On August 26, 2026, the journal Nature Computational Science published a paper by a group led by Kang Zhang, presenting the PULSE framework (Patient Unified Longitudinal Signal Engine).
This is a longitudinal self-supervised (unsupervised learning) approach that explicitly encodes personalized data about a patient's past states to reconstruct complete medical profiles based on subsequent incomplete measurements.
Unlike traditional models that focus on static or independent data types, PULSE accounts for biological interrelationships and temporal covariance of indicators. It uses historical paired data to improve consistency and generate information at the current visit, which is especially important for fragmented (mosaic) longitudinal measurements in real clinical practice.
Application of the model to UK Biobank data showed that PULSE accurately generates metabolomic and proteomic profiles based on routine blood tests. Compared to reference data (251 biomarkers), the generated profiles outperformed all baseline reference models, and algorithms trained on these synthetic data achieved an AUC (area under the ROC curve) of 0,72–0,83 for six common diseases. This accuracy is comparable to using real proteomic data.
The PULSE methodology is based on the fundamental principle that different clinical modalities represent biologically related, temporally covarying indicators of an individual's evolving physiological state. This allows the model to work effectively with incomplete observations typical of real medicine, where patients do not undergo all tests simultaneously.
In the landscape of artificial intelligence in biology and medicine, PULSE stands out among other approaches such as AlphaFold or models for single-cell data analysis. It focuses specifically on temporal dynamics and multimodal integration of electronic health records (EHRs), retinal images, and laboratory markers, distinguishing it from static foundational AI models.
This achievement opens opportunities for more accurate disease prediction and personalized treatment based on routine data, reducing the need for costly specialized tests. However, the question remains about the model's ability to generalize to other populations and data types not represented in the UK Biobank.
Independent verification on new cohorts and comparison with alternative self-supervised methods will show how robust PULSE's advantages are in real clinical scenarios.
The development of such frameworks underscores that the key to progress in medical artificial intelligence lies in the proper consideration of temporal dynamics and data incompleteness, not just in simply increasing model sizes.
