Graph Network Feature Space Fusion for Predicting Irregularly Sampled Medical Time-Series Data: Deep Learning Model Development and Validation Study

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Single-target classification: accuracy 0.74–0.88, AUC 0.70–0.87; Single-target regression: MAE 0.038–0.061; Multi-target classification: mean accuracy 0.75–0.91, mean AUC 0.70–0.75

• Primary Outcome: Predictive performance of the proposed GNN-based model on single-target classification, single-target regression, and multi-target classification tasks across four clinical datasets

• Nnt Or Sensitivity: Sensitivity and specificity not individually reported; AUC range 0.70–0.87 across tasks suggests moderate to good discriminative performance, but calibration and clinical threshold performance are not provided

• Confidence Interval: Not reported

Clinical Application

Feasibility for clinical deployment is not assessed. The model requires graph construction from raw time-series data, GNN infrastructure, and integration with existing EHR pipelines. Computational requirements, inference time, and IT integration burden are not described. No prospective clinical validation has been conducted. This model is not currently approved by the TGA as a Software as a Medical Device (SaMD) and has no PBS-listed application. Australian emergency departments and ICUs (e.g., those using EMR systems such as Epic, Cerner, or Metavision) generate precisely the type of irregularly sampled data this model targets. RACGP and ACEM guidelines do not currently endorse GNN-based clinical prediction tools. Before any Australian clinical application, prospective validation in Australian cohorts, TGA SaMD regulatory assessment, and integration with local governance frameworks (including the Australian Digital Health Agency's AI in Health framework) would be required. The ANZICS and AIHW datasets could serve as appropriate validation cohorts. Potentially applicable to emergency department patients and critically ill hospitalised patients where clinical data is collected at irregular intervals with frequent missing values — a near-universal characteristic of real-world ICU and ED electronic health records

Abstract

BACKGROUND: Irregularly sampled data, as a common data structure in the medical field, is frequently observed in emergency clinical datasets. It poses problems such as unequal sampling time intervals and frequencies, making it difficult to align the data without losing information for input into models. Meanwhile, due to information loss in the dataset, it is also difficult for the model to effectively analyze the overall data variation and predict the complex and dynamic conditions of emergency patients in the future. OBJECTIVE: This study aimed to address the impact of issues such as missing values and irregularly sampled time intervals on prediction results. This paper proposes a method that converts time-series data into a graph network structure and develops a prediction model for graph data to better support complex tasks. METHODS: In this paper, feature channels and measurement times in time-series data are constructed as nodes, and feature measurements at time points are constructed as edge weights. The point feature space and edge weight space are extracted by a graph convolutional neural network and a gated convolutional attention mechanism network, respectively, and feature fusion is carried out under the action of the learnable head. RESULTS: The proposed method was tested on 4 datasets and compared with both irregularly sampled time-series predictors and state-of-the-art methods for complete data. In addition, a dedicated analysis was conducted leveraging data from 5 key clinical treatment phases. In single-target classification and single-target regression tasks, the method proposed in this study outperforms other comparative methods in most experiments (accuracy=0.74-0.88, area under the curve=0.70-0.87, and mean absolute error=0.038-0.061). Meanwhile, this study constructs clinical phenotypes during patients' hospitalization as multitarget prediction tasks, and the proposed model also outperforms comparative methods in multitarget classification tasks (mean accuracy=0.75-0.91 and mean area under the curve=0.70-0.75). Ablation experiments verify the effectiveness of fusion between the node feature space and the edge feature space. The interpretability analysis of time points identifies key features that require focused attention for patients across different target tasks and time periods. CONCLUSIONS: The proposed model can effectively analyze medical data with missing values and uneven sampling. It achieves high prediction accuracy in both regression prediction tasks and classification tasks. The graph network-based model structure endows the model with better interpretability, enabling effective exploration of factors that trigger changes.

References

1. 1.Hong, T., Ren, Z., Fang, J., Quan, S., Pan, J., & Lin, Y. (2026). Graph network feature space fusion for predicting irregularly sampled medical time-series data: Deep learning model development and validation study. JMIR Medical Informatics. https://doi.org/10.2196/81145