Development of Diabetes Mellitus Diagnosis Method Based Random Forest Classifier with Particle Swarm Optimization

Diabetes Mellitus (DM) is a chronic metabolism that has symptom in a rise in blood sugar level. In terms of the total number of DM patients worldwide, Indonesia comes in at the seventh position out of 10 nations. DM is a disease that is very dangerous and causes various complications. To mitigate the effects and complications of DM, a diagnosis early on DM is necessary. The diagnosis of DM is generally carried out by an expert doctor who requires a high cost and the diagnosis result tends to be bias in between expert doctors. DM diagnosis method-based machine learning needs to be created to obtain a diagnostic method at an affordable cost. One of the known machine learning methods that can be used to develop DM diagnostic method is Random Forest (RF). It is relatively fast, simple, resistant to outliers and noise, easy to parallelize, avoids overfitting and is a good solution. However, the performance of the RF method is influenced by the selection of its hyperparameters. Appropriate selection of RF hyperparameters will improve the performance of all RF methods. RF hyperparameter selection can be solved by using global algorithm optimization. The Swarm Intelligence (SI) algorithm is one of the optimization methods for solving the optimization with many local optima points. It has many advantages, such as it is resilient, simple to use, does not require derivatives, and can identify a global optimal from multiple local optimums. Particle Swarm Optimization (PSO) is one of SI method that is widely used and has advantages in the balance of exploration and exploitation. To solve this problem, this works proposes a diagnosis DM by using RF-based approach and PSO. From the experimental results, it can be concluded that the proposed method is superior to the conventional RF and for DM diagnosis. The smaller number of particles results the faster in computational time. The different number of particles results not different performance significantly. The experimental results also show that the DM Diagnosis Method Based on the RF with PSO can avoid overfitting condition.