Does Model Compression Amplify Clinical Bias? A Subgroup Fairness Audit of Quantized ICU Risk Models on MIMIC-IV

<title>Abstract</title> <p>Model compression is a prerequisite for deploying clinical risk-prediction models on resource-constrained hardware, but its effect on subgroup fairness is rarely audited alongside its effect on aggregate accuracy. We train a compact multilayer perceptron (5,313 parameters) to predict 72-hour acute kidney injury (AKI) onset in a MIMIC-IV cohort of 52,178 adult first ICU stays, audit its fairness across gender, insurance, and agegroup subgroups, and re-audit three compressed variants: INT8 dynamic quantization, 30%/50% magnitude pruning, and ∼1/4parameter knowledge distillation. We report two methodological findings alongside the fairness results: (1) naive per-tensor dynamic quantization collapsed AUROC to chance level (0.500) due to a small number of physiologically implausible raw vitals/labs readings dominating the activation-quantization range– fixed at the source with a documented plausibility filter, recovering AUROC to 0.699; and (2) diffing an out-of-fold baseline against in-sample-evaluated compressed models produces a spurious ”compression improves fairness” artifact, corrected with a sameprotocol baseline. Using bootstrap significance testing (1,000 resamples), we find every compression method shifts subgroup-level AUROC by 1–4 points, with 50% pruning producing the largest and most consistently significant degradation, while insuranceand age-based equalized-odds gaps are large in absolute terms both before and after compression, driven partly by small-n subgroups. We discuss deployment implications for CPU-only clinical hardware and the limitations of proxy fairness axes.</p>