Integrating Hierarchical Structures in Medical Data Classification: A Kernel-Based Method

Abstract

Medical diagnosis systems frequently rely on structured information collected during physician–patient interviews. These data naturally follow a hierarchical organization, where general questions are followed by more specific sub-questions. Such a structure should be explicitly incorporated into similarity measures used in classification algorithms, as it reflects dependencies
between symptoms and contributes essential diagnostic information. In this work, we introduce a kernel that simultaneously accounts for (i) the hierarchical structure linking main questions to their subordinate items and (ii) interactions among sub-variables. The kernel is integrated into the pgpDA classification framework, allowing the method to embed prior
knowledge on how variables are organized and how symptoms interact. The proposed kernel is designed for binary data arranged in two-level tree structures and supports interaction modeling of any given order. Experiments conducted on simulated data and a real verbal autopsy dataset from Senegal demonstrate consistent improvements over classical kernels, and a deep-learning benchmark confirms that the structured kernel retains strong predictive power even in modern architectures. The methodology may be extended to mixed data types or adapted to graph-structured symptom networks.