📄 medicine

Knowledge-Augmented Spatiotemporal Modeling of Electronic Medical Records for Heart Failure Prognosis Prediction: Development and Validation of KA-STNet

This study presents KA-STNet, an interpretable knowledge-augmented spatiotemporal framework that integrates external medical knowledge with graph neural networks to effectively predict in-hospital mortality and discharge cardiac function grades for heart failure patients using heterogeneous electronic medical records.

Original authors: Yichen Li, Zhongrang Xie, Muyu Wang, Xin Xin, Jun Huang, Congmin Zhu, Lan Wei, Xiaolu Fei, Hui Chen

Published 2026-08-31

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yichen Li, Zhongrang Xie, Muyu Wang, Xin Xin, Jun Huang, Congmin Zhu, Lan Wei, Xiaolu Fei, Hui Chen

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Heart failure is a condition where the heart struggles to pump enough blood to meet the body's needs. It is a serious, progressive illness that affects millions of people worldwide, often leading to frequent hospital visits and a high risk of death. For doctors, the challenge lies in the fact that every patient is different; their symptoms, medical history, and how their body reacts to treatment vary wildly. To manage this, hospitals rely on electronic medical records, which are vast digital archives containing everything from a patient's age and past diagnoses to daily blood test results. While these records hold the keys to predicting who might get worse, they are incredibly complex. They mix static facts, like a patient's age or past surgeries, with dynamic data, like blood test results that change over time. Traditional computer programs often struggle to connect these two types of information effectively, treating them as separate lists rather than parts of a single, living story.

Researchers at Capital Medical University in Beijing have developed a new computer system designed to read these medical records with a level of understanding that mimics a seasoned clinician. They call their creation KA-STNet. Instead of just looking at numbers in isolation, this system builds a map of relationships between different medical facts. It combines two powerful approaches: it uses established medical knowledge about how diseases and treatments relate to one another, and it learns from the actual patterns found in thousands of patient records. The system is designed to do two specific things: predict whether a patient with heart failure might die during their hospital stay, and estimate the severity of their heart function when they are finally discharged. By weaving together what is known about medicine with what is observed in the data, the researchers aimed to create a tool that is not only accurate but also understandable, showing doctors exactly which factors are driving the prediction.

The team tested their new system on two large groups of patients. One group came from a local hospital in Beijing, containing over 5,000 records, and the other came from a massive public database of intensive care patients in the United States. They asked the computer to predict outcomes for these patients and compared its performance against six other advanced computer models that are currently considered the best in the field. The results were clear: KA-STNet outperformed all the other models. On the local dataset, it correctly identified the risk of death with a score of 0.968, a measure of accuracy where 1.0 is perfect. On the more difficult and diverse public dataset, it still achieved a strong score of 0.897. When predicting the severity of heart function at discharge, the system also led the pack, achieving a score of 0.925. These numbers suggest that the new method is particularly good at handling the messy, real-world complexity of patient data, where information is often incomplete or arrives at irregular times.

To understand why this system worked so well, the researchers broke it down into parts, much like taking apart a watch to see how the gears interact. They found that the system's success relied heavily on its ability to use static information—things like a patient's age, gender, and past medical history—to guide its understanding of the changing data. When they removed this static information, the system's accuracy dropped significantly, proving that knowing a patient's background is crucial for interpreting their current blood tests. The system also uses a special technique to learn from medical textbooks and databases, using this external knowledge to build a foundation before it even looks at a single patient's record. This "knowledge-augmented" approach allowed it to make smarter connections than systems that rely solely on statistics. Furthermore, the system successfully modeled how different laboratory tests influence each other over time, recognizing that a change in one blood marker might signal a problem with another organ system days later.

Perhaps most importantly, the researchers ensured that the system could explain its own reasoning. In medical settings, a computer cannot simply give a number; doctors need to know why that number was chosen. The team used special analysis tools to trace the system's decisions back to specific factors. They discovered that the system correctly identified anemia, kidney function issues, and signs of infection as the most critical warning signs for patients with heart failure. It also revealed complex patterns, showing how problems in the body's acid-base balance, inflammation, and liver function tend to cluster together in patients who have poor outcomes. By grouping these related laboratory tests, the system highlighted how different parts of the body interact during a crisis. This ability to point to specific, clinically meaningful groups of data means that doctors can trust the system's predictions and use them to make better decisions about who needs urgent care and how to manage a patient's recovery.

