Feature fusion is an important component of deep learning models, which improves feature representation and enhances classification accuracy. The study proposes AutoFusionNet, a learned algorithm for automatic feature fusion that optimizes multimodal data integration for Parkinson’s disease diagnosis. Unlike the previously developed fusion methods that have relied on manual feature selection, AutoFusionNet dynamically learns the best feature aggregation strategy through adaptive attention mechanisms and multi-scale feature encoding. The algorithm is designed with a transformer-based architecture combined with convolutional feature extractors that capture local and global dependencies of multimodal datasets, including motor assessments, speech data, and neuroimaging. Benchmark experiments were conducted on Parkinson’s disease datasets, and AutoFusionNet outperformed the conventional fusion techniques. On the classification task of Parkinson’s disease, AutoFusionNet obtained a classification accuracy of 92.7%, outperforming ResNet-50 (85.4%) and DenseNet-121 (88.1%). These results show that AutoFusionNet is a highly efficient and robust solution for integrating heterogeneous data sources, especially in the diagnosis of Parkinson’s disease. With AutoFusionNet, the fusion process is automated, eliminating the need for manual feature selection.