Feature-Centric AI for Breast Cancer Metastasis Analytics: A Systematic Review and Evidence-Based Framework for Feature Engineering and Multimodal Integration

Abstract

The importance of understanding how feature engineering contributes to breast cancer metastasis analytics is growing as artificial intelligence (AI) technologies rapidly transform the field of precision oncology. This systematic review collated, thematically synthesised, and analysed the evidence related to feature engineering processes, multimodal integration techniques and explainability mechanisms used in AI-driven breast cancer metastasis analytics. Following PRISMA 2020, studies from 2020 to 2026 were retrieved from MEDLINE, Scopus, Web of Science, Embase and IEEE Xplore. A total of 50 empirical studies investigating AI, feature engineering, radiomics and multimodal approaches for breast cancer metastasis prediction were included. The research findings reveal that contemporary AI-based breast cancer metastasis prediction is predominantly based on imaging-derived features, while clinical, pathological, biomarker and molecular variables increasingly enhance multimodal model development. Frequently applied techniques were observed across feature extraction, preprocessing, reproducibility assessment, dimensionality reduction, feature integration and explainability, all of which helped to create more interpretable, robust and clinically meaningful predictive systems. The study crafts an evidence-informed conceptual 5Fs framework comprising feature extraction or generation, feature preparation and transformation, feature selection and dimensionality reduction, feature construction and integration, and feature interpretability and explainability. The framework highlights the feature-centric nature of AI-driven breast cancer metastasis analytics and emphasises the importance of interpretable, multimodal and potentially clinically translatable predictive systems.

1. Introduction

Breast cancer remains the most commonly diagnosed cancer among women worldwide and continues to pose a significant public health challenge due to its high rates of incidence, recurrence and metastatic progression. Breast cancer caused about 694,000 deaths worldwide in 2024 and it was also the most frequently diagnosed cancer among women in 164 of 186 countries [1]. In 2024, approximately 2.4 million women were diagnosed with breast cancer, which highlights the global burden and the need for effective new solutions [1]. Breast cancer is a condition in which abnormal cells in the breast grow out of control and can form tumors that can be deadly if left untreated. It typically starts in the milk ducts or lobules and the early stage (in situ) is generally not life-threatening and often detectable early if screening is sought [2]. Cancer cells that spread into nearby breast tissue form a lump or result in thickening of the breast tissue. In more advanced cases, invasive cancers can invade other organs or lymph nodes by metastasis, which is life-threatening and can lead to death [1]. Lymphadenopathy, bone, liver, lung and brain metastasis significantly worsen prognosis and treatment. It is thus crucial to identify the risk of metastasis at an early stage to improve prognosis, to better inform personalised treatment and to better inform clinical decisions.

Currently, the most common methods for predicting breast cancer metastasis are based on a histopathological examination, clinical staging, radiological analysis and molecular typing. While these strategies remain clinically relevant, they continue to have several limitations in predictive reliability and translational efficiency. Traditional imaging interpretation is often radiologist-dependent and can miss subtle quantitative signatures of intratumoral heterogeneity as metastatic progression unfolds [3,4,5]. Furthermore, the traditional statistical prediction models lack the capacity to incorporate multimodal data from heterogeneous imaging, biomarkers, pathological and molecular evidence [6]. Histopathological evaluation and clinicopathological scoring systems also have low sensitivity for early detection of metastatic transformation and are often difficult to interpret due to inter-observer variability [7]. In addition, traditional statistical prediction models are limited by their low-dimensional data-handling capacity and lack of integration between heterogeneous multimodal evidence from imaging, biomarkers, pathology and molecular datasets [6]. The shortcomings of these approaches have underscored the need for new computational systems capable of providing actionable clinical insights from large, complex oncology datasets.

The use of artificial intelligence (AI), machine learning (ML), deep learning (DL), and radiomics has revolutionised the analysis of breast cancer metastasis. The ability of AI-driven systems to extract high-dimensional quantitative imaging features, identify non-linear relationships and integrate heterogeneous data sources for metastasis prediction and prognosis stratification is gaining in capability. Random Forest, support vector machine, XGBoost and logistic regression models are ML methods that have proven to be highly useful for predicting metastasis from clinical and radiomics data [8]. In a similar vein, DL setups including convolutional neural networks (CNNs), DenseNet, EfficientNet, and transformer-based systems have enhanced automated feature extraction, lesion characterisation and metastasis classification using imaging as well as pathology data [9,10,11]. Radiomics has also proven to be a translational direction that enables the extraction of non-invasive quantitative markers from routine medical images, which characterise tumor heterogeneity, microenvironmental dynamics and metastatic potential [12]. Artificial intelligence (AI) is also an emerging transformative technology and WHO has acknowledged its potential to enhance healthcare diagnostics, precision oncology and evidence-based decision-making in the healthcare sector, if used appropriately and clinically validated [13].

While there has been increasing interest in the scope of AI in breast cancer metastasis analytics, the currently available review studies remain fragmented and methodologically limited. Some recent reviews have largely centered on radiomics applications in breast imaging, overlooking the important discussion of the feature engineering processes underlying the development of metastasis prediction systems [14,15]. Other reviews focused predominantly on DL algorithms, imaging modalities, or specific metastatic endpoints, such as lymph node metastasis, while providing limited synthesis of multimodal feature integration and explainability frameworks [10,16,17]. Additional reviews have covered AI use in cancer, but none have specifically explored the feature-engineering pipelines used to analyse breast cancer metastasis [18,19]. Existing literature also demonstrates limited conceptual integration of feature extraction, preprocessing, dimensionality reduction, multimodal fusion and explainability into a unified analytical framework. Moreover, there is relatively limited research work to synthesise how imaging radiomics can be integrated with clinicopathological, molecular, biomarker and pathology-derived evidence inside metastasis prediction ecosystems.

Recent reviews by [20,21,22,23] similarly emphasised ongoing challenges related to reproducibility, standardisation, external validation, data harmonisation and explainability in AI-driven oncology analytics. These methodological challenges suggest the need for a comprehensive synthesis specifically on feature engineering processes in breast cancer metastasis analytics. Consequently, the present study has been designed to systematically synthesise the feature processing techniques, feature engineering strategies, feature types, multimodal integration approaches and explainability mechanisms used in AI-based breast cancer metastasis analytics. This study also aims to provide a conceptual understanding of how diverse imaging, molecular, biomarker, pathological and clinical characteristics are integrated into clinically relevant evidence for predicting metastasis. The following research questions were addressed in this study:

• What feature engineering techniques are employed in AI-based breast cancer metastasis analytics?
• Which feature types are most commonly used in AI models for breast cancer metastasis prediction?
• What evidence-informed conceptual feature-centric framework can be derived from feature engineering pipelines used in AI-driven breast cancer metastasis analytics?

Technical Contributions

This study makes the following technical contributions:

• It develops a feature-centered synthesis of AI-driven breast cancer metastasis analytics by showing how raw imaging, clinical, biomarker, molecular and pathology data are transformed into metastasis-relevant predictive features.
• It proposes the 5Fs Framework as a structured feature engineering pipeline comprising feature extraction or generation, feature preparation and transformation, feature selection and dimensionality reduction, feature construction and integration and feature interpretability and explainability.
• It identifies the key feature engineering techniques employed in breast cancer metastasis prediction across diverse methodological contexts.
• It maps the major feature categories employed in AI-based metastasis analytics, encompassing imaging modalities, clinicopathological variables, molecular and biomarker indicators, pathology-derived features and site-specific labels.
• It demonstrates that multimodal feature fusion is a central technical pathway for improving metastasis prediction by integrating radiomics, clinical, pathological, molecular and biomarker evidence.
• It provides a technical foundation for future AI-based breast cancer metastasis prediction systems by emphasising reproducibility, data harmonisation, external validation, multimodal integration and interpretable model deployment.

2. Materials and Methods

2.1. Review Design

The study design was based on the systematic review method to provide a synthesis of evidence related to feature engineering processes, feature types, multimodal integration strategies and explainability mechanisms used in AI-driven breast cancer metastasis analytics. Studies related to AI, ML, DL, radiomics and multimodal predictive systems for metastasis analytics and prediction in breast cancer were investigated in the review. The review process followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, which promote methodological transparency, reproducibility and rigor in reporting. A structured review protocol was therefore adhered to in a manner that guided the identification of literature, screening, eligibility assessment, quality appraisal, data extraction and thematic synthesis to ensure that it was aligned with the research questions of this study. The PRISMA 2020 study selection process is presented in Figure 1, while the PRISMA 2020 for Abstracts Checklist and the PRISMA 2020 Checklist are provided in Appendix A.1 and Appendix A.2, respectively.

Figure 1. PRISMA study identification and selection flow diagram.

