A multimodal electronic health record data processing system for predicting the cumulative risk of clinical events across multiple time points.

VN128458APending Publication Date: 2026-08-17ĐÀO XUÂN CƠ +1
0 Cites 0 Cited by

Patent Information

Application Number
VN1202606338
Authority / Receiving Office
VN · VN
Patent Type
Applications
Current Assignee / Owner
Filing Date
2026-07-27
Publication Date
2026-08-17

Smart Images

  • Figure VN1202606338_0
    Figure VN1202606338_0
Patent Text Reader

Abstract

The invention relates to a multimodal electronic health record data processing system for predicting the cumulative risk of clinical events occurring at multiple time points. The system acquires a sequence of the subject's historical data prior to the prediction point, including structured data of multiple visits and corresponding clinical notes when available. The structured data, including diagnostic codes, test values ​​and masks, demographic information, and time information, are encoded to create a structured clinical state for each visit. Clinical notes are encoded at the token level. The structured clinical state is used as a query in a cross-attendance mechanism to select relevant textual context. The textual context is projected and added along residual links to the structured state to create a merged state. When no qualifying notes are available, the merged state is retained as the structured state.The merged state sequence is processed along a chronic pathway to create a chronic evidence representation reflecting the cumulative disease burden. For each visit, the previously formed chronic state is used to generate the predicted state; the difference between the current merged state and the predicted state is identified as the acute innovation. The innovations are synthesized considering clinical content and time intervals to create the acute evidence representation. For each predictive milestone, the system individually determines the weights of acute and chronic evidence based on the subject's history and time-mile representation information, thereby creating a unique evidence representation for each subject-time-mile pair and the corresponding cumulative probability of the event occurring. Training can utilize observational label masks, cumulative risk order constraints, difficult negative pattern ranking, and evidence allocation orientation.Predictive milestones may include 7 days, 30 days, 90 days, 180 days, and 365 days; clinical events may include stroke, acute myocardial infarction, and acute kidney injury.
Need to check novelty before this filing date? Find Prior Art