Geometric Deep Learning for Multimodal Data in CKD

CKD affects 14% of all adults and is a leading cause of mortality. Early diagnosis and appropriate treatment are crucial in reducing this increasing burden. The etiologies of CKD can be variable and multifaceted. Comorbidities (diabetes and hypertension), genetics (e.g., ApoL1-mediated CKD), social determinants of health, behavioral health, and other etiologies can all contribute and interact with one another to affect pathogenesis and variation in individual outcomes.¹ Identifying disease subphenotypes, predicting disease trajectories, and identifying novel biomarkers and physiologic drug–ligand interactions are required for optimal individualized treatment. An artificial intelligence paradigm accomplishing these objectives would ideally process not only clinical (electronic health record, clinical values) and histopathology data but also interact with and draw from existing knowledge.

Conventional machine learning (ML) models (Supplemental Table 1)² are limited in their ability to holistically interpret and model dynamic, multimodal data with complex interrelationships. ML is constrained to modeling data distributed within strictly planar dimensions: Data exist as isolated entities defined by numeric values in tabular format where each row represents an independent instance and each column represents a distinct and human-curated feature. When learning from tabular data, traditional ML models construct decision boundaries that partition the input distribution space into output regions corresponding to different classes or predictions. Deep learning (DL) expands on this approach, learning more sophisticated decision boundaries and partitions as DL models automatically learn higher-order features from the data, escaping the limitations of human-curated features. Architectural details of DL models influence the features they can learn (Supplemental Table 2). The added learning capacity comes with steeper data requirements, and models remain confined by the limitations of planar space. This reduces the range of clinically relevant information presented to the model even before features can be learned. In CKD, dynamic relationships between etiologies inform disease phenotype. These relationships are challenging to represent in tabular format.

By contrast, the geometric paradigm explicitly models data as interconnected nodes and edges, where the network structure itself, including its temporal evolution, contains vital information. In geometric learning, predictions depend not just on a node's features but also on its relationships and position within the broader network topology. The geometric paradigm naturally encodes temporal information, integrates multimodal data, and incorporates holistic and contextual information.

Graphs, the basic structure of geometric data, are a flexible way to represent complex, interconnected data. In a graph, nodes represent entities (patients, clinical concepts, etc.) and edges represent relationships between nodes (Figure 1A). These relationships can take many forms.³ For example, two patient nodes might be connected if they share a clinical feature, or a patient node can be linked to a diagnosis node or concept in a medical knowledge graph. Graphs are especially useful in clinical applications. For instance, patterns in node representations can be learned to predict disease states. Edge probabilities (likelihood of two nodes to be connected) can identify drug–disease or drug–drug interactions. Graphs can also easily model 3D molecular structures and predict drug–ligand interactions. This has already shown potential in drug development and repurposing.⁴ In CKD, graphs readily capture temporal changes in disease presentation, allowing for ML of disease trajectory for better prediction and clinical management.

Figure 1.

Overview of components, structure, and functioning of GNNs. (A) Graph of clinical data with nodes representing either patients or disease conditions or procedures, with edges representing interconnections of patients with each other as well as disease states that they are experiencing or clinical procedures they might have undergone. (B) A model showing how GNNs can act on graph data to achieve downstream tasks, e.g., prediction of risk for disease progression. (C) GNNs with representation data wherein patient data including omics, imaging, and clinical data can be worked on by a GNN, which can then be represented as a computer readable embedding (patient vector) and then used for multiple downstream tasks. GNN, graph neural network.

Graph neural networks (GNNs) extend the feature-learning capacity of DL to graphs (deep geometric learning), often including common mechanisms from DL (Supplemental Table 2). GNNs flexibly learn how and what relational information in the graph should be integrated into model inferences. For example, GNNs can include more neighbor information for nodes with fewer connections or limit inclusion from neighbors when a node has certain qualities. GNNs can also operate on graphs consisting of or including established biomedical ontology information (knowledge graphs). GNNs therefore have a high potential for producing clinically reliable outputs and predictions (Figure 1B).⁵ The data from a GNN can further be embedded into a representation that embodies salient features learned from input data (Figure 1C).⁵

Johnson et al.⁵ used heterogeneous graph networks to learn representations for medical concept codes. In brief, representations of relational structures from a clinical knowledge graph of medical codes were able to reproduce important clinical relationships between medical codes, e.g., the representations position disease sequela into similar positions. These encodings (and graph structures) can in turn be included in other biomedical graphs to infuse prior knowledge and standardization into a new task. Although GNNs seem ideally suited to model a dynamic condition such as CKD, they are not yet widely used.