2.2. Information Sources and Search Strategy

A comprehensive literature search was conducted between 1 May 2026 and 15 May 2026 across multiple multidisciplinary and healthcare-related electronic databases including MEDLINE, Scopus, Web of Science, Embase and IEEE Xplore. These databases were chosen as they cover a wide range of research in oncology, radiology, radiomics, AI, medical imaging and healthcare analytics. The reference lists of the eligible articles and relevant review studies were manually screened to identify additional studies. The search strategy was developed using keywords derived from the research questions and the thematic concepts central to breast cancer metastasis analytics. The search incorporated combinations of Boolean operators and keywords such as “breast cancer”, “metastasis”, “lymph node metastasis”, “distant metastasis”, “artificial intelligence”, “machine learning”, “deep learning”, “radiomics”, “feature engineering” and “multimodal integration”. An example of the search string used was: (“breast cancer” OR “breast neoplasm”) AND (“metastasis” OR “metastatic progression”) AND (“artificial intelligence” OR “machine learning” OR “deep learning” OR “radiomics”) AND (“feature engineering” OR “feature extraction” OR “feature processing” OR “feature selection” OR “prediction analytics”).

2.3. Eligibility Criteria

Studies that examined the use of AI, ML, or DL in breast cancer metastasis analytics and prediction or multimodal analytical or feature engineering techniques were included. Studies were eligible for inclusion if they incorporated feature engineering processes, imaging-derived features, clinicopathological variables, molecular integration, biomarker integration and/or explainability processes related to metastasis analytics. Inclusion of articles was limited to full-text peer-reviewed articles published in English in the 2020–2026 period. Articles focusing on cancers other than breast cancer, non-metastatic outcomes, purely laboratory experiments without AI analytics, editorials, conference abstracts, protocols, dissertations and review articles were excluded. Additionally, studies that did not provide adequate methodological detail on feature engineering and/or predictive analytics were excluded from the synthesis.

2.4. Study Selection Process

Selection of studies was based on a sequential screening procedure that was based on predefined criteria. All studies retrieved from the various electronic literature databases were exported to a reference management system, where duplicate records were identified and removed electronically and manually before screening commenced. Subsequently, the titles and abstracts were screened independently to identify potentially relevant studies related to AI-driven breast cancer metastasis analytics. Studies that passed the initial screening were then assessed in full text to determine eligibility for final use in the synthesis. The screening process focused on relevance to feature engineering techniques, type of features, multimodal integration, explainability methods and outcomes of metastasis prediction. Any disagreements that arose during the screening and eligibility process were discussed, and a consensus was reached to ensure a consistent and reliable methodological approach in the selection process.

2.5. Quality Appraisal

The Mixed Methods Appraisal Tool (MMAT) was used to evaluate the methodological quality of the included studies. The MMAT was selected because the evidence comprised heterogeneous empirical quantitative studies, including retrospective diagnostic, prognostic, radiomics, machine-learning, deep-learning, computational pathology and biomarker-based studies. Thus, the appraisal was confined to the applicable MMAT criteria, within which relevant aspects of AI study design, data processing, measurement, analysis and validation were considered integral components of methodological quality rather than separate appraisal criteria. Since majority of the studies were quantitative non-randomised, the tool was used to ensure consistent appraisal of various designs within one framework and operationalised through MMAT screening questions and criteria 3.1–3.5: clarity of research questions, appropriateness of data, participant representativeness, measurement validity, completeness of outcome data, control of confounders and suitability of statistical analysis.

Each MMAT criterion was rated “Yes”, “No”, or “Can’t tell”. A “Yes” rating indicated that the criterion was satisfied, a “No” rating indicated that it was not satisfied, and “Can’t tell” indicated that the reported information was insufficient to determine whether the criterion was satisfied. In accordance with MMAT guidance, no overall numerical score or percentage was calculated. Instead, author-defined narrative categories were used to summarise the pattern of MMAT responses. Studies were classified as high quality when both screening questions were rated “Yes”, and three to five of the criteria 3.1–3.5 were rated “Yes.” Studies were classified as moderate quality when both screening questions were rated “Yes,” but no more than two of the criteria 3.1–3.5 were rated “Yes.” Only criteria rated “Yes” were regarded as satisfied when assigning the narrative quality category. These classifications are narrative summaries developed for this review and should not be interpreted as official MMAT categories. The detailed quality appraisal results are outlined in Appendix B.

2.6. Data Extraction and Synthesis

A structured extraction table was developed based on the study’s research questions. Extracted data included author details, publication year, study aim, study methods, feature engineering techniques, feature types and AI approaches, as well as information regarding the results such as model performance indicators. The summary table of data extracted from the articles included is presented in Table 1 and Appendix C. Thematic synthesis was used to analyse the extracted evidence and find common methodological and analytical patterns found in the studies. The thematic synthesis was guided by thematic codes that were progressively refined to basic themes, organising themes and grand themes, following the method developed by [24] called thematic network analysis. Appendix C shows the data extraction summary table. The thematic interpretation also guided the development of the conceptual framework representing the feature-engineering pipeline in AI-driven breast cancer metastasis analytics.

Table 1. Feature-engineering techniques employed in AI-based breast cancer metastasis analytics and the feature types.

Grand Theme	Organising Theme	Basic Theme	Thematic Code	References
Feature-processing and engineering	Feature extraction or generation	Imaging radiomics extraction	PyRadiomics	[25,26,27,28,29,30,31,32]
			Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) radiomics	[25,31,33,34,35,36,37,38]
			Ultrasound radiomics	[27,28,39,40,41]
		Texture-feature extraction	Gray level co-occurrence matrix (GLCM) texture features	[25,29,30,31,32,33,37,38,42,43]
			Gray level run length matrix (GLRLM) texture features	[25,29,30,31,32,33,38,42]
			Gray level size zone matrix (GLSZM) texture features	[25,29,30,31,32,33,38,42]
			Gray level dependence matrix (GLDM) texture features	[25,29,31,32,33,35,38,42,43]
			Neighbourhood gray tone difference matrix (NGTDM) texture features	[25,29,30,32,38,43]
		Deep-feature extraction	CNN-derived deep features	[29,36,39,41,44,45,46,47,48,49]
			VGG/DenseNet/MobileNet/EfficientNet deep-feature extraction	[36,41,44,47,50]
		Pathology feature extraction	WSI patch extraction	[45,48,49,50,51,52]
	Feature preparation and transformation	Region of Interest (ROI) and image preparation	Manual ROI/Volume of Interest (VOI) segmentation	[25,27,28,30,31,33,36,37,38,42,43,51]
		Image/data standardisation	Z-score normalisation	[27,31,37,39,40,53,54,55,56,57,58]
		Reproducibility testing	Intra-class correlation coefficient (ICC) feature reproducibility filtering	[25,27,28,31,32,33,37,40,41,46]
		Data balancing	SMOTE oversampling	[26,36,42,59,60]
	Feature selection and dimensionality reduction	Penalized regression	Least absolute shrinkage and selection operator (LASSO) feature selection	[25,27,28,30,31,33,37,38,40,46,51,57,58,61,62,63,64,65]
		Redundancy reduction	Pearson correlation filtering	[30,33,37,43,51]
			Spearman correlation filtering	[40,51,55,57,58]
			MRMR feature selection	[27,28,40,61,62]
		ML-based feature ranking	Random Forest feature ranking or mean decrease accuracy/Gini	[29,54,66,67,68]
	Feature construction and integration	Radiomics-score construction	Rad-score/radiomics signature construction	[25,27,28,29,30,33,36,37,38,43]
		Nomogram construction	Clinical-radiomics or multimodal nomogram	[25,28,30,36,37,43,44,51,54,69]
		Clinical-imaging fusion	Imaging radiomics integrated with clinical/pathological variables	[25,27,28,30,35,36,37,38,44,46,60,70]
	Feature interpretability and explainability	SHAP explainability	SHAP feature attribution	[27,31,41,56,64,68,71]
		Visual explainability	Gradient-weighted Class Activation Mapping (Grad-CAM), class activation map (CAM), attention heatmaps or prototype visualisation	[29,39,41,45,47,49,52]
		Clinical utility testing	Decision curve analysis	[25,27,28,29,30,33,43,51,54,55,57,58,68,70]
Feature types	Imaging-derived feature types	MRI feature types	DCE-MRI features	[25,29,31,33,34,35,36,37,38,44,53]
			Diffusion-weighted imaging (DWI)/apparent diffusion coefficient (ADC) features	[29,31,37,42,53,60]
			Contrast-enhanced T1-weighted imaging (T1 + C)/T2-weighted imaging (T2WI) multiparametric MRI features	[29,46,60,62,64,65,71]
		Ultrasound feature types	B-mode ultrasound features	[27,28,39,40,41]
	Clinical and clinicopathological feature types	Tumor and nodal variables	Tumour size or lesion diameter	[28,38,51,58,68]
		Pathology variables	ER, PR, human epidermal growth factor receptor (HER2) or Ki-67 proliferation index (Ki-67) status	[29,38,40,43,48,51,58,64,65,68]
	Digital pathology and pathomics feature types	WSI pathology features	Haematoxylin and eosin (H&E) whole-slide image features	[45,48,49,50,51,52]
	Metastasis-site and outcome-related feature types	Lymph node metastasis outcomes	Axillary lymph node metastasis (ALNM) prediction	[25,27,28,30,31,33,39,40,48,51,53,62,65,70]
		Distant metastasis outcomes	Distant metastasis prediction	[49,55,56,57,58,61,69,72,73]
		Organ-specific metastasis outcomes	Bone metastasis prediction	[43,57,58,63,67,68]
			Bone/liver/lung/brain metastasis site labels	[32,43,57,58,67,68,74]

3. Results

3.1. Study Selection

A search of the electronic literature databases identified 1772 records from PubMed/MEDLINE, Scopus, Web of Science Core Collection, Embase and IEEE Xplore. A total of 314 duplicate records were removed, and 1458 studies were screened by title and abstract, of which 1260 were excluded. A total of 198 reports were retrieved and evaluated for eligibility. After full-text evaluation, 148 studies were excluded for the following reasons: no information on metastasis, no breast cancer-specific focus on metastasis, insufficient information on the features being engineered, non-empirical designs, review papers only, or preclinical studies. In the end, 50 studies were included in this systematic review.

