Adaboost-Enabled Type-2 Fuzzy Classifier For Disease Staging
Shambhab Chaki, Sanlap Banerjee, Arka Roy, Pratyusha Rakshit · 2025
This paper presents a novel approach for disease staging in new patients with limited clinical records. Each disease stage is treated as a distinct class, characterized by two feature sets taken from electronic health records (EHRs), including quantitative laboratory test results and electrocardiogram (ECG) signal features. To model variations in feature measurements of a patient, a Gaussian type-1 fuzzy primary membership (PM) is constructed for each feature. The union of PMs from patients in the same disease stage forms an interval type-2 fuzzy set (IT2FS), representing uncertainty through its footprint of uncertainty (FOU). The accuracy in PM assignments of the given IT2FS is next assessed by general type-2 fuzzy set (GT2FS). A novel approach is proposed here to evaluate the secondary membership (SM) of the concerned GT2FS, emphasizing the midpoint of the FOU. These SMs refine existing PMs, shrinking the FOU and enhancing estimation precision. This process generates a fuzzy feature space for the proposed modified type-2 fuzzy classifier (MT2FC). A new patient is classified into a disease stage with the highest support based on fuzzy max-min and min-max operations and generating support measures for each class. To boost classification performance, adaptive boosting (AdaBoost) is applied to refine classification boundaries iteratively, leveraging the proposed belief on outcome (BOC) measure to avoid disturbances due to weak classification. The proposed method is evaluated against four baselines using data from MIMIC-III v1.4, eICU-CRD, and MIMIC-IV v3.1, covering ten diseases (seven mental disorders, two heart diseases, and chronic kidney disease). Experiments reveal that the proposed method outperforms its competitors, achieving a peak average Matthew’s correlation coefficient (MCC) of 0.94 for CKD.