Precision oncology has fundamentally transformed cancer research and treatment by adopting personalized genetic, molecular, and clinical approaches for individual patients [1,2]. A key principle of this paradigm is patient similarity, defined as the identification of patterns and commonalities among patients based on tumor genomics, disease progression, and therapeutic response [3]. Recognizing biologically similar patients enables clinicians to select targeted therapies more precisely, improving treatment outcomes and minimizing adverse effects [4,5].

The idea of learning from similarities between patients has deep historical roots. Since the time of Hippocrates, who systematically observed symptom patterns to guide diagnosis and treatment, comparing similar cases has been a cornerstone of medical practice [6]. Over the centuries, this principle evolved from purely clinical observations to sophisticated molecular and genomic analyses, ultimately forming the foundation of modern precision medicine [7,8].

In oncology, the concept of patient similarity expanded notably in the 1970s and 1980s, when researchers began classifying cancers into subtypes based on histopathological features. This approach recognized significant heterogeneity in tumor behavior, prognosis, and treatment responses even within cancers originating from the same tissue [2,3]. The 1990s marked a turning point with the advent of molecular profiling techniques, such as gene expression analysis, which allowed for more precise subtyping and highlighted the relevance of molecular similarity in guiding therapeutic strategies [9].

The clinical utility of molecular-based patient similarity was exemplified by the development of targeted therapies. Landmark successes, such as trastuzumab for HER2-positive breast cancer and imatinib for chronic myeloid leukemia, demonstrated that patients sharing specific genetic profiles could benefit from tailored treatments [10,11]. These advances firmly established the importance of integrating individual molecular characteristics into treatment planning, laying the groundwork for personalized oncology [4,12]. More recently, large-scale genomic initiatives such as The Cancer Genome Atlas and the International Cancer Genome Consortium have enabled comprehensive comparisons of genetic and molecular profiles across diverse patient populations [13]. Progress in bioinformatics and computational methods throughout the 2010s further facilitated detailed analyses of multidimensional biological data, including genetic mutations, epigenetic modifications, gene expression patterns, and features of the tumor microenvironment [14-17]. These technological advances have greatly enhanced our ability to stratify patients based on complex molecular characteristics. However, despite these achievements, consistently defining and applying patient similarity in research and clinical practice remains a major challenge. Current approaches often rely on heterogeneous criteria, ranging from specific genetic mutations to broader clinical phenotypes, leading to inconsistent findings and complicating the translation of results into clinical decision-making [18]. In addition, the integration of diverse data types (eg, genomic, clinical, and imaging) poses significant technical and methodological difficulties [14]. Furthermore, the lack of consensus on how to determine patient similarity, due to the use of varying algorithms and models, leads to inconsistent findings [15]. Moreover, the inherent heterogeneity of cancer, including variations within the same tumor type, makes it even more difficult to identify similar patients [19].

To address these challenges, further research is needed to develop standardized metrics and methodologies for assessing patient similarity, which could open crucial research opportunities in precision cancer care. To underline this need and to provide a current understanding and application of patient similarity in the context of cancer, a scoping review was conducted. This review aims to offer a comprehensive understanding of the current applications and methodologies of patient similarity in cancer research, with the hypothesis that a multilayered perspective is essential for enhancing precision oncology and improving patient outcomes.

This scoping review focuses on the concept of patient similarity in oncology across several interrelated aspects. It investigates methods and approaches for identifying and analyzing patient similarities by exploring diverse data types, such as genetic, molecular, and clinical information, to uncover patterns among patients with cancer. The review highlights the most commonly studied cancer types, identifies research trends, and pinpoints overlooked areas requiring further exploration. Despite growing interest in patient similarity, no prior scoping review has systematically examined both the qualitative and quantitative dimensions of this concept. While individual studies propose various definitions and methodological approaches, a comprehensive overview of these perspectives is lacking. Furthermore, the field lacks a synthesis of the most frequently investigated aspects of patient similarity, such as commonly used data types, analytical techniques, and dominant research trends. This review addresses these gaps by providing an integrated analysis of how patient similarity is defined, measured, and applied in oncology, offering insights that may enhance its clinical utility.

