Application of cross attention-based heterogeneous data fusion framework in alcohol dependence classification

Alcohol dependence (AD) has become the third-largest public health issue worldwide, trailing only cardiovascular diseases and malignancies, with its adverse effects on society and individual health intensifying. Effective early detection is critical for timely intervention and treatment of AD. This study compares the performance of Extended Long Short-Term Memory (xLSTM) and Transformer models in classifying AD and control groups using genetic information based on Single Nucleotide Polymorphisms (SNPs). Furthermore, we propose a Cross-Attention-based framework for integrating SNP and demographic information through representation-level interactions. Under model-level fusion, the Transformer-based Cross-Attention model yielded mean accuracy, precision, sensitivity, specificity, and AUC values of 83.94%, 86.36%, 85.03%, 88.41%, and 0.93, respectively. In the formally tested single-source settings, no between-model differences remained statistically significant after correction for dependence between repeated runs and for multiple comparisons. Comparisons involving the fusion settings were not formally tested and are therefore interpreted descriptively. These findings demonstrate the technical feasibility of using Cross-Attention to jointly model SNP and demographic features and provide a basis for further evaluation in larger independent AD cohorts. An additional experiment on a public ATR-FTIR saliva dataset further examined the applicability of the framework to a different biomedical classification setting.