Designing a Meta Learning Classifier for Sensor-Enabled Healthcare Applications

• Published in: SN computer science 2024-01, Vol.5 (1), p.59, Article 59
• Main Authors: Patikar, Srabani, Saha, Anindita, Neogy, Sarmistha, Chowdhury, Chandreyee
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Breast cancer; Business metrics; Cervical cancer; Classification; Classifiers; Computer Imaging; Computer Science; Computer Systems Organization and Communication Networks; Data Structures and Information Theory; Datasets; Feature selection; Health care; Information Systems and Communication Service; Liver cancer; Liver diseases; Lung cancer; Machine learning; Methods; Original Research; Pattern Recognition and Graphics; Performance prediction; Software Engineering/Programming and Operating Systems; Standardization; Vision
• ISSN: 2661-8907; 2662-995X; 2661-8907
• DOI: 10.1007/s42979-023-02373-0

Being one of the fastest growing industries in recent times, healthcare has witnessed a complete transition in the last decade. Rapid technological and communication advancements in machine learning have contributed to predict fatal diseases with datasets available from various sources. Single-layer ensemble learning has been applied in literature extensively, to perform predictive analysis mostly in the form of bagging, boosting, and stacking. In this work, a meta learning model is proposed that combines two single-layered ensemble models to render the final prediction. The two layers of the ensemble model explore different sets of features subject to different base learners. Weighted majority voting is performed to make a decision at the two candidate ensemble models based on different feature sets that are combined at a later stage. Interestingly, one feature may be important to classify a particular instance to a class, while another feature indicates distinguishing pattern for a different class. So, the effect of important features would be captured in the prediction result of a single-layered ensemble. However, combining two ensemble models subject to selected feature and all feature sets adds robustness to the design while reducing the chances of overfitting. The role of dimensionality on the choice of ensemble layers is also considered in the work. The first layer deals with all features of the datasets and the second layer works on datasets after feature scaling and feature selection is applied on them. The proposed method is implemented for multiple benchmark datasets of the healthcare domain. Experimentally, it has been observed that double ensemble methods provide a better predictive analysis in comparison to single-layered ensemble models if the dataset dimensionality is moderate. The problem of overfitting is also significantly reduced in double-layered ensemble approach. The combination of more than one classifier also eliminates the dependency on a single classifier to perform predictions.