3.2. Study Characteristics

All 50 included studies were related to breast cancer metastasis analytics, prediction, classification, or feature identification. The majority of the studies focused on lymph-node metastasis outcomes (30/50, 60%), such as axillary lymph node metastasis (ALNM), sentinel lymph node metastasis and high axillary lymph node burden. An additional 18 studies (36%) examined distant, organ-specific, recurrence or metastasis or survival-related outcomes, such as bone, liver, lung, brain, distant metastasis, metabolic response, recurrence/metastasis risk and 5-year survival. Of these, bone metastasis was the most common organ-specific distant metastasis focus, reported in 6 studies (12%), followed by broader distant metastasis or prediction of recurrence/metastasis seen in 7 studies (14%). A small number of studies used comparator or mixed metastatic settings (2/50; 4%): one differentiated breast-cancer bone metastases from non-breast-cancer bone lesions and one classified the primary tumor origin of brain metastases while including breast cancer brain metastases.

Across publication years, the included studies were published between 2021 and 2025: 2 studies (4%) in 2021, 12 studies (24%) in 2022, 10 studies (20%) in 2023, 11 studies (22%) in 2024 and 15 studies (30%) in 2025. Thus, the overall evidence is heavily concentrated in the most recent years, with 37 of 50 studies (74%) published from 2023 onward. By study design, the evidence base was mostly retrospective empirical work (43/50; 86%).

A variety of AI algorithms were used across the included studies. The most frequently used predictive algorithms were support vector machine [SVM] (23 studies, 46%), logistic regression [LR] (21 studies, 42%), random forest [RF] (20 studies, 40%) and gradient-boosting algorithms ([XGBoost, LightGBM, GBM, CatBoost, AdaBoost, gradient boosting] 18 studies, 36%). Decision tree (DT) models were used in 24% of the studies and k-nearest neighbor (KNN) was employed in 22% of the studies. Deep-learning architectures, including CNN-based models, ResNet, DenseNet, VGG, MobileNet, EfficientNet, vision transformer and attention-based multiple instance learning (MIL) frameworks, were used in 13 studies (26%), mainly in imaging and digital pathology-based metastasis prediction.

3.3. Quality Appraisal Results

The quality appraisal process showed that 43 of 50 studies (86%) were rated high quality, while 7 studies (14%) were rated moderate quality. The high-quality studies showed strong methodological features, such as strong preprocessing, feature selection, cross-validation, external/multicenter validation, explainable AI, calibration analysis and clinically relevant prediction performance. The moderate-quality studies remained methodologically useful but were limited by retrospective single-center design, small sample size, manual segmentation, moderate performance, or limited external validation. The appraisal overall demonstrated that the included studies provided sufficiently rigorous and relevant methodological evidence to support synthesis in this review.

3.4. Feature-Engineering Techniques and Feature Types

The evidence retrieved demonstrates a rapidly expanding application of AI, radiomics and ML, as well as multimodal data analytics in breast cancer metastasis analytics and metastasis-related clinical decision support. The findings reveal themes around extraction, preparation, transformation, integration and interpretation of heterogeneous features derived from imaging, clinicopathological, molecular and biomarker datasets to improve metastasis detection, prediction, classification and prognostic stratification in breast cancer. The findings are grouped into two broad themes: feature-processing and engineering techniques used in AI-based breast cancer metastasis analytics and the types of features used in AI-based breast cancer metastasis analytics.

3.4.1. Feature-Processing and Engineering Techniques in AI-Based Breast Cancer Metastasis Analytics

Feature extraction and generation emerged as the most dominant analytical process across the extracted evidence. The evidence revealed that imaging radiomics extraction is a key component of most of the pipelines for metastasis prediction, notably those used in magnetic resonance imaging (MRI), computed tomography (CT), ultrasound imaging, PET-CT and diffusion-weighted imaging modalities. Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) radiomics, contrast-enhanced magnetic resonance imaging (CE-MRI) radiomics, CT radiomics, DWI radiomics, ADC feature extraction and multimodal MRI radiomics were commonly used to characterise tumor heterogeneity, microenvironmental alterations and metastatic behavior. High-dimensional quantitative descriptors were extracted repeatedly from regions of interest (ROIs) for the tumor, vertebral bone marrow regions, lymph nodes and metastatic lesions on radiomics platforms such as PyRadiomics, Darwin Scientific Research Platform, 3D Slicer, ITK-SNAP medical image segmentation software and PyRadiomics libraries. Typical feature-generation processes were the first-order statistical descriptors, shape-based, texture-based and matrix-based radiomics features, such as gray level co-occurrence matrix (GLCM) features, gray level run length matrix (GLRLM) features, gray level dependence matrix (GLDM) features, gray level size zone matrix (GLSZM) features and neighborhood gray tone difference matrix (NGTDM) features. The primary aim of radiomics extraction was to predict ALN metastasis, sentinel lymph node metastasis, distant metastasis, bone metastasis, liver metastasis, lung metastasis and brain metastasis.

The results also underscored the importance of feature preparation and transformation as a key methodological step in the analysis of breast cancer metastasis. Most research applied extensive preprocessing algorithms in order to optimise the reproducibility, standardisation and robustness of the features prior to the subsequent analytical modelling. Commonly used preprocessing methods are normalisation techniques such as z-score normalisation, maximum–minimum normalisation, pixel intensity standardisation, standard normal distribution transformation and unit standardisation. Voxel resampling, image reconstruction standardisation, ROI contour review and segmentation refinement, inter-observer reproducibility analysis (using intraclass correlation coefficients (ICC) or Cohen’s kappa statistics) were often included in imaging research. Additional techniques employed also encompass redundancy reduction and correlation filtering, notably Pearson correlation filtering, Spearman correlation analysis, variance thresholding and redundancy mitigation. The widespread use of dimensionality preprocessing techniques such as principal component analysis (PCA), missing-value imputation, outlier detection and harmonisation of imaging parameters was also evident. These preprocessing techniques demonstrate a focus on maintaining the stability of the analytical process, enabling robust and reliable prediction of metastasis, as well as minimising noise in AI-driven metastasis prediction pipelines.

Feature selection and dimensionality reduction further emerged as an increasingly important methodological consideration. The use of feature reduction methods based on statistical and ML approaches to identify metastasis-relevant predictors from high-dimensional data was reported. The most commonly mentioned dimensionality reduction method is the Least Absolute Shrinkage and Selection Operator (LASSO) regression. LASSO has been widely applied with sparse modelling, L1 regularisation, coefficient shrinkage and cross-validation optimisation to identify the most informative radiomics, biomarker and clinicopathological features associated with metastatic progression. Other studies employed SelectKBest, minimum redundancy maximum relevance (mRMR), variance filtering, sample variance F-value filtering, logistic regression-based selection, Cox regression analysis and statistical significance thresholding to reduce dimensionality and keep metastasis-related predictors. The feature selection processes frequently yield compact feature subsets that contain highly discriminative imaging signatures, biomarker combinations, or clinicopathological variables predictive of lymphatic or distant metastatic spread.

Another important finding reveals that feature integration and fusion represent a more conceptually relevant analytical component in metastasis-oriented feature engineering. The results indicate that integrating heterogeneous data sources enhances the predictive performance of metastasis prediction models and, in turn, improves their clinical applicability. In related results, radiomics signatures are often integrated with clinical variables, clinicopathological data, molecular biomarkers, blood-based biomarkers and imaging descriptors in multimodal fusion strategies, such as forming nomograms, comprehensive prediction models and multimodal analytical frameworks. Multiple studies reveal integration of combined DCE-MRI and diffusion-weighted imaging (DWI) radiomics features and other studies combined CT radiomics with DL segmentation output, or ultrasound with pathology and blood biomarker profiles. Multiple clinical-radiomic nomograms and multifactorial integration strategies, as well as multimodal feature-fusion frameworks, have been repeatedly reported to improve metastasis prediction performance relative to single-modality approaches. The results indicate that integration of heterogeneous feature domains is an important aspect of contemporary AI-based breast cancer metastasis analytics.

