A Hybrid Machine Learning Model Based on LightGBM and a Genetic Algorithm for Diagnosing Diabetes

Diabetes mellitus is one of the most common chronic diseases, and its early detection helps to prevent cardiovascular, renal and retinal complications. We address the three-class task of separating healthy, prediabetic and diabetic patients from 13 demographic, anthropometric and symptom variables that can be collected without laboratory tests, using anonymized records of 13,053 patients of an endocrinology center, in which 2.40% of the cells were missing. LightGBM was combined with two separate genetic algorithms (GA). GA-1 tunes LightGBM imputation models with a label-free fitness (reconstruction error on artificially hidden cells), so the learned rule can be applied to a new patient. GA-2 tunes nine hyperparameters of the LightGBM classifier on a validation subset and averages an elite set of five configurations. The pipeline was tested on a locked test set of 2,611 patients and by nested 5 × 5 repeated cross-validation with every step fitted inside the training folds. On the test set LightGBM+GA reached an accuracy of 0.9345 (95% CI 0.9245–0.9441) and a macro F1-score of 0.9325, compared with 0.9284 and 0.9263 for the default LightGBM. The optimized model achieved 0.9306 ± 0.0047 accuracy, outperforming default LightGBM and matching random/Bayesian tuning. GA-1 imputation achieved the lowest Gower error (0.302). SHAP identified age, BMI, blood pressure, walking difficulty, general health, and family history as key predictors. The model is a candidate approach for low-cost pre-laboratory screening, pending external validation.

Hali tarjima qilinmagan