This scoping review was conducted in accordance with the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) checklist and the Joanna Briggs Institute reviewer’s manual [20,21], with all steps guided by a preestablished protocol [22]. These provided a structured framework that ensured transparency and consistency throughout the review.

Our primary goal was to include a broad range of studies to provide a comprehensive overview of research on similarities in patients with cancer. To ensure quality and relevance, we established specific inclusion and exclusion criteria as predefined in our protocol.

Studies were included if they provided substantial evidence on similarities in patients with cancer across diverse populations, ages, genders, and cancer types. We limited the review to studies published within the last 25 years and written in English or German, the working languages of our team, to facilitate thorough analysis and to focus on contemporary developments relevant to current cancer research. Studies were excluded if they did not address similarities among patients with cancer, were bachelor’s or master’s theses, were unpublished manuscripts, or focused on noncancer conditions or animals. Pilot testing on a sample of studies helped refine the criteria, enabling the selection of the most relevant studies for our scoping review.

Our search strategy used a triangulated approach, combining keyword searches, snowball sampling, and manual review for a comprehensive literature assessment. Details regarding the search strategy, including the specific search queries for the databases, are outlined in the protocol [22]. Final searches were completed on February 7, 2024.

The selection process followed a rigorous 3-stage approach—title-abstract screening, disagreement resolution, and full-text screening—aligned with the recommendations by Tricco et al [20]. To ensure quality, each stage included pilot testing with calibrated forms for standardization [21,22]. Publications were independently reviewed by at least 2 blinded reviewers using predefined eligibility criteria. Disagreements were resolved by an independent reviewer, with all discrepancies documented for transparency [21]. Screening was conducted using Rayyan (Rayyan Systems Inc), a web-based software [23], and outcomes were summarized using the PRISMA-ScR flowchart.

In this scoping review, the research team manually extracted data using a predefined, pretested template designed to align with the research questions. This approach ensured systematic data capture, consistency across studies, and minimized errors. Pilot testing confirmed the template’s robustness and accuracy, and the original templates are available in the protocol without modifications.

The analysis process followed a structured series of standardized steps, including data preprocessing, categorization, methodological classification, and quantitative and qualitative analyses. All data generated or analyzed during this study are included in this paper and Multimedia Appendix 1 [15,18,24-158].

Figure 1 illustrates the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flowchart [20], detailing the process of narrowing down 3298 initially identified records to the 137 studies ultimately included in the review. Records were sourced from PubMed (637/3298, 19.31%), MEDLINE (484/3298, 14.67%), LIVIVO (903/3298, 27.38%), and Web of Knowledge (924/3298, 28.02%), as well as snowball sampling (95/3298, 2.88%) and manual searching (255/3298, 7.73%). The selection process consisted of 3 key stages: title-abstract screening, disagreement resolution, and full-text screening. The initial steps included removing 2066 (62.64%) of the 3298 duplicate records, followed by the exclusion of 1.48% (49/3298) of records because of publication issues, such as mismatched publication year or language. This resulted in 35.87% (1183/3298) of the records eligible for screening. During the title-abstract screening, 78.19% (925/1183) of the records were excluded for not meeting the inclusion criteria, leaving 258 (21.81%) papers for full-text screening. In this final stage, a further 46.9% (121/258) of the records were excluded, culminating in 137 (53.1%) studies included in the review (refer to Multimedia Appendix 3 for more information).

The triangulated approach to identifying, selecting, and retrieving studies ensured a robust and comprehensive review, including only the most relevant and high-quality studies. These systematically categorized studies provide a solid foundation for both quantitative and qualitative analyses, addressing key research questions on patient similarities across definitions, methodologies, data sources, and cancer types. The analysis of the resulting 137 studies provides critical insights into patterns of patient similarity, creating a comprehensive understanding of the current state of cancer research and care.

As specified in one of the research questions of this scoping review, data sources play an important role in defining patient similarity in cancer research. These sources can be seen as a distinct dimension of patient similarity or as part of an interconnected framework encompassing 4 dimensions that collectively define this concept. Given the complexity of cancer and its numerous influencing factors, a wide variety of data sources is used. To streamline this diversity, we categorized the data into 6 broad groups: genetic and genomic data, clinical data, transcriptomic data, proteomic and metabolomic data, epigenetic data, and 1 functional category—pathway and network data. Table 2 presents the classification of data types with examples, while Multimedia Appendix 4 provides a visualization of their distribution.