Feature interpretability and explainability are also more prominent methodological techniques in AI-based breast cancer metastasis analytics. The results explicitly show that explainable AI (XAI) techniques improve transparency of prediction models, feature attribution and clinical interpretability of metastasis prediction systems. Shapley Additive Explanations (SHAP) is the most commonly used explainability framework, reporting and depicting SHAP values, beeswarm plots, feature importance and variable contribution analyses to identify metastasis-associated predictors. Other interpretability tools included correlation heatmaps, feature coefficient analysis, decision curve analysis, odds ratio interpretation, confusion matrices, and radiomics score comparisons between metastasis-positive and metastasis-negative groups. By and large, the findings highlight the need for interpretable AI systems to translate complex metastasis prediction analytics into clinically actionable decision support tools in breast oncology practice.

3.4.2. Feature Types Used in AI-Based Breast Cancer Metastasis Analytics

The second grand theme reveals various types of features utilised in AI-based breast cancer metastasis analytics. The most prevalent data type comprises imaging-derived features. Extensive usage of MRI-derived radiomics, CT-derived radiomics, ultrasound imaging descriptors, PET-CT imaging variables, DWI features, ADC maps and vertebral bone marrow imaging signatures is evident. The imaging-based features that are mainly captured include tumor morphology, texture, lesion intensity distributions, volumetric features and microenvironmental changes that can occur during metastatic dissemination. Texture-based radiomics matrices such as GLCM, GLRLM, GLDM, GLSZM and NGTDM are common matrices reflected in analytics focused on tumor heterogeneity and metastasis risk stratification.

Clinical and clinicopathological feature types also appeared extensively. The most commonly used variables are tumor size, histological subtype, tumor–node–metastasis (TNM) stage, lymph node status, menopausal status, age, molecular subtype, Breast Imaging Reporting and Data System (BI-RADS) classification, treatment variables and surgical status. Such clinicopathological characteristics are often integrated with imaging features to improve the accuracy of predictions of lymph node metastasis, distant metastasis and survival. Despite the growing prominence of radiomic and AI-based analysis methods, clinicopathological variables remain foundational predictors within multimodal metastasis prediction systems.

Another important feature type is blood-based biomarkers, especially those used to study distant metastasis and predict bone metastasis. Many of the biomarkers that are commonly reported are carcinoembryonic antigen (CEA), cancer antigen 15-3 (CA153), cancer antigen 125 (CA125), cancer antigen 19-9 (CA199), alpha-fetoprotein (AFP), creatine kinase-MB isoenzyme (CK-MB), alpha-hydroxybutyrate dehydrogenase (α-HBDH), apolipoprotein B, albumin–globulin ratio, alkaline phosphatase, lipoprotein-a and gamma-glutamyl transferase. These biomarkers are often integrated with clinical or image-derived markers to further enhance predictive discrimination between metastasis-positive and non-metastatic patient groups. The integration of systemic biochemical indicators therefore reflects growing interest in combining tumor imaging phenotypes with circulating physiological signatures associated with metastatic progression.

Molecular and omics-derived feature types were also commonly represented. Estrogen receptor/progesterone receptor (ER/PR), human epidermal growth factor receptor 2 (HER2), Ki-67 proliferation index (Ki-67), molecular subtypes, luminal subtypes, HER2-overexpression profiles and triple-negative breast cancer markers are repeatedly reported as metastasis-associated markers. These molecular descriptors are often integrated into predictive models and linked to metastatic risk stratification, tumor aggressiveness and survival outcomes. The results indicate that molecular subtyping also remains an integral part of AI-driven metastasis analytics in breast cancer.

In contrast, digital pathology and pathomics feature types appear less frequently in breast cancer metastasis analytics work. Only isolated instances included biopsy pathology descriptors, histopathological examination variables, or pathology-derived features in the metastasis prediction systems. The results show that digital pathology integration remains comparatively underdeveloped relative to imaging radiomics and clinical-radiomic analytics in current breast cancer AI-based metastasis modelling.

Finally, the evidence consistently showed feature types associated with metastatic sites and outcomes. Multiple metastatic endpoints emerging from the results include metastasis to ALNs, metastasis to sentinel lymph nodes, bone metastasis, liver metastasis, lung metastasis, brain metastasis, distant metastasis, metastatic recurrence and metastasis-related survival outcomes. Various studies specifically differentiated metastasis-positive and metastasis-negative patient groups, while others focused on predicting future metastatic occurrence within predefined follow-up periods. By and large, the current evidence highlights the progressive maturation of AI-driven breast cancer metastasis analytics, transforming it into multimodal, interpretable, and comprehensive predictive systems that can aid metastasis detection and diagnostics, prognostic stratification, and personalised oncology decision-making across diverse metastatic contexts.

3.5. Evidence-Informed Conceptual 5Fs Framework for Feature Engineering in AI-Driven Breast Cancer Metastasis Analytics

The evidence extracted in this study provided a strong basis for developing a layered, feature-centric framework (Figure 2), referred to as the 5Fs Framework for the feature-engineering pipeline in AI-driven breast cancer metastasis analytics. It is worth noting that the acronym 5Fs is used to refer to five related feature-engineering steps: feature extraction or generation (FE/G), feature preparation and transformation (FPT), feature selection and dimensionality reduction (FSD), feature construction and integration (FCI) and feature interpretability and explainability (FIE). These stages collectively illustrate the 5Fs approach for feature engineering the breast cancer metastasis data pipeline from raw data to clinically useful, interpretable and metastasis-relevant predictive evidence.

Figure 2. The evidence-informed conceptual 5Fs Framework for feature engineering pipeline in AI-driven breast cancer metastasis analytics (The 5Fs framework (FE/G–FPT–FSD–FCI–FIE) provides a systematic, evidence-based roadmap for engineering high-quality, interpretable and clinically relevant features for AI-driven breast cancer metastasis analytics).

The framework outlines the five stages of evidence-based metastasis prediction, which are sequential but not mutually exclusive: FE/G, where metastasis-relevant data are extracted or generated; FSD—high-dimensional features and transformed into features suitable for reliable analysis; FSD-high-dimensional features are reduced to the most informative metastasis predictors; FCI—features are assembled into an integrated predictive signature; and FIE—feature contributions are explained for clinical interpretation. This framework elucidates a model-based approach that is fundamentally feature-driven, integrating imaging, clinical, biomarker and pathology-derived evidence to enhance metastasis prediction, diagnostic accuracy and AI-driven decision making.

• Layer 1: Feature Extraction or Generation (FE/G)

The base level of the framework is FE/G. It illustrates how to turn raw imaging, clinical, pathological, blood-based, molecular and omics data into measurable candidate predictors of metastasis modelling. This layer captures tumor characteristics, biological behavior, patient clinical status and metastatic risk patterns through radiomics, pathomics, biomarker extraction and molecular profiling.

• Layer 2: Feature Preparation and Transformation (FPT)

This layer preprocesses extracted features to make them amenable for reliable analysis by improving quality, standardisation, reproducibility and comparability. This layer converts raw extracted features into analysis-ready variables through normalisation, resampling, segmentation refinement, missing-value handling, outlier management, reproducibility checks and correlation preprocessing, among other techniques.

• Layer 3: Feature Selection and Dimensionality Reduction (FSD)

This layer detects the most metastasis-relevant features from the high-dimensional feature space. At this stage, FSD converts high-dimensional feature pools into low-dimensional feature signatures that represent features associated with metastases. It filters out the noise of features to obtain short, metastasis-relevant feature signatures for downstream integration and interpretation, while keeping only features which are most relevant to metastatic outcomes.

• Layer 4: Feature Construction and Integration (FCI)

This layer synthesises the selected features together to create higher-order predictive signatures, scores, nomograms and multimodal feature combinations. It presents the process of how refined predictors are combined into clinically relevant feature products that integrate imaging, clinical, biomarker, molecular, pathological and omics evidence for breast cancer metastasis analytics. The role of FCI in the framework is to convert the selected predictors into integrated representations of the metastasis risk that can be used for diagnosis, prediction, stratification or prognosis.

• Layer 5: Feature Interpretability and Explainability (FIE)

This is the final translational layer of the framework. It describes how engineered features can aid breast cancer metastasis prediction and how their results can inform clinical decisions. These include various feature-importance rankings, SHAP values, coefficient interpretation, odds ratios, radiomics score interpretation, nomogram visualisation, calibration assessment, decision curve analysis, heatmaps and clinical net benefit evaluation. Thus, the role of FIE is to translate engineered features into clinically meaningful evidence.

