MIRA-Net: A Missingness-Informed Residual Attention Network for Chronic Kidney Disease Prediction

Abstract

Chronic kidney disease progresses silently, and the routine laboratory panels that could reveal it early are almost never complete: entries are skipped and a clinician is left reasoning from a partial record. Most published models delete such records or fill the gaps with column statistics, discarding the information carried by the pattern of absence itself. This work presents MIRA-Net, a missingness-informed residual attention network that treats every clinical variable as a token and encodes whether that variable was actually measured. A learned absence prototype replaces the value term for unobserved entries, a reliability gate scales each token by the evidence it carries, and two residual attention blocks fuse the tokens before a temperature-scaled class-token read-out. The network holds 79 618 parameters and classifies one record in 1.42 ms on a single processor core. Two independent cohorts totalling 600 records were used, one with 10.55 per cent of its entries absent. Under repeated stratified ten-fold cross-validation, MIRA-Net reached 99.67 ± 1.07 per cent accuracy, 99.74 per cent F1 score and 99.99 per cent area under the curve on the first cohort, with a Brier score of 0.0032. The reliability gate ranked specific gravity, haemoglobin and red blood cell morphology highest, matching accepted nephrological practice and giving every decision an auditable trace.