The study concludes that combining deep medical knowledge with the ability to track changes over time creates a powerful tool for predicting heart failure outcomes. While the system showed great promise, the researchers noted that it was tested on specific datasets and that further validation in different hospitals would be necessary before it becomes a standard part of daily medical practice. The work demonstrates that when computer models are built to respect the complexity of human biology and the structure of medical knowledge, they can uncover insights that simpler methods miss. This approach offers a path forward for turning the overwhelming volume of hospital data into clear, actionable guidance that could help save lives and improve the care of millions of people living with heart failure.

Technical Summary: KA-STNet for Heart Failure Prognosis

Problem Statement

Heart failure (HF) is a progressive clinical syndrome with high mortality and frequent hospitalizations. Accurate prognosis prediction, specifically in-hospital mortality and discharge cardiac function grading, is critical for risk stratification and management. However, existing deep learning approaches face significant challenges when applied to Electronic Medical Records (EMRs):

Data Heterogeneity: EMRs contain diverse static data (demographics, diagnoses, medications) and irregular time-series data (longitudinal laboratory tests).
Lack of Explicit Relational Modeling: Conventional models (DNNs, RNNs, Transformers) often fail to explicitly model the complex dependencies among heterogeneous clinical variables.
Limitations of Existing Graph Methods: Current graph-based EMR studies typically rely on a single source of information (either fixed external knowledge or data-driven statistics) and lack mechanisms to adaptively integrate these sources. Furthermore, most spatiotemporal graph models are designed for high-frequency ICU data, making them less reliable for the sparser, irregularly sampled routine inpatient laboratory data.
Context Dependency: Relationships among laboratory variables often depend on patient-level clinical contexts (e.g., comorbidities), which existing dynamic graph models often overlook.

Methodology: KA-STNet

The authors propose KA-STNet (Knowledge-Augmented Spatiotemporal Network), an interpretable framework designed to model heterogeneous EMRs for HF prognosis. The architecture consists of five core components:

1. Knowledge-Augmented Static Representation Module (KA-Module)

This module learns patient-level static representations by constructing a knowledge-augmented graph containing patient nodes and medical concept nodes.

Graph Construction: It integrates two sources of edge information:
- External Knowledge: Semantic relationships derived from the Biomedical Informatics Ontology System (BIOS).
- Data-Driven Statistics: Associations quantified using Positive Pointwise Mutual Information (PPMI) from co-occurrence patterns in the EMR data.
Adaptive Fusion: A learnable gating mechanism adaptively fuses the semantic correlation matrix ( $C_{se}$ ) and the statistical correlation matrix ( $C_{st}$ ) to generate a fused association matrix. This allows the model to balance prior medical knowledge with observed data patterns.
Representation Learning: A Graph Attention Network (GAT) processes the unified static graph to produce patient-level static embeddings ( $h_{Stat}$ ).

2. Spatial Dependency Encoder (SD-Encoder)

This component models inter-laboratory dependencies within longitudinal time series.

Context-Guided Attention: Unlike standard STGNNs, the SD-Encoder uses the static patient representation ( $h_{Stat}$ ) from the KA-Module to guide the learning of inter-laboratory dependencies.
Mechanism: It projects the static representation into a patient-specific bias matrix, which adjusts the query-key similarity scores in an attention mechanism. This generates a patient-specific weighted adjacency matrix ( $\tilde{A}_p^S$ ) representing the dependency strength between laboratory variables for that specific patient.
Output: A Graph Attention Network aggregates these dependencies to produce spatial representations ( $h_{T1}$ ).

3. Temporal Dependency Encoder (TD-Encoder)

This component captures longitudinal dynamics in laboratory time series.

