QUANTIFYING EXPLANATION DRIFT UNDER MODEL COMPRESSION IN CLINICAL RISK PREDICTION

Abstract: Predictive models on tabular clinical data increasingly use SHAP and LIME explanations, while compression is routine. This study quantifies how compression affects post hoc explanations on five clinical benchmarks. A multilayer perceptron was trained, then subjected to L1 pruning at four sparsities (30 to 90 percent) and quantization to four bit widths (8 to 2 bits), yielding nine variants. SHAP, LIME, and permutation importance were applied to each variant and compared to the full model. The transparency cost of compression is compression-type-dependent: heavy pruning generally degrades both accuracy and explanations together, so accuracy alone catches the problem; heavy quantization more often preserves accuracy while degrading explanations, so accuracy alone misses the problem. On two of five datasets at 2 bit quantization, AUC retention exceeds 0.92 while SHAP rank correlation against the full model falls below 0.70. Explanation fidelity should be reported alongside accuracy specifically when quantization is used.

Keywords: Explainable Artificial Intelligence, SHAP, LIME, Model Compression, Clinical Risk Prediction