4. Discussion

The results revealed that AI-powered breast cancer metastasis analytics evolve from stand-alone prediction models to feature engineering ecosystems that effectively integrate imaging, clinical, pathological, biomarker and molecular evidence to build clinically interpretable metastasis prediction systems. The dominance of radiomics-based feature extraction reflects the growing recognition that tumor heterogeneity contains clinically meaningful signatures associated with metastatic progression. Evidence from previous studies also demonstrated that radiomics derived from MRI, PET-CT and ultrasound can capture intratumoral heterogeneity and microenvironmental variations linked to metastatic dissemination and survival outcomes in breast cancer [75,76]. The extensive use of texture matrices such as GLCM, GLRLM, GLSZM and GLDM further supports the premise that spatial heterogeneity within tumor regions represents an important biomarker for metastatic behavior.

The findings also establish that the preparation and transformation of features remain foundational to reliable AI-based metastasis analytics. Normalisation, segmentation refinement, ICC reproducibility testing and correlation filtering are widely adopted, pointing towards a growing awareness that poorly standardised radiomics pipelines are likely to limit reproducibility and external validity. This observation has been echoed in recent radiomics standardisation research, which has pinpointed segmentation, image acquisition and radiomics preprocessing as major barriers to clinical translation [77,78,79]. The emphasis on preprocessing therefore reflects an emerging methodological shift toward reproducible and clinically robust AI workflows.

LASSO, mRMR and correlation filtering are the most frequently used methods in the feature selection process, which highlights the difficult situation of dealing with high-dimensional metastasis datasets. The results suggest that AI-powered metastasis analytics is not just an algorithmic phenomenon, but a feature-based one as well. Recent evidence similarly suggests that compact and biologically meaningful feature subsets often outperform excessively large feature pools in oncological prediction systems due to improved generalisability and reduced overfitting [80,81]. The finding is consistent with other recommendations for AI governance by the World Health Organization, which highlight the importance of transparency, robustness, explainability and generation of clinically meaningful evidence in the context of AI-supported health systems [13].

A key finding is the growing use of clinical radiomics fusion, nomograms and multimodal predictive frameworks to integrate heterogeneous multimodal data. This indicates that metastatic progression is increasingly conceptualised as a multidimensional biological process requiring simultaneous interpretation of imaging phenotypes, molecular markers and clinicopathological characteristics. Empirical evidence from past studies has also shown that multimodal AI systems integrating both radiomics and molecular and clinical data achieve better performance in metastasis prediction or detection and survival stratification than single-modality models [82,83]. The crafted evidence-based 5Fs Framework for feature-engineering pipeline in AI-driven breast cancer metastasis analytics therefore provides an important conceptual contribution by systematically positioning feature engineering as the central translational mechanism linking raw cancer data to clinically interpretable metastasis prediction.

The increasing relevance of explainable AI approaches, Grad-CAM and SHAP further demonstrates the rising appreciation of the importance of interpretability and clinician trust for clinical use. Recent studies have found that explainable AI can enhance the transparency, clinician trust and clinical acceptance of oncology decision support systems [22,84,85]. The results thus indicate that future breast cancer metastasis analytics is likely to evolve towards an explainable, multimodal and clinically integrated AI ecosystem that will enable personalised oncology decision making, metastasis risk stratification and precision medicine.

5. Conclusions

The application of AI tools for breast cancer metastasis analysis relies on complex feature engineering workflows that convert complex imaging, clinical, molecular, biomarker and pathological data into meaningful clinical evidence in the form of predictions. The results show that radiomics-based imaging features such as MRI, DWI, Apparent Diffusion Coefficient (ADC), ultrasound and multimodal imaging signatures are the most prominent features for the identification of metastatic risk patterns and tumor heterogeneity associated with disease progression. The consistent usage of preprocessing, normalisation, reproducibility assessment, dimensionality reduction and feature selection methods further highlights the importance of analytical stability and feature robustness in metastasis prediction systems. The integration of heterogeneous feature domains through multimodal fusion frameworks, radiomics signatures and clinical nomograms reflects a growing transition toward precision-oriented oncology analytics capable of supporting individualised metastasis prediction and AI-driven decision-making. The increasing incorporation of explainable AI approaches such as SHAP, Grad-CAM and feature attribution analysis further demonstrates the need for transparency and interpretability in clinically deployable AI systems.

The proposed evidence-informed conceptual 5Fs framework provides a structured representation of how metastasis-relevant data can be systematically extracted, refined, selected, integrated and interpreted within AI-driven analytical environments. The future of AI-enabled metastasis analytics promises to be greatly shaped by advancements in data harmonisation, external validation across diverse patient populations, more standardised feature engineering workflows and stronger integration of multimodal biological evidence. These advances may contribute toward more accurate metastasis prediction, earlier detection of disease progression and more personalised therapeutic decision making in breast cancer management.

6. Strengths and Limitations

One of the strengths of this study is its thorough synthesis of recent empirical research on feature-engineering pipelines for breast cancer metastasis analytics across 50 included studies. Imaging, pathology, clinical, genomic, proteomic, metabolomic and multimodal AI studies added richness and depth to evidence synthesis and brought a more holistic picture to the emerging AI applications in breast cancer metastasis prediction. In addition, the review incorporated a multidisciplinary approach, including evidence from radiomics, ML and DL, genomics, pathomics, biomarker analytics and multimodal AI frameworks, which informed the development of the evidence-informed conceptual 5Fs Framework. However, there are limitations to point out. Some studies included in this review applied different imaging modalities, feature-processing workflows, validation methods, outcome definitions and AI algorithms, which makes direct comparison challenging. Comparability and generalisability of findings may be influenced by variability in reporting and limited external validation.

7. Future Directions

Future studies should be directed towards the development of more standardised, externally validated and clinically implementable pipelines for the creation of AI-driven breast cancer metastasis analytics. Research should prioritise standardised radiomics and pathomics extraction procedures, transparent reporting of feature-selection methods, and reproducible validation across different imaging systems, populations and clinical contexts. Future studies should also improve multimodal integration by integrating imaging, clinicopathological, blood-based, molecular, genomic and digital pathology features in the same prediction frameworks. In addition, explainable AI methods should be routinely incorporated to clarify feature contribution, improve clinical trust and support decision making. The proposed 5Fs Framework may provide a conceptual guide for future feature-centric and interpretable AI research; however, its structure, practical utility, and clinical relevance should be evaluated through expert consultation, prospective studies, external validation, and implementation research before clinical application.

Appendix A.1

Table A1. PRISMA abstract checklist.

Section and Topic	Item #	Checklist Item	Reported (Yes/No)
TITLE
Title	1	Identify the report as a systematic review.	Yes
BACKGROUND
Objectives	2	Provide an explicit statement of the main objective(s) or question(s) the review addresses.	Yes
METHODS
Eligibility criteria	3	Specify the inclusion and exclusion criteria for the review.	Yes
Information sources	4	Specify the information sources (e.g., databases, registers) used to identify studies and the date when each was last searched.	Yes
Risk of bias	5	Specify the methods used to assess risk of bias in the included studies.	Yes
Synthesis of results	6	Specify the methods used to present and synthesise results.	Yes
RESULTS
Included studies	7	Give the total number of included studies and participants and summarise relevant characteristics of studies.	Yes
Synthesis of results	8	Present results for main outcomes, preferably indicating the number of studies included and participants for each. If meta-analysis was done, report the summary estimate and confidence/credible interval. If comparing groups, indicate the direction of the effect (i.e., which group is favored).	Yes
DISCUSSION
Limitations of evidence	9	Provide a brief summary of the limitations of the evidence included in the review (e.g., study risk of bias, inconsistency and imprecision).	Yes
Interpretation	10	Provide a general interpretation of the results and important implications.	Yes
OTHER
Funding	11	Specify the primary source of funding for the review.	No
Registration	12	Provide the register name and registration number.	No

Table A1. PRISMA abstract checklist.

Section and Topic	Item #	Checklist Item	Reported (Yes/No)
TITLE
Title	1	Identify the report as a systematic review.	Yes
BACKGROUND
Objectives	2	Provide an explicit statement of the main objective(s) or question(s) the review addresses.	Yes
METHODS
Eligibility criteria	3	Specify the inclusion and exclusion criteria for the review.	Yes
Information sources	4	Specify the information sources (e.g., databases, registers) used to identify studies and the date when each was last searched.	Yes
Risk of bias	5	Specify the methods used to assess risk of bias in the included studies.	Yes
Synthesis of results	6	Specify the methods used to present and synthesise results.	Yes
RESULTS
Included studies	7	Give the total number of included studies and participants and summarise relevant characteristics of studies.	Yes
Synthesis of results	8	Present results for main outcomes, preferably indicating the number of studies included and participants for each. If meta-analysis was done, report the summary estimate and confidence/credible interval. If comparing groups, indicate the direction of the effect (i.e., which group is favored).	Yes
DISCUSSION
Limitations of evidence	9	Provide a brief summary of the limitations of the evidence included in the review (e.g., study risk of bias, inconsistency and imprecision).	Yes
Interpretation	10	Provide a general interpretation of the results and important implications.	Yes
OTHER
Funding	11	Specify the primary source of funding for the review.	No
Registration	12	Provide the register name and registration number.	No

Appendix A.2. The Summarised PRISMA 2020 Checklist

Table A2. PRISMA 2020 checklist. Topic: Feature-Centric AI for Breast Cancer Metastasis Analytics: A Systematic Review and Evidence-Informed Conceptual Framework for Feature Engineering and Multimodal Integration.

