An Intelligent Healthcare Framework for Early Chronic Kidney Disease Screening using Swarm-Optimized Feature Selection

Abstract

Chronic kidney disease progresses silently, and the laboratory panel used to detect it is broad, so any screening model that reaches a diagnosis from fewer measurements has practical value. This work develops a five stage selection pipeline and pairs it with a supervised support vector machine. Mutual information first removes attributes that carry almost no information about the class. A hybrid weighted recursive elimination stage then ranks the survivors by fusing rank normalised importances taken from a linear support vector machine, a random forest and a gradient boosting model, each weighted by its own inner cross validation accuracy. Profiling the ranked prefixes supplies a starting point for a chaotic adaptive binary particle swarm that searches the attribute mask and the kernel parameters at the same time, and the attributes retained across the elite band of visited solutions form the final subset. Every stage is fitted inside the training partition of a 10 times 5 fold cross validation. The model reached 99.30 percent accuracy with an area under the curve of 0.9987 while using 14 of the 24 attributes, a result statistically indistinguishable from using the complete panel and ahead of recursive elimination, a genetic algorithm and a conventional binary swarm evaluated under the identical objective. On an independent cohort of 200 records the same subset returned 98.00 percent accuracy.