Interactive companion to our final report

Effects of Model Footprint Reduction on Chest X-ray Performance and Fairness

An interactive reading of our ECE/CSC 591 report. Move the slider to see the reported behavior of the DenseNet-121 model at different sparsity levels.

Read the full report on GitHub

Choose model sparsity

80%

More weights removed

Stable compression range
Reported diagnostic performance stays close to baseline through 80% sparsity.

Overall AUROC change from baseline - the marker shows the selected pruning level. Select the chart to enlarge it.

Diagnostic performanceResilientNegligible AUROC change.

Model behaviorUsefulStill discriminates between images.

Fairness resultUnchangedExisting subgroup gaps remain.

0% · Baseline80% · Stable in report90%+ · Failure region100% · Random-like limit

What this means at 80%

The report supports substantial model-footprint reduction without a meaningful loss in diagnostic capability. This does not remove the need for separate fairness work.

Research insights

What the experiment adds

The slider summarizes the main result. These notes show how the study reached it and what it means for real deployment decisions.

01 · RedundancyMost weights were not equally important.

AUROC, recall, and specificity remained stable through 80% sparsity. This indicates substantial redundant capacity in the pretrained model.

02 · Failure modeExtreme pruning creates a misleading signal.

Beyond 95% sparsity, recall moves toward 1.0 while specificity approaches 0.0 because the model trends toward a constant “abnormality” prediction.

03 · FairnessCompression is not bias mitigation.

Sex and age gaps remained until global failure. Narrower group differences at collapse do not represent more equitable diagnostic care.

Model

DenseNet-121 from TorchXRayVision, pruned with unstructured L1 weight pruning across convolutional layers.

Data

2,797 unique NIH ChestX-ray8 images. The analysis focused on four abnormalities and removed repeated patient records.

Audit

Eleven sparsity levels from 0% to 100%, with AUROC, recall, specificity, worst-group accuracy, and true-positive-rate disparity.

Evidence from the report

Key charts

Original figures extracted from the submitted final report. Select a chart to view it at full size. Click the GitHub report link above for the complete figure set and interpretation.

AUROC change from baseline across pruning levels for four chest X-ray findings — **Overall AUROC change**Performance remains near baseline until high sparsity, then declines sharply.

Recall across pruning levels for four chest X-ray findings — **Recall by pruning level**Recall stays stable early, then rises sharply during prediction collapse.

AUROC by gender across pruning levels for four chest X-ray findings — **AUROC by gender**Both groups follow the same broad pruning pattern, while their baseline gaps persist.

AUROC heatmap by age decade across pruning levels — **AUROC by age cohort**Age-group performance is stable through 80% sparsity before the failure region.