It is important to note that data types in cancer research are rarely used in isolation; instead, they are often combined to achieve a more comprehensive and meaningful representation of patient similarity. The choice of data types to be combined depends largely on the research objectives and the methodologies applied. Therefore, we analyzed the data from different perspectives, considering not only their absolute abundances but also the frequencies of specific combinations (Figure 4). The most frequently used data type is transcriptomic data, accounting for 67.1% (92/137) of the studies. This category includes expression data for microRNA, messenger RNA, and exons [85,173,174], as well as gene signatures and microarray data [73,86], reflecting its critical relevance in cancer research, particularly in biomarker discovery and subtype identification [53-55,73,86]. These data are most often used in combination with other types, particularly clinical data (37/137, 27%) [53,174] and epigenetic data (34/137, 24.8%) [54,55]. The frequent overlap between transcriptomic and clinical data reflects the relevance of gene expression patterns in relation to patient-specific clinical features, providing molecular insights that complement phenotypic observations. Similarly, transcriptomic data are often combined with genetic and genomic data (30/137, 21.9%; Figure 4), highlighting the influence of genetic alterations, such as mutations and copy number variations, on downstream gene expression. This integration enhances the understanding of molecular mechanisms driving cancer progression and underscores the importance of combining these data types for comprehensive similarity analyses. Clinical data, the second most common category, is highly heterogeneous, limiting the interpretive significance of its frequency alone. In addition to image data, this category includes data on the activity of chemical compounds [56], tumor size, and blood counts [5]. The methodological approaches described in the associated papers are often explicitly tailored to a data group in this category. While clinical data are often used in combination with transcriptomic data (Figure 4), they also appear as the most frequent stand-alone category (20/137, 14.5%). The versatility of clinical data lies in its ability to provide a contextual framework for other molecular data types, despite the analytical challenges posed by heterogeneity [56,87]. Another important category is genetic and genomic data. This includes studies investigating copy number variations [57], gene interactions [88], and molecular profiles [49]. With a frequency of 35% (48/137), this data group is the third most common and often appears in combination with clinical data (20/137, 14.6%) [44,49] and transcriptomic data (30/137, 21.9%) [28,29,59]. This may be because genetic alterations, such as copy number variations and mutations, can directly influence gene expression. For example, increased gene copy number (amplification) can lead to overexpression of an oncogene, which promotes tumor growth [63]. Therefore, when both types of data are considered together, the chance of obtaining a more detailed picture of mutational influences for similarity analysis is significantly higher. Epigenetic data, which account for approximately 27% (37/137) of the studies, are mainly used in combination with transcriptomic data (34/137, 24.8%; Figure 4) [54] and are rarely used as a stand-alone data source (2/137, 1.4%) [64]. DNA methylation profiles, which are part of this category, are key epigenetic features that influence cellular phenotypes and patient similarity. However, the impact of methylation often becomes more apparent when integrated with other types of molecular data, highlighting the complex interplay between epigenetic and genetic factors [65]. Proteomic and metabolomic data (10/137, 7.3%) and pathway and network data (18/137, 13.1%) make up the smallest proportion of the types of data sources used. These include proteomics [66], metabolomics [67], and mass spectrometry data [68], as well as PPI data [163] and pathway aberration profiles [164]. Notably, pathway and network data are only used in combination with other data types (Figure 4). In other words, they serve in a supporting or augmenting capacity but are not used independently to determine similarity. However, very innovative approaches in this data category have led to progress, especially in the areas of simulation and prediction of survival [163,164]. Quantitatively, transcriptomic data, that is, expression data, were analyzed particularly frequently, while genetic and genomic data were studied significantly less often. This may be because of the higher clinical relevance of expression data [61]. This prioritization may indicate that transcriptomic data are considered a more direct indicator of disease status and treatment response. In addition, the methodological approaches for the analysis of expression data are often more mature and better established, which supports their frequent use. In contrast, genetic and genomic data provide deeper insights into the underlying mechanisms of carcinogenesis but are often only fully interpreted in combination with other data types. The heterogeneity of clinical data underscores their versatility but also makes them difficult to analyze in a consistent manner. The variety of data sources used reflects the complexity of cancer research, which requires a multidisciplinary approach.

