The editorial “Dynamics-driven medical big data mining: A dynamic approach to early disease prediction and personalized care” intelligent medicine (February 2026, Volume 6, Issue 1) was written by Lu Wang (Tianjin Medical University), Han Lyu (Beijing Friendship Hospital of Capital Medical University), and Bin Sheng (Shanghai Jiao Tong University). We argue that the future of medical AI lies not only in diagnosing diseases after they become visible, but also in detecting early dynamic changes that occur before symptoms are fully manifest. By analyzing how health data evolves over time, from omics and medical records to imaging and wearable devices, AI could help identify the “tipping points” where the body moves toward disease. The authors also emphasize that these systems must be rigorously validated and used to support, not replace, clinical judgment.
From population average to individual tipping point
At the heart of this framework is the dynamic network biomarker (DNB) theory, which detects impending disease transitions by monitoring sudden increases in fluctuations and correlations within biomolecular networks. Previous studies summarized in the editorial have validated the DNB-based approach across two clinically important scenarios. The aim is to flag increased gene expression instability in influenza infection days before symptoms appear, and to identify genomic tipping points where cells transition from a benign to a malignant state with greater than 80% accuracy in predicting tumor progression.
For busy clinicians, the most directly relevant advancement may be individual-specific edge network analysis (iENA), which transforms molecular data into edge networks without the need for control groups and assesses important transitions using a single patient’s own longitudinal data. For transcriptome applications, this single-sample approach achieves area under the curve (AUC) values greater than 0.9, making it the first method in this class to provide dynamic evaluation applicable at the bedside in real time.
Hybrid AI closes the gap between models and patients
This editorial also presents evidence that combining mechanistic physiological knowledge with deep learning can significantly improve clinical utility, rather than relying solely on data-driven models. In type 1 diabetes management, a physiologically based long short-term memory (LSTM) network reduced the average absolute error of blood glucose prediction to 35.0 mg/dL, compared to 79.7 mg/dL with a traditional simulator, a reduction of more than 55%. These models create patient-specific digital twins that can be used to test treatment strategies in silico before clinical application.
The editorial discusses parallel advances across data modalities beyond metabolic diseases. Time graph neural networks applied to EHRs improved diagnostic prediction accuracy by 10-15% on the MIMIC-III dataset. Dynamic graph model derived from functional MRI predicts tinnitus treatment outcome. Transformer-based architectures trained on longitudinal EHRs demonstrate the ability to predict risk for multiple diseases, including diabetes and hypertension, through hierarchical attention mechanisms.
Enhance, rather than replace, clinical judgment
“These dynamics-driven approaches are designed to enhance, rather than replace, clinical expertise,” said corresponding author Professor Bin Sheng, a professor at the School of Computer Science at Shanghai Jiao Tong University. “These provide timely early warning signals that enable proactive intervention, moving healthcare from reactive to true prevention, while preserving the irreplaceable role of human judgment in complex medical decision-making.”
Current restrictions require careful deployment
The editorial is equally candid about the challenges that must be overcome for these tools to provide fair real-world benefits. Data heterogeneity and missing values can lead to false positive detections of important transitions, increasing network variability and generating false alerts. A more fundamental challenge is that current methods, while good at identifying statistical associations, cannot reliably distinguish between correlation and causation unless they incorporate medical knowledge and experimental validation. Interpretability remains a major barrier. Although tools such as SHAP and LIME provide partial explanations for model decisions, full transparency in deep architectures has not yet been achieved, and opaque predictions risk undermining the clinical confidence needed for adoption.
Ethical and regulatory concerns also require attention. Despite its decentralized training architecture, privacy risks remain with federated learning, and algorithmic bias is a particular concern when models trained on specific populations are deployed to underrepresented groups, which could widen rather than reduce health disparities.
The way forward: Multimodal integration and future validation
Looking ahead, the editorial identifies two priorities. The first is multimodal integration. Through advanced transformers, graph neural networks, and causal inference techniques including instrumental variable and counterfactual simulations, we fuse omics, imaging, EHR, and wearable data to build comprehensive causal models of individual disease trajectories. Second, and perhaps more important, is rigorous prospective validation. The authors emphasize that the gap between theoretical expectations and clinical implementation can only be bridged by well-designed prospective clinical trials and real-world deployment studies in diverse populations and healthcare settings.
Published as open access, this editorial serves as both a cutting-edge reference and a practical roadmap for clinicians, researchers, and healthcare leaders working at the intersection of medicine and artificial intelligence.
sauce:
Reference magazines:
Wang, L. Others. (2025). Dynamics-driven medical big data mining: A dynamic approach to early disease prediction and personalized care. intelligent medicine. DOI: 10.1016/j.imed.2025.10.001. https://www.sciencedirect.com/science/article/pii/S2667102625001068?via%3Dihub
