Hyperparameter-Tuned Machine Learning Models for Complex Medical Datasets Classification

Abstract

Medical datasets are unavoidable for disease classification at an early stage. Still, it is challenging to systemize the medical datasets because of some complex relations among them, and machine learning (ML) algorithms provide the opportunity to classify medical datasets accurately. This research paper interested in four types of frequently occurred diseases, such as diabetic retinopathy, lower back pain, heart failure, and breast cancer. After completing initial data preprocessing, we used seven traditional ML classifiers and then enacted the stacking and voting ensemble method for boosting the classification performance. The top three performing classifiers acted as base estimators for stacking and voting classifiers. After that, we used the random-search method and grid-search method to expose the best parameters of the model, where these parameters were indispensable to building the model for better performance results. Later on, after tuning the hyper-parameters, we collected the performance result from the model via 10-fold cross-validation techniques in the research work. After that, we collated the model's performance and perceived that, in the diabetes retinopathy medical dataset, the stacking classifier displayed the highest accuracy of 74.63%. The stacking classifier also outperformed other models with an accuracy of 89.76 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup> in the lower back pain dataset. In the heart failure dataset, once again, the stacking classifier outperformed other models with an accuracy of 88.08%. Finally, In the breast cancer dataset, the Extra Tree classifier performed marginally well with an accuracy of 94.96%. Therefore, our proposed prediction models can be used as an alternative or supportive tool for medical practitioners to identify those diseases with high accuracy.

Topics & Concepts