model based on history and physical examination (Hx + PE) improved diagnosis.

Study Setup:

Reference standard: Ultrasound imaging applied independently of the index test.
Base model: Medical history and physical signs.
Extended model: Base model plus D-dimer.

The Area Under the Curve (AUC) improved:

Base model: 0.72
Extended model: 0.87This jump of 0.15 represents strong added diagnostic value.

Limits of AUC Difference Alone

While the AUC gives a sense of overall discrimination, it has limitations:

It doesn’t reflect actual patient reclassification across decision thresholds.
Gains in AUC may appear small despite substantial improvements in clinical decision-making.
It lacks direct interpretability in terms of who gets diagnosed or treated differently.

To address this, reclassification analysis is needed.

Reclassification Tables and Clinical Thresholds

Reclassification tables show how patients shift across a diagnostic cutoff when a new test is added. In clinical settings, such thresholds often guide treatment or further testing.

Example Threshold: 25% Risk

Patients with a predicted DVT probability above 25% are referred for further management. Reclassification tables then categorize how patients move:

Upward (from ≤25% to >25%)
Downward (from >25% to ≤25%)

Reclassification in DVT Study:

Group	Total Patients	Upward Shift	Downward Shift
With DVT (n=416)	416	123	26
Without DVT (n=1670)	1670	116	227

These movements represent a 36% total reclassification rate in cases and 21% in non-cases, meaning many patients were reclassified by the new model.

Net Reclassification Improvement (NRI)

NRI quantifies whether reclassifications were beneficial or harmful:

NRI Formula:

NRI =[P(up | Disease) − P(down | Disease)] +[P(down | No Disease) − P(up | No Disease)]

Using the DVT data:

P(up | D = 1) = 123 / 416 = 0.30
P(down | D = 1) = 26 / 416 = 0.06
P(down | D = 0) = 227 / 1670 = 0.14
P(up | D = 0) = 116 / 1670 = 0.07

So:

NRI = (0.30 − 0.06) + (0.14 − 0.07) = 0.24 + 0.07 = 0.31

This suggests a 31% net improvement in correct classification after adding the new test.

Limitations of NRI

NRI depends heavily on the chosen probability threshold. Changing the threshold can significantly alter the result.

Problem:

A 25% threshold yields a certain NRI.
A 50% or 10% threshold might give very different NRIs.

This sensitivity can complicate comparisons across studies or settings.

Integrated Discrimination Improvement (IDI)

To overcome the threshold dependency, the IDI was developed. It reflects how much the new model improves the separation of predicted probabilities between those with and without disease—without relying on arbitrary cutoffs.

IDI Formula:

IDI =[(Mean P in Disease: new model − old model)] −[(Mean P in No Disease: new model − old model)]

Example Calculation:

Disease group: 0.49 − 0.13 = 0.36
No disease group: 0.28 − 0.18 = 0.10So:IDI = 0.36 − 0.10 = 0.26 (95% CI: 0.23–0.28)

This means the new model improves group separation by 26 percentage points.

Conclusion

Diagnostic added-value research is essential when evaluating whether a new test or biomarker enhances decision-making beyond current practice. While AUC differences provide an overview, reclassification metrics like NRI and IDI offer richer, patient-centered insights. They reveal whether the test meaningfully reclassifies patients, aids in treatment decisions, and improves clinical outcomes.

As healthcare grows more complex, using these methods ensures that new diagnostics are adopted based on value—not just novelty.

Let me know if you'd like a tabulated comparison of AUC, NRI, and IDI features—or a practice example to work through interactively.