Explainability Aware Heterogeneous Temporal Graph Transformer for Automated Genetic Disorder Detection from Genomic Data

Genetic disorders affect millions worldwide, posing major challenges for accurate diagnosis and effective treatments. New developments in genomic technologies, especially Next-Generation Sequencing (NGS), have enhanced the classification of genetic modification and improved diagnostic abilities. Moreover, the Machine Learning (ML) and Deep Learning (DL) methods emerged as efficient tools for evaluating the intricate genomic data and determining the patterns, which can be generally complex to identify by employing conventional analytical techniques. Additionally, the Explainable Artificial Intelligence (XAI) has developed as a significant constituent in the analysis of genomic data by enhancing the interpretability of DL methods and showing the contribution of main genetic features involved in the detection of genetic disorders. This study presents an Explainability-Aware Heterogeneous Temporal Graph Transformer for Automated Genetic Disorder Detection (XHTGT-AGDD) approach. The primary objective of the proposed model is to improve the reliability and accuracy of genetic disorder detection from genomic data by integrating advanced techniques. Initially, preprocessing is performed using quantile transformation to normalize the input data for robust analysis. Following preprocessing, feature selection is performed using the fire hawk optimizer to identify the most informative feature subsets from the genomic data. Besides, the classification task is accomplished using a heterogeneous temporal graph transformer, which captures complex relationships and temporal dependencies among genomic features. To further refine the learning process, hippopotamus optimization algorithm is employed for model optimization. In addition, an explainability mechanism based on SHAP is incorporated to interpret the model predictions and highlight the individual genetic features toward the final decision. The experimental result analysis of the XHTGT-AGDD model is performed using a benchmark dataset. The simulation results reported the enhanced results of the XHTGT-AGDD technique in terms of several measures, achieving an accuracy of 99.12%. Therefore, the proposed model provides reliable and effective performance for automated genetic disorder detection from genomic data.