MPC-Net: Detection of Chronic Kidney Disease Using Missingness-Pattern Conditioned Deep Learning

Abstract

Renal impairment is usually discovered late, because the biochemical panels capable of exposing it are ordered irregularly and returned with gaps. Standard practice fills those gaps with a column statistic and hands the repaired vector to a classifier, so a value the laboratory produced and a value the algorithm invented become indistinguishable. We take the opposite view and treat the pattern of what was left unmeasured as a second input channel. The proposed MPC-Net encodes that binary pattern through its own subnetwork and uses the resulting code to apply a feature-wise affine transform to the value pathway, so absence modulates how the recorded numbers are interpreted instead of impersonating them. The value pathway itself carries an explicit second-order interaction term, letting variable pairs contribute without an enumerated expansion, and two gated residual blocks refine the fused representation. Five independently initialised replicas vote, and a logistic recalibrator maps the pooled score onto a probability scale. Evaluation covers 600 patient records in two cohorts under five-fold partitioning repeated five times. On the first cohort MPC-Net attained 99.80 ± 0.58 per cent accuracy, 100.00 per cent area under the curve and a Brier score of 0.0013, and on the second 98.00 per cent accuracy. Permutation testing identified specific gravity and haemoglobin as the dominant evidence.