Using data from the India ML Repository of patients with and without CKD, Rao et al.⁶ developed a fusion model using GNN and traditional tabulated data to predict CKD development. Here, each patient was a node on the graph, with edges being relationships between patients based on feature similarities (hypertension, diabetes, obesity, anemia, consumption of unhealthy foods, excessive drug use, self-medication, exposure to water pollution, and history of kidney stones). They compared the performance of the fusion model with a convolutional neural network, which is a DL method that uses input data from a grid-like interconnected matrix then connected to downstream layers (Supplemental Table 2). They found that while including tabular information enhanced overall performance (accuracy of 95.1%), GNNs alone could achieve most of this (92.9%). By contrast, combining the tabular model with a convolutional neural network (CNN) achieved only 90.9% accuracy, highlighting the value of the relational structures learned by GNNs in biomedical tasks.

These performance gains are consistent across applications. Yin et al. used multi-instance learning on kidney ultrasound data from children with mild hydronephrosis (no CKD) versus children with congenital abnormalities of the kidney and urinary tract. When compared with a multi-instance learning with CNN-based instance-level classification, GNNs significantly increased mean accuracy (84.9% from 64.8%), mean sensitivity (85.8% versus 59.7%), and mean specificity (83.8% versus 71.4%).⁷ GNNs have also been used to develop a clustering-based spatial analysis model to extract key visual features from whole slide images of kidney biopsies. Predictions of kidney functions based on biopsy features derived by this framework had an accuracy of 94% when predicting function at the time of biopsy and 84% accuracy at predicting function 1 year postbiopsy. GNNs are ideally suited for histopathology tasks as individual features are best interpreted in the context of their surroundings (spatial relationships),⁸ while CNNs cannot encode this contextual information.

Although we have suggested a critical role for GNNs, and geometric learning broadly, in artificial intelligence–based CKD applications, their full potential in accelerating research will be best realized with large-scale, diverse, and multimodal data. The Kidney Precision Medicine Project⁹ comprises one such dataset with demographics, comorbidities, longitudinal clinical data, biopsies, peripheral, and tissue-specific omics data for patients with CKD due to diabetes and hypertension. However, these datasets with multimodal data are often limited by sample size (Kidney Precision Medicine Project has only 350 patients with CKD as of April 2025).⁹ Recent developments in graph-based federated learning allow for patients at different institutions to be inserted into a common and anonymized decentralized graph, increasing data volume and interinstitutional harmonization.¹⁰ This unique aspect of geometric learning can therefore accelerate progress in key research objectives toward individualized CKD treatment. Geometric DL holds the potential to establish a new paradigm in precision medicine, improving outcomes for the millions affected by CKD worldwide.

Disclosures

Disclosure forms, as provided by each author, are available with the online version of the article at http://links.lww.com/JSN/F258.

Funding

None

Author Contributions

Conceptualization: Girish N. Nadkarni, Madhumitha Rajagopal.

Funding acquisition: Girish N. Nadkarni.

Methodology: Justin Kauffman, Girish N. Nadkarni.

Supervision: Girish N. Nadkarni.

Writing – original draft: Justin Kauffman, Girish N. Nadkarni, Madhumitha Rajagopal.

Writing – review & editing: Justin Kauffman, Girish N. Nadkarni, Madhumitha Rajagopal.

Supplemental Material

This article contains the following supplemental material online at http://links.lww.com/JSN/F257.

Supplemental Table 1. A summary of conventional ML with associated algorithms, use cases, and estimated computing requirements.

Supplemental Table 2. A summary of different DL (artificial neural network) architectures describing information flow and examples of use cases.