A review of the types of data commonly used in similarity studies involving patients with cancer emphasizes the intrinsic link between data types and methodological approaches. The classification of methodological aspects of patient similarity is inherently dependent on the type of data used. The type of data directly affects how similarity can be calculated and interpreted. This relationship can be understood by examining the underlying principles of methods, their fields of application, and the specific types of data they use. Consequently, it is important to analyze the interaction between methodological categories and the data used in the context of patient similarity, highlighting the strong correlation between data parameters and methodological parameters. After examining this interrelation, we conclude that machine and deep learning methods are most prominently applied to transcriptomic, clinical, and genomic data. This analysis also shows that network- and graph-based approaches are applied to the same data as machine learning models, while slightly favoring pathway and interaction data.

To analyze the cancer types that have been frequently studied in relation to patient similarity, we evaluated the number of papers using datasets from specific cancer types. Many papers, especially those focused on methodological innovation, use large and diverse cancer datasets that include >5 major cancer types [24,33,89,90]. However, several papers focus extensively on 1 or 2 major cancer types, or even on 1 or 2 cancer subtypes [40,91-93]. These studies often use methods specifically designed for these entities to obtain specific information [40,91-93].

All these publications share 1 commonality: several main cancer types and subtypes appear frequently, either as part of a larger dataset or on their own. We categorized the cancer types by primary tumor site using OncoTree [159]. We identified 5 commonly reported primary tumor sites: kidney, bowel, central nervous system, that is, brain, lung, and breast, as outlined in Table 3.

With a conspicuously large quantity of 52.5% (72/137) mentions in different publications [28,66,94,175], breast cancer is the most studied cancer regarding patient similarity. This is logically consistent with the fact that breast cancer is also widely studied in different cancer research fields because of its high prevalence in the population [176] and the abundance of well-categorized data available [177]. The gap between breast cancer and the second most studied type, lung cancer (38/137, 27.7%) [30,41,95,96], underscores the prominence of breast cancer in cancer research. This is further confirmed when looking at specific cancer subtypes, where breast invasive carcinoma (BRCA) is generally the most researched subtype (49/137, 35.8%), although in most cases, the specific BRCA type was not mentioned in the publications [24,25,97,98]. Interestingly, the second and third most studied specific cancer types are glioblastoma multiforme (GBM) and renal clear cell carcinoma (clear cell renal cell carcinoma). GBM was studied in 20.4% (28/137) of the publications [26,42,57,99] and represents the majority of brain cancers studied. The same can be observed with kidney renal clear cell carcinoma or clear cell renal cell carcinoma: 83% (24/29) of the total mentions of bowel cancer refer to this cancer [25,34,41,100].

The specific lung cancers studied are divided into 2 main types—lung squamous cell carcinoma (LUSC) and lung adenocarcinoma—with LUSC being more common, with 16% (22/137) mentions, than lung adenocarcinoma, with only 8.7% (12/137) mentions [30,41,96]. In bowel cancer, colorectal adenocarcinoma was the most commonly reported specific type with 13.9% (19/137) mentions [32,34,41,90]. It is also striking that ovarian serous cystadenocarcinoma and acute hematologic cancers (eg, lymphoid and myeloid cancers), which are less common than some other cancers but have been the focus of significant research, are also listed (15/137, 10.9% and 8/137, 5.8%, respectively). However, in most cases, they were only part of a larger dataset that included >5 different cancers (11 of 15 and 6 of 8, respectively), which relativizes the true importance of the findings in this context.

Overall, the most commonly studied cancer types regarding patient similarity are breast cancer (especially BRCA), followed by central nervous system or brain cancer (especially GBM) and lung cancer (especially LUSC). Including kidney and bowel cancer, the top 5 primary affected tumor sites account for >50% of the total number of mentions of cancer types in the studies.