Section and Topic	Item #	Checklist Item	Location Where Item Is Reported
TITLE
Title	1	Identify the report as a systematic review.	Title block, page 1: the article title identifies the report as “A Systematic Review” and specifies the feature-centric breast cancer metastasis analytics focus.
ABSTRACT
Abstract	2	See the PRISMA 2020 for Abstracts checklist.	Abstract, page 1: reports PRISMA 2020 guidance, databases searched, publication period, 50 included empirical studies, synthesis focus, main findings and the 5Fs framework. Section 2.1, paragraph 1 also points to Appendix A.1 for the PRISMA 2020 for Abstracts Checklist.
INTRODUCTION
Rationale	3	Describe the rationale for the review in the context of existing knowledge.	Section 1, paragraphs 1–5: explains breast cancer burden, metastatic dissemination, limitations of conventional staging/imaging, growth of AI/radiomics, fragmented prior reviews and the need for feature-engineering synthesis in breast cancer metastasis analytics.
Objectives	4	Provide an explicit statement of the objective(s) or question(s) the review addresses.	Section 1, Research Questions 1–3: states the three review questions on feature-engineering techniques, feature types and the evidence-based feature-centric framework. Section Technical Contributions further specifies the technical contributions, including the 5Fs framework.
METHODS
Eligibility criteria	5	Specify the inclusion and exclusion criteria for the review and how studies were grouped for the syntheses.	Section 2.3, paragraph 1: states the inclusion and exclusion criteria. Section 2.6, paragraph 1 and Section 3.4, paragraph 1 explain how studies were grouped for thematic synthesis by feature-processing and engineering techniques and feature types, with Table 1 presenting the grouped synthesis.
Information sources	6	Specify all databases, registers, websites, organisations, reference lists and other sources searched or consulted to identify studies. Specify the date when each source was last searched or consulted.	Section 2.2, paragraph 1: identifies MEDLINE, Scopus, Web of Science, Embase and IEEE Xplore as the electronic databases searched between 1 May 2026 and 15 May 2026 and also reports manual screening of reference lists from eligible articles and relevant review studies.
Search strategy	7	Present the full search strategies for all databases, registers and websites, including any filters and limits used.	Section 2.2, paragraph 1: reports the main keywords, Boolean operators and an example search string.
Selection process	8	Specify the methods used to decide whether a study met the inclusion criteria of the review, including how many reviewers screened each record and each report retrieved, whether they worked independently, and if applicable, details of automation tools used in the process.	Section 2.4, paragraph 1: describes export to a reference-management system, electronic and manual duplicate removal, title/abstract screening, full-text eligibility assessment, independent screening and consensus resolution of disagreements.
Data collection process	9	Specify the methods used to collect data from reports, including how many reviewers collected data from each report, whether they worked independently, any processes for obtaining or confirming data from study investigators, and if applicable, details of automation tools used in the process.	Section 2.6, paragraph 1: states that a structured extraction table aligned with the research questions was used to collect author details, publication year, study aim, methods, feature-engineering techniques, feature types, AI approaches and model-performance results.
Data items	10a	List and define all outcomes for which data were sought. Specify whether all results compatible with each outcome domain in each study were sought and, if not, the methods used to decide which results to collect.	Section 2.3, paragraph 1; Section 3.2, paragraph 1; and Section 3.4.2, paragraphs 1 and 6: metastasis-related outcomes included axillary lymph node metastasis, sentinel lymph node metastasis, high axillary lymph-node burden, distant metastasis, bone/liver/lung/brain metastasis, recurrence/metastasis risk, metabolic response and survival-related endpoints. Section 2.6, paragraph 1 states that model performance indicators were extracted when reported.
Data items	10b	List and define all other variables for which data were sought (e.g., participant and intervention characteristics, funding sources). Describe any assumptions made about any missing or unclear information.	Section 2.6, paragraph 1 and Appendix C: extracted variables included author details, publication year, study aim, study methods, feature-engineering techniques, feature types, AI approaches and model-performance results.
Study risk of bias assessment	11	Specify the methods used to assess risk of bias in the included studies, including details of the tool(s) used, how many reviewers assessed each study and whether they worked independently, and if applicable, details of automation tools used in the process.	Section 2.5, paragraphs 1–3; Section 3.3, paragraph 1; and Appendix B: methodological quality was assessed using the Mixed Methods Appraisal Tool (MMAT), operationalised through screening questions and criteria 3.1–3.5, with additional appraisal of processing pipelines, feature selection, validation, explainable AI, calibration and multimodal integration.
Effect measures	12	Specify for each outcome the effect measure(s) used in the synthesis or presentation of results.	Section 2.6, paragraph 1 states that model-performance indicators were extracted.
Synthesis methods	13a	Describe the processes used to decide which studies were eligible for each synthesis.	Section 2.3, paragraph 1; Section 2.6, paragraph 1; Section 3.4, paragraph 1; and Table 1: studies were eligible for synthesis if they met the empirical AI-based breast cancer metastasis analytics criteria and contributed evidence to the planned synthesis domains of feature-processing and engineering techniques, feature types, multimodal integration, explainability and framework development.
Synthesis methods	13b	Describe any methods required to prepare the data for presentation or synthesis, such as handling of missing summary statistics, or data conversions.	Section 2.6, paragraph 1 and Section 3.4, paragraph 1: extracted evidence was organised into thematic codes, basic themes, organising themes and grand themes.
Synthesis methods	13c	Describe any methods used to tabulate or visually display results of individual studies and syntheses.	Section 3.1, paragraph 1 and Figure 1 display the study-selection process; Section 3.4, paragraph 1 and Table 1 display the thematic synthesis of feature-engineering techniques and feature types; Section 3.5, paragraphs 1–2 and Figure 2 display the 5Fs framework; Appendix B presents quality appraisal; and Appendix C presents the extracted study summary table.
Synthesis methods	13d	Describe any methods used to synthesise results and provide a rationale for the choice(s). If meta-analysis was performed, describe the model(s), method(s) to identify heterogeneity, and software package(s) used.	Section 2.6, paragraph 1: thematic synthesis was guided by Attride-Stirling’s thematic network analysis, with thematic codes refined into basic themes, organising themes and grand themes.
Synthesis methods	13e
Synthesis methods	13f	Describe any sensitivity analyses conducted to assess robustness of the synthesised results.	Sensitivity analyses were not conducted because no pooled quantitative synthesis was undertaken.
Reporting bias assessment	14	Describe any methods used to assess risk of bias due to missing results in a synthesis arising from reporting biases.	Reporting bias due to missing results was not statistically assessed because the review used heterogeneous narrative/thematic synthesis; note that MMAT quality appraisal and reporting transparency were considered when interpreting the evidence.
Certainty assessment	15	Describe any methods used to assess certainty or confidence in the body of evidence for an outcome.	Section 2.5, paragraphs 1–3; Section 3.3, paragraph 1; Section 6, paragraph 1; and Appendix B support narrative confidence assessment using MMAT-informed methodological quality, validation, reproducibility, reporting completeness and convergence across feature domains.
RESULTS
Study selection	16a	Describe the results of the search and selection process, from the number of records identified in the search to the number of studies included in the review, ideally using a flow diagram.	Section 3.1, paragraph 1 and Figure 1: database searches identified 1772 records from PubMed/MEDLINE (n = 412), Scopus (n = 536), Web of Science Core Collection (n = 287), Embase (n = 391) and IEEE Xplore (n = 146); 314 duplicates were removed; 1458 records were screened; 1260 records were excluded; and 50 studies were included.
Study selection	16b	Cite studies that might appear to meet the inclusion criteria, but which were excluded, and explain why they were excluded.	Section 3.1, paragraph 1 and Figure 1: full-text exclusion reasons included no information on metastasis/not AI/ML-driven metastasis analytics, no breast cancer-specific metastasis focus, insufficient/no feature-engineering details, non-empirical designs, review-only publications and non-human/preclinical studies.
Study characteristics	17	Cite each included study and present its characteristics.	Section 3.2, paragraphs 1–3 summarises the 50 included studies by metastasis endpoint, publication year, study design and AI algorithms. Appendix C contains the data extraction summary of included articles, including author/year, aim, methods, feature-engineering techniques, feature types, AI approaches and results.
Risk of bias in studies	18	Present assessments of risk of bias for each included study.	Section 2.5, paragraphs 1–3; Section 3.3, paragraph 1; and Appendix B present the study-level methodological quality/risk-of-bias appraisal using MMAT. Section 3.3, paragraph 1 summarises 43 high-quality studies (86%) and 7 moderate-quality studies (14%).
Results of individual studies	19	For all outcomes, present, for each study: summary statistics for each group where appropriate and an effect estimate and its precision, ideally using structured tables or plots.	Section 2.6, paragraph 1 and Appendix C present extracted study-level data, including model-performance indicators where reported. Section 3.2, paragraph 3 summarises AI algorithm distributions, while Section 3.4 and Table 1 synthesise feature-engineering techniques and feature types.
Results of syntheses	20a	For each synthesis, briefly summarise the characteristics and risk of bias among contributing studies.	Section 3.2, paragraphs 1–3; Section 3.3, paragraph 1; Section 3.4, paragraph 1; Table 1; and Appendix B summarise the characteristics and quality appraisal of contributing studies.
Results of syntheses	20b	Present results of all statistical syntheses conducted. If meta-analysis was done, present the summary estimate and precision and measures of statistical heterogeneity. If comparing groups, describe the direction of the effect.	No meta-analysis or pooled statistical synthesis was conducted. Narrative/thematic synthesis results are presented in Section 3.4.1, paragraphs 1–5; Section 3.4.2, paragraphs 1–6; and Section 3.5, paragraphs 1–2.
Results of syntheses	20c	Present results of all investigations of possible causes of heterogeneity among study results.	Section 3.2, paragraphs 1–3; Section 3.4.1, paragraphs 1–5; Section 3.4.2, paragraphs 1–6; and Section 4, paragraphs 1–5 report heterogeneity narratively across metastasis outcome types, feature domains, feature-engineering workflows, imaging modalities, multimodal integration approaches, explainability methods and AI algorithms.
Results of syntheses	20d	Present results of all sensitivity analyses conducted to assess the robustness of the synthesised results.	No sensitivity analyses were conducted because there was no pooled quantitative synthesis.
Reporting biases	21	Present assessments of risk of bias due to missing results arising from reporting biases for each synthesis assessed.	Reporting bias due to missing results was not statistically assessed because evidence was heterogeneous and synthesised narratively/thematically.
Certainty of evidence	22	Present assessments of certainty or confidence in the body of evidence for each outcome assessed.	Section 3.3, paragraph 1; Section 2.5, paragraphs 1–3; Appendix B; and Section 6, paragraph 1 can be used to report narrative confidence in the evidence.
DISCUSSION
Discussion	23a	Provide a general interpretation of the results in the context of other evidence.	Section 4, paragraphs 1–5: interprets the dominance of radiomics extraction, preprocessing, LASSO/mRMR selection, multimodal integration and explainable AI in relation to prior oncology, radiomics, explainability and AI governance evidence.
Discussion	23b	Discuss any limitations of the evidence included in the review.	Section 6, paragraph 1 discusses methodological heterogeneity, variability in reporting and limited external validation across included studies. Section 4, paragraph 2 also discusses reproducibility and standardisation barriers in radiomics pipelines.
Discussion	23c	Discuss any limitations of the review processes used.	Section 2.3, paragraph 1 identifies review-process restrictions to English-language, full-text, peer-reviewed empirical studies from 2020–2026. Section 6, paragraph 1 discusses comparability and generalisability limitations arising from heterogeneity and limited external validation.
Discussion	23d	Discuss implications of the results for practice, policy, and future research.	Section 4, paragraph 5; Section 5, paragraphs 1–2; and Section 7, paragraph 1 discuss implications for standardised feature-engineering pipelines, prospective multicenter validation, radiomics/pathomics harmonisation, multimodal integration, explainable AI, clinical trust and decision-support deployment.
OTHER INFORMATION
Registration and protocol	24a	Provide registration information for the review, including register name and registration number, or state that the review was not registered.	This systematic review was not registered.
Registration and protocol	24b	Indicate where the review protocol can be accessed, or state that a protocol was not prepared.	Section 2.1, paragraph 1 states that a structured review protocol was followed, but no public protocol location is reported.
Registration and protocol	24c	Describe and explain any amendments to information provided at registration or in the protocol.	The review was not registered and no public protocol was prepared.
Support	25	Describe sources of financial or non-financial support for the review, and the role of the funders or sponsors in the review.	Funding statement after Section 7: states that the research did not receive external financial support.
Competing interests	26	Declare any competing interests of review authors.	Conflicts of Interest statement after Section 7: states that the authors declare no conflicts of interest.
Availability of data, code and other materials	27	Report which of the following are publicly available and where they can be found template data collection forms; data extracted from included studies; data used for all analyses; analytic code; any other materials used in the review.	Data Availability Statement after Section 7: states that no datasets were generated or analysed during the current study. Appendix B and Appendix C are identified as locations for quality appraisal and for the extraction of study data.

