A patient similarity network (PSN) is a mathematical model based on graph theory that allows for the visual and analytical exploration of complex relationships among individuals. In this graph representation, nodes represent single patients, while edges measure the degree of clinical, phenotypic, or molecular similarity between them.
PSNs are increasingly utilized in clinical research to support clinical decision-making, shifting the focus from general statistical averages to highly personalized treatments, the core concept of precision medicine. By employing machine learning techniques, these networks enable the prediction of patient outcomes and improve patient stratification, facilitating the identification of novel subtypes in complex diseases.
The ability to map each patient in relation to others introduces a paradigm where individual complexity is treated as unique, with specific characteristics and therapeutic needs. Furthermore, these networks are intuitive and easy to visualize, offering a natural way to inspect complex, high-dimensional data. This confers explainability for the predictions generated by machine learning models, while effectively maintaining the privacy that each patient deserves.
History and context
Historically, clinical practice and medical research have relied on a "one-size-fits-all" model, which was based on generalized treatments designed for the "average patient". In this paradigm, a standard therapy was prescribed to everyone sharing the same disease label. While this approach has proven successful for many common conditions, it fails to account for the inter-individual variability and underlying molecular differences among patients. Consequently, commonly prescribed drugs may only benefit a small fraction of patients, leaving others as non-responders or even causing adverse side effects.
Accounting for these unique features is especially critical in complex diseases which arise from the interplay of multiple biological and environmental factors. To address these limitations, modern healthcare is shifting toward precision medicine. This approach aims to fragment broad diagnostic categories into specific subtypes, tailoring diagnostics and treatments to each patient's unique genetic, molecular and environmental profile. This conceptual shift coincides with the post-genomic era and the advent of high-throughput technologies, which generate massive volumes of heterogeneous patient data at high velocity. Precision medicine relies on integrating these "Big Data", acquired from electronic health records (EHRs), wearable devices, biomarkers, and genetic data. As researchers collect increasingly large datasets combining biological, clinical, and environmental information, advanced computational tools are required to process them.
To manage and interpret this scale of data, artificial intelligence and network-based methods have become essential. PSNs were developed within this context as a tool to integrate large amounts of heterogeneous data. By grouping patients based on shared clinical and molecular traits, PSNs enable the identification of highly specific and precise patient subgroups, providing a solid foundation for individualized care.
Network construction procedure
The construction of a patient similarity network involves transforming individual patient profiles into a mathematical graph. In this structure, each patient is represented as a node, and the similarities between their medical or biological profiles are represented as weighted edges. This process generally follows four main steps: data collection, similarity calculation, data integration, and graph sparsification. Data collection The first step for the construction of the network involves gathering features that could accurately describe the clinical and biological profile of a patients. In order to have a complete understanding of the different subtypes, precision medicine relies on highly heterogeneous datasets, which can be broadly divided in the following categories:
Electronic health records Electronic health records (EHRs) are digital archives including structured clinical data, such as demographics, vital signs, laboratory test results, prescribed medications and formal diagnoses, as well as unstructured data such as clinical notes. Analizing EHRs often represents a challenge, due to the irregular and incomplete nature of data. However, they can be useful to researchers to track clinical trajectories over time about the evolution of the patient state. This makes it possible to map not only a patient's current health but also their entire medical history and disease progression.
Multi-omics data
Advances in high-throughput molecular analysis have revolutionized our understanding of biological processes. The term "omics" refers to the holistic study of the molecules within a biological system, with each discipline focusing on a specific cellular aspect. The complexity of a biological system arises from the interactions among these various omics layers: genomics, epigenomics, transcriptomics, proteomics, and metabolomics. These omics layers can be analyzed individually or integrated.
Multi-omics integration yields a more comprehensive and precise patient molecular profile (taking into consideration both static and dynamic cellular aspects), enabling the identification of specific patient subtypes that may share the same phenotypic profile but are driven by different underlying biological processes. This generation of comprehensive data enhances the understanding of disease mechanisms and advances precision medicine approaches. However, integrating these high-dimensional data vectors introduces extreme computational complexity. This difficulty is primarily driven by magnitude imbalances across different omics layers and technical background noise dependent on sequencing platforms. Building a PSN with this data reduces background noise by focusing on the relationships between patients rather than just their raw measurements. Grouping patients this way captures their complete molecular profile, making predictions about survival or treatment responses much more accurate than models relying on a single omics layer.