We identified several similarity metrics frequently used in cancer research to compare patient data. Patient similarity is context dependent, influenced by specific methods, goals, and expert perspectives [101,102]. Common approaches represent patients as vectors in feature spaces, with similarity expressed as a distance metric [5,12,18,178]. The multidimensional feature spaces are often divided into different data categories that are considered differently in the similarity calculation, such as numerical data, binary data, and classification data or structured data (eg, gender and age) and unstructured data (eg, diagnostic texts) [12,102]. Using this data concept, patient similarity can be understood as the distance between 2 vectors in the feature space, which can be intuitively described as the edge weight on the edge between 2 nodes in a graph, as is the case in methods such as the PSN [60,88,101,179]. Depending on the approach, it is possible to examine all features together or to evaluate multiple features individually and ultimately merge them [5]. This representation can be used to define distances between the vectors of individual patients using mathematical methods.

There are several distance or similarity measures that are used to compute similarities. The most common metrics are the Euclidean distance [27,60,87,88,101,179], cosine similarity [18,60,87,88,101,103,178,179], Jaccard similarity [87,88,101,178], and Pearson correlation [60,85,88,101,103,104,179]. Cosine similarity calculates the cosine of the angle between 2 patient vectors in feature space. The calculation can be simplified using the dot product described by Brown [87]. The result is in a range between −1 and 1, where 1 is the maximum similarity [87,178]. The Jaccard similarity measures the similarity of 2 sets compared to each other. It describes the coefficient of the size of the intersection and the size of the union of 2 sets. The result ranges from 0 to 1, where 1 represents the optimal similarity between 2 patients. This concept can be used to compare symptom phenotypes of patients [27,173]. The Pearson correlation (or Pearson correlation coefficient) is a measure of the linear relationship between 2 variables. It ranges from −1 to 1, where 1 indicates a perfect positive correlation, −1 indicates a perfect negative correlation, and 0 indicates no correlation [104,174]. This similarity metric is particularly robust to missing data within the patient vectors, making it an ideal tool for analyzing incomplete datasets [60,179]. Other similarity measures, such as Hamming, City Block, Minkowski, and Chebyshev distances, are also mentioned [88,101]. Overall, the specific measures are usually the basis for individually defined distances, which are also adapted to the data sources used. For example, when using gene and microRNA expression levels, the distance between 2 samples can be calculated using the Euclidean distance in combination with other metrics [18,27,87,103]. However, depending on the type of data available, new distance measures may need to be used. For example, Virgolin et al [86] describes distance measures that focus on the physical evaluation of image data. They define different similarities based on organ deformation, organ overlap, organ shape, and organ constellation. Even these approaches often use known coefficients and metrics, such as the Hausdorff distance, as a starting point [86].

When creating a similarity score, it is very important to consider the characteristics of the data and to optimize the metric according to the underlying goals. For example, if the goal is to identify driver genes and compare patients for similar mutations, the metric must be able to represent the similarity of mutated genes in patients [105]. In this specific case studied by Zhang et al [73], the similarity of the Gaussian interaction profile kernel was used to score functional similarity. When numerical data alone are not available, as is often the case with clinical data, ways must be found to incorporate both continuous and binary features. This can be done by defining a similarity measure for each data category [5] and then scoring them individually, or by using metrics such as the Gower similarity coefficient for the categories and merging them using the geometric mean calculation [88]. However, the more diverse the data types, the more difficult it becomes to make an all-encompassing statement limited to 1 value, which is why the development of methodological approaches for determining similarity between patients plays such an important role. Overall, defining patient similarity in cancer research involves understanding the interplay between different clinical, biological, and treatment data types. The complexity and heterogeneity of cancer underscore the need for a multifaceted approach, ensuring that similarity metrics adequately reflect the nuanced relationships between patients.

Our ongoing analysis has identified several key objectives frequently applied to define similarity in patients with cancer, as shown in Multimedia Appendix 5. Most of our publications addressed the identification of cancer subtypes, with 51.1% (70/137) of the publications focusing on this objective. As cancer is a highly heterogeneous disease, there are many subgroups that may respond differently to various therapies, allowing different treatment methods to be chosen [24,180]. Patients within a subtype are, by definition, similar and can therefore benefit from the treatment outcomes of other patients. To identify these subtypes, machine and deep learning methods (44/70, 63%) as well as network- and graph-based approaches (46/70, 66%) were used predominantly, as shown in Figure 5 [24,25,27,31].