Table A2. PRISMA 2020 checklist. Topic: Feature-Centric AI for Breast Cancer Metastasis Analytics: A Systematic Review and Evidence-Informed Conceptual Framework for Feature Engineering and Multimodal Integration.

Section and Topic	Item #	Checklist Item	Location Where Item Is Reported
TITLE
Title	1	Identify the report as a systematic review.	Title block, page 1: the article title identifies the report as “A Systematic Review” and specifies the feature-centric breast cancer metastasis analytics focus.
ABSTRACT
Abstract	2	See the PRISMA 2020 for Abstracts checklist.	Abstract, page 1: reports PRISMA 2020 guidance, databases searched, publication period, 50 included empirical studies, synthesis focus, main findings and the 5Fs framework. Section 2.1, paragraph 1 also points to Appendix A.1 for the PRISMA 2020 for Abstracts Checklist.
INTRODUCTION
Rationale	3	Describe the rationale for the review in the context of existing knowledge.	Section 1, paragraphs 1–5: explains breast cancer burden, metastatic dissemination, limitations of conventional staging/imaging, growth of AI/radiomics, fragmented prior reviews and the need for feature-engineering synthesis in breast cancer metastasis analytics.
Objectives	4	Provide an explicit statement of the objective(s) or question(s) the review addresses.	Section 1, Research Questions 1–3: states the three review questions on feature-engineering techniques, feature types and the evidence-based feature-centric framework. Section Technical Contributions further specifies the technical contributions, including the 5Fs framework.
METHODS
Eligibility criteria	5	Specify the inclusion and exclusion criteria for the review and how studies were grouped for the syntheses.	Section 2.3, paragraph 1: states the inclusion and exclusion criteria. Section 2.6, paragraph 1 and Section 3.4, paragraph 1 explain how studies were grouped for thematic synthesis by feature-processing and engineering techniques and feature types, with Table 1 presenting the grouped synthesis.
Information sources	6	Specify all databases, registers, websites, organisations, reference lists and other sources searched or consulted to identify studies. Specify the date when each source was last searched or consulted.	Section 2.2, paragraph 1: identifies MEDLINE, Scopus, Web of Science, Embase and IEEE Xplore as the electronic databases searched between 1 May 2026 and 15 May 2026 and also reports manual screening of reference lists from eligible articles and relevant review studies.
Search strategy	7	Present the full search strategies for all databases, registers and websites, including any filters and limits used.	Section 2.2, paragraph 1: reports the main keywords, Boolean operators and an example search string.
Selection process	8	Specify the methods used to decide whether a study met the inclusion criteria of the review, including how many reviewers screened each record and each report retrieved, whether they worked independently, and if applicable, details of automation tools used in the process.	Section 2.4, paragraph 1: describes export to a reference-management system, electronic and manual duplicate removal, title/abstract screening, full-text eligibility assessment, independent screening and consensus resolution of disagreements.
Data collection process	9	Specify the methods used to collect data from reports, including how many reviewers collected data from each report, whether they worked independently, any processes for obtaining or confirming data from study investigators, and if applicable, details of automation tools used in the process.	Section 2.6, paragraph 1: states that a structured extraction table aligned with the research questions was used to collect author details, publication year, study aim, methods, feature-engineering techniques, feature types, AI approaches and model-performance results.
Data items	10a	List and define all outcomes for which data were sought. Specify whether all results compatible with each outcome domain in each study were sought and, if not, the methods used to decide which results to collect.	Section 2.3, paragraph 1; Section 3.2, paragraph 1; and Section 3.4.2, paragraphs 1 and 6: metastasis-related outcomes included axillary lymph node metastasis, sentinel lymph node metastasis, high axillary lymph-node burden, distant metastasis, bone/liver/lung/brain metastasis, recurrence/metastasis risk, metabolic response and survival-related endpoints. Section 2.6, paragraph 1 states that model performance indicators were extracted when reported.
Data items	10b	List and define all other variables for which data were sought (e.g., participant and intervention characteristics, funding sources). Describe any assumptions made about any missing or unclear information.	Section 2.6, paragraph 1 and Appendix C: extracted variables included author details, publication year, study aim, study methods, feature-engineering techniques, feature types, AI approaches and model-performance results.
Study risk of bias assessment	11	Specify the methods used to assess risk of bias in the included studies, including details of the tool(s) used, how many reviewers assessed each study and whether they worked independently, and if applicable, details of automation tools used in the process.	Section 2.5, paragraphs 1–3; Section 3.3, paragraph 1; and Appendix B: methodological quality was assessed using the Mixed Methods Appraisal Tool (MMAT), operationalised through screening questions and criteria 3.1–3.5, with additional appraisal of processing pipelines, feature selection, validation, explainable AI, calibration and multimodal integration.
Effect measures	12	Specify for each outcome the effect measure(s) used in the synthesis or presentation of results.	Section 2.6, paragraph 1 states that model-performance indicators were extracted.
Synthesis methods	13a	Describe the processes used to decide which studies were eligible for each synthesis.	Section 2.3, paragraph 1; Section 2.6, paragraph 1; Section 3.4, paragraph 1; and Table 1: studies were eligible for synthesis if they met the empirical AI-based breast cancer metastasis analytics criteria and contributed evidence to the planned synthesis domains of feature-processing and engineering techniques, feature types, multimodal integration, explainability and framework development.
Synthesis methods	13b	Describe any methods required to prepare the data for presentation or synthesis, such as handling of missing summary statistics, or data conversions.	Section 2.6, paragraph 1 and Section 3.4, paragraph 1: extracted evidence was organised into thematic codes, basic themes, organising themes and grand themes.
Synthesis methods	13c	Describe any methods used to tabulate or visually display results of individual studies and syntheses.	Section 3.1, paragraph 1 and Figure 1 display the study-selection process; Section 3.4, paragraph 1 and Table 1 display the thematic synthesis of feature-engineering techniques and feature types; Section 3.5, paragraphs 1–2 and Figure 2 display the 5Fs framework; Appendix B presents quality appraisal; and Appendix C presents the extracted study summary table.
Synthesis methods	13d	Describe any methods used to synthesise results and provide a rationale for the choice(s). If meta-analysis was performed, describe the model(s), method(s) to identify heterogeneity, and software package(s) used.	Section 2.6, paragraph 1: thematic synthesis was guided by Attride-Stirling’s thematic network analysis, with thematic codes refined into basic themes, organising themes and grand themes.
Synthesis methods	13e
Synthesis methods	13f	Describe any sensitivity analyses conducted to assess robustness of the synthesised results.	Sensitivity analyses were not conducted because no pooled quantitative synthesis was undertaken.
