Early Detection of Chronic Kidney Disease Using Logistic Regression with SMOTE Balancing and Grid Search Optimisation: Evidence of Benchmark Saturation and the Case for a Reduced Primary Care Panel

Abstract

Chronic kidney disease advances silently until most kidney function is lost, and in low resource systems patients present when dialysis is the only option. This study develops and stress tests an early prediction model using logistic regression, a classification technique widely applied in clinical research. The model was implemented in Python with the Chronic Kidney Disease dataset from the UCI Machine Learning Repository, holding 400 patient records, 24 clinical and demographic predictors and one binary outcome, with 250 cases and 150 non cases. Exploratory analysis identified 1,012 missing entries, led by red blood cells at 38.00 percent and red blood cell count at 32.75 percent of records. Numeric attributes were imputed with the median and categorical attributes with the mode. The data was divided 70 to 30, and the Synthetic Minority Oversampling Technique was applied to the training partition alone, balancing it at 175 records per class. The base model returned 98.33 percent accuracy, 98.40 percent weighted precision, 98.33 percent weighted recall, an F1 score of 98.34 percent and a ROC AUC of 100.00 percent. Grid search over the penalty term and regularisation strength selected an l2 penalty with C equal to 2.8072, yielding no improvement on held out data at 96.67 percent accuracy. Ten fold cross validation returned 99.14 percent. A reduced panel of eight inexpensive predictors averaged 97.11 percent across 30 repeated splits, against 98.53 percent for the full 24 predictor model. The benchmark is saturated, and practical gain lies in cutting input cost rather than chasing accuracy