The second most common aim of the analyzed publications was biomarker discovery [46,49,64,65,72,97,106], which was the focus of 29.9% (41/137) of the publications. According to the National Cancer Institute, biomarkers are biological molecules that serve as indicators of normal or abnormal processes or diseases. They play an important role in the diagnosis and treatment of patients with cancer [181], and the discovery of new biomarkers is crucial for improving the success rate of therapies [68]. The transition between the identification of subtypes and biomarkers is fluid, as subtypes are often defined by specific biomarkers. This explains the significant overlap detected between these 2 patterns of similarity identification (Figure 6). A notable overlap was observed between subtype identification and the measurement of patient similarity, as subtypes often reflect patient-specific features that define similarity patterns. This connection highlights the interdependence of these 2 concepts in patient stratification efforts. The distribution of machine and deep learning, as well as network- and graph-based approaches, shows a relatively balanced representation. Similarly, a notable overlap was identified between subtype identification and data integration, which is logical, as integrating diverse data types often aids in identifying complex subtype-specific characteristics. Moreover, overlaps between subtype identification and simulation or prediction models were prominent, reflecting the reliance of predictive analyses on subtype-specific features (Figure 6). Predictive models frequently used these features to predict survival outcomes or drug responses. In addition, overlaps were observed between simulation and prediction models and survival analysis, indicating that survival outcomes are often subtype dependent (Figure 6). Simulations and prediction models were also frequent, with 27% (37/137) of the publications using them. In particular, machine and deep learning approaches [48,49,54,169] were used (23/37, 62%; Figure 5). Different approaches to predictive models were used. For example, in some cases, survival time was predicted [38,55,69,107]. Other prediction models included drug response prediction [48,62,80,108] and diagnosis prediction [53]. Because predictive analyses frequently rely on subtype-specific features, a notable overlap was identified between subtype identification and simulation or prediction models (Figure 6). Measurement of patient similarity emerged as another crucial area of research, with 25.5% (35/137) of the publications focusing on this aspect. A clear overlap with subtype identification was detected, as subtypes are often defined by shared similarity metrics across patient groups. Simulations and prediction models were followed at some distance by pathway analysis, but with considerably lower frequency, with 16.8% (23/137) of the publications. Many of these publications focused on the identification of cancer-related pathways [74,92,109]. Data integration and interoperability were addressed in 16% (22/137) of the publications, while clinical decision support was the focus in 3.6% (5/137) of the publications. Furthermore, overlaps were observed between pathway analysis, subtype identification, and survival analysis, emphasizing the role of pathway-specific mechanisms in subgroup survival outcomes. As shown in Figure 6, overlaps were also observed between pathway analysis and both biomarker discovery and subtype identification. This indicates the importance of pathway analysis in identifying biomarkers and understanding mechanisms within subgroups.

This scoping review offers a comprehensive synthesis of how patient similarity is defined, operationalized, and applied in current cancer research. The analysis identified 4 central dimensions that collectively shape the conceptualization and implementation of the patient similarity model: research objectives, analytical methods, cancer type representation, and data availability. Rather than representing a fixed or uniform metric, this study highlights that patient similarity is a dynamic construct shaped by the interaction of multiple elements and influenced by methodological choices and clinical intent.

A key finding of this review is the multifaceted nature of patient similarity in oncology. The analysis of this multidimensional modeling approach revealed that similarity is applied across a range of research objectives, including subtype identification, biomarker discovery, prediction modeling, survival analysis, and pathway analysis. Among these, subtype identification and biomarker discovery were the most prevalent, appearing in more than half of the reviewed studies. This emphasis highlights the central role of patient similarity in optimizing stratification and informing therapeutic interventions, an imperative that aligns with the heterogeneity of cancer and the need for individualized treatment strategies. This focus can be explained by the fact that identifying molecular subtypes and actionable biomarkers remains a foundational step in precision oncology. It allows for the classification of patients into clinically relevant categories that may respond differently to therapies, thus improving outcomes and resource allocation [24,25,46,49,64].