Medical imaging and wearables
Advanced networks may also integrate features extracted from medical imaging modalities and continuous physiological monitoring from wearable devices, providing dynamic and real-time clinical information. Wearables (such as continuous glucose monitors, sleep quality trackers, and wrist-worn cardiovascular telemetry systems) generate massive time-series datasets. While these longitudinal trajectories successfully capture circadian variations and dynamic stress responses, they require robust signal processing to remove background noise and harmonize varying sampling frequencies across individuals. Furthermore, medical imaging serves as another crucial data source. By extracting specific quantitative features, such as dimensions, textures, and signal intensities, it is possible to uncover vital information regarding the patient's current clinical status and the phenotypic manifestation of the pathology.
Concurrently, medical imaging modalities like magnetic resonance imaging (MRI) and computed tomography (CT) provide critical anatomical context. By extracting specific quantitative features from these images (such as spatial dimensions, volumes, tissue textures, and pixel intensities) algorithms can accurately evaluate a patient's current physical status and the phenotypic manifestation of a underlying pathology. The fusion of real-time wearable telemetry with high-resolution anatomical imaging creates a comprehensive, dynamic clinical profile.
Similarity calculation After patient features have been selected, filtered, and transformed into numerical representations, the next step in the construction of a patient similarity network is the computation of scores of similarity between pairs of patients. This process converts complex and often heterogeneous patient profiles into similarity matrices, which represent the strength of the relationships between individuals and gives the basis for network construction and subsequent integration methods.
The choice of similarity metric is a critical step, as different measures capture different aspects of patient relatedness and can substantially influence the resulting network structure. For continuous variables, such as laboratory measurements Euclidean distance has traditionally been used to quantify overall differences between patient profiles, although its sensitivity to scaling and outliers often requires careful preprocessing such as normalization procedures. When dealing with transcriptomics profiles, while Euclidean distance can capture differences in the absolute expression levels, cosine similarity is frequently preferred for these high-dimensional and sparse datasets (as well for embeddings and graph neural network features), because it effectively measures the similarity of overall expression patterns independetly of tehir magnitude. Similarly, Pearson correlation distance is widely utilized to assess the linear relationship between continuous profiles, making it highly effective at capturing parallel trends (such as gene co-expression or longitudinal clinical trajectories) regardless of baseline shifts in the data. For binary data, such as the presence or absence of genetic mutations, clinical diagnoses, or adverse events, the Jaccard coefficient is commonly employed because it emphasizes shared characteristics while ignoring the large number of absent features that typically provide limited biological information. In clinical datasets containing mixed data types, such as electronic health records (EHRs), Gower similarity offers a flexible solution by combining different similarity calculations for continuous, categorical, and binary variables within a unified framework while also accommodating missing values.
These patient-by-patient similarity matrices can then be analyzed independently or integrated through network fusion approaches to generate a comprehensive representation of patient relationships across multiple biological and clinical domains.
Integration methods The construction of patient similarity networks from heterogeneous biomedical data requires the integration of multiple data modalities. Integration strategies can be broadly categorized into early, intermediate, and late fusion approaches, which differ in the stage at which data are combined and in the nature of the resulting representation. Within these paradigms, several methodological frameworks have been developed, including kernel-based learning, graph-based diffusion models, and ensemble-based aggregation techniques.
Early fusion strategies Early fusion methods represent the most straightforward integration approach. Data from various sources are concatenated into a single, high-dimensional vector before the training of any model. Similarity between patients is then computed in this unified space using standard distance metrics. A key advantage of early fusion is its ability to capture direct correlations across data types in a single representation. However, these methods typically require careful normalization to avoid scale dominance effects, where features with larger numerical ranges disproportionately influence the similarity estimation. Furthermore, early fusion is sensitive to missing data and may suffer from the "curse of dimensionality" when the number of features greatly exceeds the number of patients, increasing the risk of overfitting.
Representative approaches in this category include classical feature concatenation pipelines and dimensionality reduction methods such as principal component analysis (PCA) and matrix factorization, which aim to project heterogeneous data into a shared latent space prior to similarity computation.
Late fusion strategies