Reporting bias assessment	14	Describe any methods used to assess risk of bias due to missing results in a synthesis arising from reporting biases.	Reporting bias due to missing results was not statistically assessed because the review used heterogeneous narrative/thematic synthesis; note that MMAT quality appraisal and reporting transparency were considered when interpreting the evidence.
Certainty assessment	15	Describe any methods used to assess certainty or confidence in the body of evidence for an outcome.	Section 2.5, paragraphs 1–3; Section 3.3, paragraph 1; Section 6, paragraph 1; and Appendix B support narrative confidence assessment using MMAT-informed methodological quality, validation, reproducibility, reporting completeness and convergence across feature domains.
RESULTS
Study selection	16a	Describe the results of the search and selection process, from the number of records identified in the search to the number of studies included in the review, ideally using a flow diagram.	Section 3.1, paragraph 1 and Figure 1: database searches identified 1772 records from PubMed/MEDLINE (n = 412), Scopus (n = 536), Web of Science Core Collection (n = 287), Embase (n = 391) and IEEE Xplore (n = 146); 314 duplicates were removed; 1458 records were screened; 1260 records were excluded; and 50 studies were included.
Study selection	16b	Cite studies that might appear to meet the inclusion criteria, but which were excluded, and explain why they were excluded.	Section 3.1, paragraph 1 and Figure 1: full-text exclusion reasons included no information on metastasis/not AI/ML-driven metastasis analytics, no breast cancer-specific metastasis focus, insufficient/no feature-engineering details, non-empirical designs, review-only publications and non-human/preclinical studies.
Study characteristics	17	Cite each included study and present its characteristics.	Section 3.2, paragraphs 1–3 summarises the 50 included studies by metastasis endpoint, publication year, study design and AI algorithms. Appendix C contains the data extraction summary of included articles, including author/year, aim, methods, feature-engineering techniques, feature types, AI approaches and results.
Risk of bias in studies	18	Present assessments of risk of bias for each included study.	Section 2.5, paragraphs 1–3; Section 3.3, paragraph 1; and Appendix B present the study-level methodological quality/risk-of-bias appraisal using MMAT. Section 3.3, paragraph 1 summarises 43 high-quality studies (86%) and 7 moderate-quality studies (14%).
Results of individual studies	19	For all outcomes, present, for each study: summary statistics for each group where appropriate and an effect estimate and its precision, ideally using structured tables or plots.	Section 2.6, paragraph 1 and Appendix C present extracted study-level data, including model-performance indicators where reported. Section 3.2, paragraph 3 summarises AI algorithm distributions, while Section 3.4 and Table 1 synthesise feature-engineering techniques and feature types.
Results of syntheses	20a	For each synthesis, briefly summarise the characteristics and risk of bias among contributing studies.	Section 3.2, paragraphs 1–3; Section 3.3, paragraph 1; Section 3.4, paragraph 1; Table 1; and Appendix B summarise the characteristics and quality appraisal of contributing studies.
Results of syntheses	20b	Present results of all statistical syntheses conducted. If meta-analysis was done, present the summary estimate and precision and measures of statistical heterogeneity. If comparing groups, describe the direction of the effect.	No meta-analysis or pooled statistical synthesis was conducted. Narrative/thematic synthesis results are presented in Section 3.4.1, paragraphs 1–5; Section 3.4.2, paragraphs 1–6; and Section 3.5, paragraphs 1–2.
Results of syntheses	20c	Present results of all investigations of possible causes of heterogeneity among study results.	Section 3.2, paragraphs 1–3; Section 3.4.1, paragraphs 1–5; Section 3.4.2, paragraphs 1–6; and Section 4, paragraphs 1–5 report heterogeneity narratively across metastasis outcome types, feature domains, feature-engineering workflows, imaging modalities, multimodal integration approaches, explainability methods and AI algorithms.
Results of syntheses	20d	Present results of all sensitivity analyses conducted to assess the robustness of the synthesised results.	No sensitivity analyses were conducted because there was no pooled quantitative synthesis.
Reporting biases	21	Present assessments of risk of bias due to missing results arising from reporting biases for each synthesis assessed.	Reporting bias due to missing results was not statistically assessed because evidence was heterogeneous and synthesised narratively/thematically.
Certainty of evidence	22	Present assessments of certainty or confidence in the body of evidence for each outcome assessed.	Section 3.3, paragraph 1; Section 2.5, paragraphs 1–3; Appendix B; and Section 6, paragraph 1 can be used to report narrative confidence in the evidence.
DISCUSSION
Discussion	23a	Provide a general interpretation of the results in the context of other evidence.	Section 4, paragraphs 1–5: interprets the dominance of radiomics extraction, preprocessing, LASSO/mRMR selection, multimodal integration and explainable AI in relation to prior oncology, radiomics, explainability and AI governance evidence.
Discussion	23b	Discuss any limitations of the evidence included in the review.	Section 6, paragraph 1 discusses methodological heterogeneity, variability in reporting and limited external validation across included studies. Section 4, paragraph 2 also discusses reproducibility and standardisation barriers in radiomics pipelines.
Discussion	23c	Discuss any limitations of the review processes used.	Section 2.3, paragraph 1 identifies review-process restrictions to English-language, full-text, peer-reviewed empirical studies from 2020–2026. Section 6, paragraph 1 discusses comparability and generalisability limitations arising from heterogeneity and limited external validation.
Discussion	23d	Discuss implications of the results for practice, policy, and future research.	Section 4, paragraph 5; Section 5, paragraphs 1–2; and Section 7, paragraph 1 discuss implications for standardised feature-engineering pipelines, prospective multicenter validation, radiomics/pathomics harmonisation, multimodal integration, explainable AI, clinical trust and decision-support deployment.
OTHER INFORMATION
Registration and protocol	24a	Provide registration information for the review, including register name and registration number, or state that the review was not registered.	This systematic review was not registered.
Registration and protocol	24b	Indicate where the review protocol can be accessed, or state that a protocol was not prepared.	Section 2.1, paragraph 1 states that a structured review protocol was followed, but no public protocol location is reported.
Registration and protocol	24c	Describe and explain any amendments to information provided at registration or in the protocol.	The review was not registered and no public protocol was prepared.
Support	25	Describe sources of financial or non-financial support for the review, and the role of the funders or sponsors in the review.	Funding statement after Section 7: states that the research did not receive external financial support.
Competing interests	26	Declare any competing interests of review authors.	Conflicts of Interest statement after Section 7: states that the authors declare no conflicts of interest.
Availability of data, code and other materials	27	Report which of the following are publicly available and where they can be found template data collection forms; data extracted from included studies; data used for all analyses; analytic code; any other materials used in the review.	Data Availability Statement after Section 7: states that no datasets were generated or analysed during the current study. Appendix B and Appendix C are identified as locations for quality appraisal and for the extraction of study data.