Beyond identifying research objectives, this review provides insight into how patient similarity is measured and conceptualized. It was observed that the selection of similarity metrics varied significantly across studies, depending largely on specific research aims and the type of data available. Because no single metric can universally capture the complexity of interpatient relationships, metric selection must be purpose specific. Some metrics emphasize molecular proximity, while others are better suited to capture phenotypic or clinical resemblance. This variability underscores the need for methodological transparency and for tailoring metric selection to the specific context of the study.

This heterogeneity in metric use was mirrored in the analytical methods used. Machine and deep learning approaches were used in more than half of the studies, marking a significant methodological trend. These methods offer notable advantages for processing large-scale, high-dimensional datasets, facilitating predictive modeling and identifying complex, nonlinear patterns that may not be accessible via traditional statistical techniques. This suggests that the growing adoption of machine learning is a response to both the scale and complexity of omics data in oncology. In particular, clustering techniques, such as spectral and hierarchical clustering, were frequently used to define patient subgroups based on genetic and clinical features, supporting the hypothesis that certain molecular signatures may correlate with treatment outcomes [24-39]. Despite their strengths, machine learning methods often lack interpretability, which remains a significant challenge in their application to clinical practice. These models are also prone to overfitting, performing well on training data but failing to generalize to new cases. In addition, imbalanced test-to-training ratios in oncology datasets can introduce bias, further limiting their clinical reliability.

Network- and graph-based approaches were also prominently represented in the reviewed studies. That can be explained by the fact that these methods are especially important when modeling the relationships and interactions among biological entities and when working with data that encode complex dependencies. It was shown that techniques such as PSNs and PPI networks provide an intuitive way to represent patient similarities in the context of shared genetic or clinical features [28,29,49,53-59]. Nevertheless, network-based models are also sensitive to data weighting and metric definitions, which can significantly alter network topology and affect reproducibility. This reinforces the importance of standardizing methodological choices and validating models across diverse datasets.

The integration of different analytical approaches, particularly in hybrid models, shows promising potential for future research. For instance, combining GCNs with statistical validation techniques represents an emerging strategy for enhancing both accuracy and interpretability [53,54,61,69,70]. These hybrid approaches may offer a path forward in overcoming the limitations of individual methods while leveraging their respective strengths.

In contrast, our analysis showed that methodological choices were rarely arbitrary. Instead, they were closely aligned with research objectives: clustering was typically used for subgroup identification, deep learning and GCNs for prediction, and multimodal integration for stratification and exploratory biological analysis. These patterns suggest that patient similarity is best viewed not as a universal measurement but as a flexible framework tailored to specific goals.

In addition, the analysis of data dimensions brought additional insight by highlighting that transcriptomic, clinical, and genetic and genomic data are the most frequently used for assessing patient similarity [182]. The high use of transcriptomic data can be attributed to its direct relevance in understanding gene expression patterns that characterize cancer heterogeneity and its abundance. The integration of multiple data types, such as combining transcriptomic and clinical data, emerged as a significant trend, reflecting the need for comprehensive data to fully capture patient similarity.

Finally, the frequent study of breast, lung, and brain cancers in the context of patient similarity underscores the prevalence of these cancers and the availability of well-characterized datasets [25,66,94,175]. However, this focus also points to gaps in the study of less common cancers, highlighting opportunities for future research to explore underrepresented cancer types.

In comparison with previous literature, this scoping review provides a substantial advancement in the conceptual and methodological understanding of patient similarity by focusing specifically on its application in oncology. Earlier reviews, such as the one by Sharafoddini et al [183], primarily examined patient similarity in the context of prediction models based on electronic health record data across various medical domains. Their focus remained largely on structured clinical features and statistical similarity metrics, without a disease-specific emphasis. Similarly, the systematic review by Parimbelli et al [1] explored patient similarity within the broader framework of precision medicine, outlining its relevance to predictive modeling and treatment stratification, though without a dedicated focus on oncology.

In contrast, this review focuses exclusively on cancer research and provides a multidimensional synthesis that integrates biological, methodological, and clinical perspectives. This disease-specific emphasis enables a more detailed analysis of how patient similarity is conceptualized and operationalized in the context of tumor heterogeneity and precision oncology.

