The logistic regression model achieves an AUC of 0.8666 (86.7%), indicating good discrimination between diabetic and non-diabetic patients across all thresholds. At the Youden-optimal threshold of 0.3227, the model reaches sensitivity of 82.8% and specificity of 78.2%. The strongest single predictor of diabetes onset is Glucose. Zero-coded missing values were detected and imputed in 652 observations across five clinical variables before model fitting.

8 of 8 clinical measurements differ significantly between diabetic and non-diabetic patients (p < 0.05 by Welch t-test). The most significant predictor is Pregnancies (mean diabetic = 4.866 vs. non-diabetic = 3.298, p = 0). Among the 268 diabetic patients and 500 non-diabetic patients, nearly all continuous measurements are elevated in the diabetic group.

Zero-coded missing values are most prevalent in Insulin with 374 zeros (48.7% of patients). In total, 652 zero-coded values were detected across 5 clinical variables before model fitting. All zeros were replaced with class-conditional medians (computed separately for diabetic and non-diabetic patients) prior to analysis.

Box plots show standardised (z-score) clinical measurements for diabetic vs. non-diabetic patients across the top 5 predictors ranked by model importance. Standardisation places all variables on a common scale, enabling direct visual comparison of distributional separation. Glucose shows the greatest separation between groups, consistent with its position as the strongest predictor in the logistic regression model.

6 of 8 clinical predictors have odds ratio confidence intervals that exclude 1.0, indicating statistically significant associations with diabetes at the 95% confidence level. Diabetes Pedigree Function has the highest odds ratio of 2.198 (95% CI: 1.2235 – 3.9853). An odds ratio above 1.0 indicates increased odds of diabetes per unit increase in that predictor; a CI crossing 1.0 means the effect is not statistically distinguishable from zero.

The ROC curve plots sensitivity (true positive rate) against 1 - specificity (false positive rate) across all possible classification thresholds. The model achieves an AUC of 0.8666 (86.7%), indicating good discriminative ability for clinical screening. At the Youden-optimal threshold, sensitivity is 82.8% and specificity is 78.2%, meaning the model correctly identifies 82.8% of true diabetic patients while correctly ruling out 78.2% of non-diabetic patients.

At the Youden-optimal threshold of 0.323, the model correctly classifies 222 diabetic patients (true positives) and 391 non-diabetic patients (true negatives), for an overall accuracy of 79.8%. 46 diabetic patients are missed (false negatives) and 109 non-diabetic patients are incorrectly flagged (false positives). Sensitivity of 82.8% means the model catches most true cases, which is important for a clinical screening tool where missed diagnoses carry high cost.

Glucose is the most important predictor with an absolute standardised log-odds of 0.9265, meaning a one-standard-deviation increase corresponds to the largest shift in diabetes log-odds among all clinical variables. The gap to the second-ranked predictor is 0.4459. Standardising by predictor scale puts all coefficients on a fair footing: variables with different measurement units (e.g., glucose in mg/dL vs. pregnancies as a count) are directly comparable in this ranking.