Dual-Path Processing: It processes multivariate time series through two paths: one embedding laboratory variables and another embedding temporal sequences.
Fusion and Modeling: The representations are fused via a residual gating mechanism and fed into a Bi-LSTM to capture long-range temporal dependencies, resulting in temporal representations ( $h_{T2}$ ).

4. Dual-Stream Dynamic Feature Aggregator (DDF-Aggregator)

This module coordinates the spatial and temporal streams to ensure consistency in learned dependency structures.

Contrastive Alignment: It computes a structure-level contrastive loss by comparing the graph-structure embeddings derived from the spatial stream (attention matrices) and the temporal stream (Pearson correlation matrices of temporal representations).
Objective: The symmetric contrastive loss encourages the spatial and temporal streams to learn consistent dependency structures for the same patient while preserving patient-specific differences.

5. Fusion and Prediction Module

The final patient representation is an adaptive weighted sum of the static ( $h_{Stat}$ ), spatial ( $h_{T1}$ ), and temporal ( $h_{T2}$ ) embeddings. This fused vector is passed to a linear prediction layer for downstream tasks (mortality or cardiac function grade).

Key Contributions

Adaptive Knowledge Integration: The paper introduces a gating mechanism that adaptively fuses external medical ontologies (BIOS) with data-driven statistical associations (PPMI) for static graph construction, addressing the limitation of relying on a single information source.
Context-Aware Spatiotemporal Modeling: KA-STNet uniquely incorporates patient-level static representations to guide the learning of inter-laboratory dependencies in the temporal stream, acknowledging that lab relationships vary by patient context.
Dual-Stream Contrastive Alignment: The framework employs a structure-level contrastive loss to align spatial and temporal dependency structures, promoting consistency between the two streams without requiring paired ground-truth graphs.
Interpretability: The model supports interpretability through SHAP for temporal features, GNNExplainer for static concepts, and Leiden community detection to identify clinically meaningful laboratory communities and network reorganization patterns.

Results

The model was validated on a local real-world cohort (Beijing) and the public MIMIC-IV dataset across two tasks: in-hospital mortality prediction and discharge cardiac function grade prediction.

Performance: KA-STNet outperformed six state-of-the-art baselines (including MT-GRU, TA-RNN, MSTD-GNN, and Graph-spa) across all metrics.
- In-hospital Mortality (Local Dataset): Achieved an AUROC of 0.968, surpassing the best baseline by 0.017.
- In-hospital Mortality (MIMIC-IV): Achieved an AUROC of 0.897, surpassing the best baseline by 0.020.
- Discharge Cardiac Function Grade: Achieved a weighted AUROC of 0.925, with significant improvements over baselines (e.g., +0.048 over the best competitor).
Ablation Studies: Removing any major component (static features, time-series features, SD-Encoder, TD-Encoder, or the DDF-Aggregator) resulted in performance degradation. Notably, removing static features caused the largest drop in performance, highlighting their critical role in guiding dependency modeling.
Interpretability Findings:
- Key Predictors: Anemia-related variables, renal function indicators, and infection-related markers were identified as primary contributors to adverse outcomes.
- Network Patterns: Leiden community detection revealed that adverse outcomes were associated with network reorganization involving acid–base/electrolyte, inflammatory, and hepatic indicators, suggesting multi-organ interaction patterns.

Significance and Claims

The paper claims that KA-STNet provides an effective and interpretable framework for predicting HF outcomes using heterogeneous EMRs. The authors assert that their findings support the value of combining knowledge-augmented static representation learning with spatiotemporal dependency modeling.

Specifically, the study demonstrates that:

Integrating external medical knowledge with data-driven statistics adaptively improves the construction of task-specific graph structures.
Using static patient context to guide inter-laboratory association modeling enhances the representation of irregular time-series data.
The proposed dual-stream contrastive alignment effectively synchronizes spatial and temporal learning, leading to superior performance in complex, heterogeneous clinical prediction tasks compared to existing RNN, Transformer, and STGNN baselines.

The work concludes that this approach offers a robust solution for risk stratification and post-discharge management optimization in heart failure patients.