A major point of distinction lies in the methodological approaches reported. While previous studies largely discussed traditional distance-based similarity metrics and simple clustering algorithms, our review highlights the widespread adoption of more advanced computational techniques. These include, among others, GCN, SNF, and federated learning models, which allow the integration of heterogeneous data types while preserving patient privacy and ensuring model robustness. This evolution reflects the increasing sophistication of oncology research and the growing need for methods capable of handling high-dimensional multiomics datasets [15,52,53,165].

Moreover, our review extends prior work by emphasizing the interdependence between data types and methodological frameworks. Although previous literature focused primarily on structured clinical or demographic data, the findings here underscore the central role of transcriptomic, genomic, and multiomics data. These data types are often analyzed using machine learning and network-based approaches to uncover complex biological patterns [49,55].

Importantly, while earlier reviews identified patient similarity as a promising concept, they mostly discussed its potential applications. In contrast, our findings demonstrate that patient similarity is already being actively applied in oncology for critical tasks such as cancer subtype identification [39], biomarker discovery [49], and predictive modeling of survival or treatment outcomes [55]. These applications signify a clear transition from theoretical potential, as discussed in earlier work, to real-world implementation in precision oncology.

Despite its strengths, this scoping review has several limitations.

First, it relies on published literature, meaning the findings are constrained by the scope and quality of the included studies. Potential biases such as selective reporting and variability in study quality may have influenced the results. In particular, publication bias poses a concern: studies reporting positive findings, especially those successfully identifying patient similarities, may be overrepresented, while studies with negative or inconclusive results (eg, in specific cancer types) might be underrepresented. This pattern is theoretically plausible, given the tendency to publish predominantly significant findings [1,183]. Although we used a comprehensive search strategy across multiple databases, we did not include gray literature, which may further limit the breadth of the included evidence. Consequently, the inherent biases of published data remain a challenge. Future research should aim to incorporate unpublished studies and negative findings to provide a more balanced and comprehensive perspective.

Second, we limited the temporal scope to research published within the last 25 years, which could have led to the omission of earlier foundational studies.

Third, the focus on common cancers (eg, breast, lung, and brain) limits the generalizability of the findings to rarer cancers, where patient similarity remains underexplored [39,49,53,101]. In addition, synthesizing findings across studies using different definitions and metrics of patient similarity was challenging. The lack of a universally accepted definition led to inconsistencies in how similarity was measured, ranging from clinical characteristics to molecular profiles. This heterogeneity introduced biases and limited generalizability. In particular, studies relying on different data types, such as clinical records versus genomic data, varied in their ability to make strong claims about patient similarity. Consequently, findings from certain methodological approaches may carry more weight than others, introducing an inherent imbalance in the available evidence. Furthermore, the variability in methodologies, from machine learning to network-based and statistical approaches, hinders standardized clinical recommendations [24,25,28,46,182].

Finally, while the review emphasizes the importance of integrating multiomics and clinical data, it does not delve into practical guidance for addressing challenges such as data heterogeneity and interoperability, as this was beyond its scope. However, emerging solutions, such as data harmonization efforts and federated learning approaches, are increasingly being explored to tackle these issues, offering avenues for future research and application [52,61,69].

This scoping review provides a comprehensive overview of the current understanding of the concept of patient similarity in cancer research and treatment.

The findings indicate that patient similarity is best conceptualized as a relational construct, reflecting the specific objectives, data constraints, and clinical priorities of each study. Rather than representing a single metric, it constitutes a multidimensional modeling paradigm, shaped by research goals, data architecture, cancer-specific contexts, and methodological design. Its successful application hinges on purpose-driven model development, rigorous validation, and the thoughtful integration of heterogeneous data sources.

To further advance the field, future research must emphasize standardization, including the development of clear and consistent criteria for defining, measuring, and evaluating similarity. In parallel, efforts should focus on expanding available data diversity and designing models that are both biologically grounded and clinically interpretable. These advancements will support the broader clinical integration of patient similarity and catalyze progress toward personalized oncology, enabling more precise, effective, and individualized cancer treatments that reflect the unique characteristics of each patient and improve outcomes across diverse populations.