Distribution-aware neural network: A novel patient representation learning algorithm for Traditional Chinese medicine diagnosis prediction

Traditional Chinese medicine (TCM) diagnosis involves complex and implicit associations between heterogeneous symptoms and diagnostic patterns, and distributional heterogeneity across diseases and patients. Existing intelligent diagnostic models focus primarily on architectural optimization but lack explicit modeling of underlying symptom-diagnosis distributions, resulting in limited robustness, cross-disease generalization, and interpretability. To address these challenges, we propose a distribution-aware neural network (DANN) for diagnostic representation learning. The proposed framework incorporates explicit representations of both global and class-conditional feature distributions and integrates discriminative pruning and latent structure decomposition to capture population-level diagnostic regularities and fine-grained differential variations. In addition, we introduce a cross-disease clinical dataset (TCM-Chronic) covering 15 chronic diseases, 5876 clinical cases, and 97 diagnostic labels to simulate real-world comorbidity scenarios. Experiments on both a public multilabel dataset (TCM-Lung) and a cross-disease dataset demonstrate that the DANN consistently outperforms state-of-the-art machine learning, deep learning, and large language model baselines. With respect to TCM-Lung, the F1-score of the DANN is 0.5112, which is 3.6 percentage points greater than that of the strongest baseline. With respect to TCM-Chronic, the DANN achieves an F1-score of 0.7146, outperforming the random forest by 7.21 percentage points. Ablation and expert evaluations further confirm that distribution-aware modeling contributes to increased diagnostic robustness and better interpretability. These results indicate that explicitly modeling diagnostic feature distributions provides an effective paradigm for intelligent diagnosis, with potential applicability beyond TCM to broader clinical decision-support tasks. • A cross-disease diagnostic dataset with 15 diseases and 97 labels is constructed. • A distribution-aware neural network models symptom-diagnosis distributions. • State-of-the-art performance is shown across multilabel and small-sample settings. • Ablation studies and expert evaluation confirm enhanced interpretability.