A formal explanation space for the simultaneous clustering of neurologic diseases based on their signs and symptoms Authors: Raghu Yelugam, Daniel B. Hier, Tayo Obafemi-Ajayi, Michael D. Carrithers, Donald C. Wunsch II Published: 2025 ABSTRACT ObjectiveClustering is widely used to identify meaningful subgroups in biomedical data, but interpretation remains challenging, especially in the absence of ground-truth labels. Moreover, clustering often produces multiple plausible solutions without a single correct answer. Using dementia phenotypes as a case study, we introduce a Formal Explanation Space (FES) to improve interpretability and facilitate comparison across competing cluster solutions.MethodsWe used spectral coclustering and spectral biclustering to cluster a dataset of dementia patients based on clinical phenotypes (signs and symptoms). To enhance interpretability, we constructed an FES to explain algorithm behavior, assess cluster quality, identify influential features, and characterize cluster composition. Although simultaneous clustering is unsupervised, interpretation was aided by diagnostic labels, which we used for external validation of cluster composition.ResultsSpectral coclustering and spectral biclustering each identified five biologically plausible dementia subgroups, though subgroup composition differed by method. The FES provided a structured framework for comparing these divergent outputs.ConclusionsClustering complex biomedical data often produces multiple biologically plausible solutions. Retaining and comparing such solutions within a formal explanation space enhances interpretability and supports the discovery of complementary insights across methods.Supplementary informationThe online version contains supplementary material available at 10.1186/s12911-025-03297-w. Introduction The proliferation of artificial intelligence (AI) in medicine has produced promising results [1–3], yet a persistent gap remains: clinical domain experts are not consistently integrated into model development in ways that optimize design and evaluation. As a result, technically sophisticated systems can overlook subtle but clinically important patterns, and their outputs may appear opaque to clinicians who lack formal AI training. Limited clinician involvement weakens safeguards against bias, reducing confidence in model reliability and trustworthiness. These challenges are compounded by the sensitivity of many models to small changes in training data, underscoring the need for systematic, clinician-centered evaluation before adoption. A practical way to address this gap is to make model outputs more accessible and interpretable to clinicians with modest AI training so that they can validate findings, identify biases, and inform model refinement. We propose that a Formal Explanation Space (FES) can make outputs of complex AI and machine learning models more interpretable to domain experts. As a use case, we apply FES to the interpretation of unsupervised learning results–specifically, the biclustering and coclustering of dementia patients by their phenotypic profiles. These methods often produce multiple, equally plausible solutions. By comparing them within a common explanation space, we show how FES enables domain experts to assess biological plausibility and extract complementary clinical insights from divergent clustering outputs. The interpretation of unsupervised learning results is especially challenging when no single partition of the data emerges as best [4–7]. Expert judgment becomes essential in evaluating whether identified groups are clinically meaningful [8–11]. Biclustering methods are particularly well-suited to neurological phenotype data, as they allow simultaneous clustering of patients (rows) and phenotypes (columns). This approach has the potential to uncover clinically relevant subgroups with distinct diagnostic or prognostic profiles [12–15]. Phenotypic heterogeneity within neurological diseases poses challenges analogous to those in gene-expression analysis. In classical genomic biclustering, rows are genes and columns are experimental conditions; in phenotypic datasets, rows are patients and columns are symptoms. Biclustering seeks coherent submatrices where subsets of patients share subsets of phenotypes, thereby revealing subtypes and symptom modules. When diagnostic labels are available, they can be used post hoc to assess which diseases co-cluster and which phenotypes are enriched, providing external validation. Despite this promise, adoption in healthcare has been limited by uncertainty in algorithm selection and by the difficulty of interpreting and validating bicluster outputs [16]. Dementia provides an instructive test case. It encompasses a heterogeneous group of disorders—including Alzheimer’s disease (AD), frontotemporal dementia (FTD), Lewy body dementia (LBD), Huntington’s disease (HD), Creutzfeldt–Jakob disease (CJD), and normal pressure hydrocephalus (NPH)—that share the core symptom of cognitive impairment but differ in features such as tremor, hyperkinesia, or sphincter disturbance [17]. Symptom variability occurs within a disease (e.g., some AD patients present with speech defects, others do not), and symptom overlap is common across diseases (e.g., sleep disturbance occurs in both LBD and AD). This mix of within-disease heterogeneity and between-disease overlap makes dementia an ideal candidate for clustering approaches. In this study, we applied spectral coclustering and spectral biclustering to dementia phenotypes, excluding diagnostic labels from the algorithms but retaining them for validation. We then constructed a FES to systematically compare solutions and improve interpretability. By simultaneously clustering patients (rows) and symptoms (columns), this framework creates opportunities to answer biologically meaningful questions such as: Which diseases cluster together? Which signs and symptoms for clusters? And what diagnoses and what signs and symptoms enrich each dementia subtype? These questions guided our analysis of the cluster solutions and the assessment of their biological plausibility. Our contributions are threefold: (1) we demonstrate the feasibility of constructing a formal explanation space for biclustering, (2) we show how FES increases the interpretability of cluster solutions, and (3) we argue for retaining multiple plausible outputs rather than enforcing a single best solution [18, 19]. The remainder of this paper is organized as follows. Methods introduces our naming convention for biclusters and coclusters and describes the pipeline that converts lists of signs and symptoms into vectors, clusters these vectors by spectral methods, and constructs the explanation space. Results presents the cluster solutions derived from the dementia dataset. Discussion considers how the FES supports assessment of biological plausibility and side-by-side comparison of competing cluster outputs. We conclude with limitations and future directions. To promote reproducibility and transparency, the dataset and methods are publicly available at https://github.com/ACIL-Group/formal_explanation_space.git. Methods Data acquisition and pre-processing We analyzed 87 dementia cases extracted from neurology textbooks [20, 21]. No protected health information was used. Each case was assigned a ground-truth diagnosis—Alzheimer’s disease (AD), frontotemporal dementia (FTD), Huntington’s disease (HD), Creutzfeldt–Jakob disease (CJD), Lewy body dementia (LBD), or normal pressure hydrocephalus (NPH)—based on the textbook description. Signs and symptoms were mapped to standardized neuro-ontology terms, converting expressions such as slowness to bradykinesia [22, 23]. The ontology contained approximately 1,600 neurological terms drawn from SNOMED CT and the Human Phenotype Ontology [23, 24]. To reduce complexity, we applied a roll-up strategy that consolidated specific terms into broader categories [25]. For example, chorea, dyskinesia, and athetosis were grouped under hyperkinesia. Where possible, this grouping followed the hierarchical structure of the ontology [23]. The process reduced the feature set to 15 categories. Each case was then represented as a vector of category frequencies, normalized to the interval [0, 1] to account for differences in reporting frequency across cases. Vectors were labeled with case number and ground-truth diagnosis and exported as a CSV file for clustering analysis. Cluster naming convention To enable consistent comparison across clustering methods, we adopted a uniform naming convention for cluster partitions. Row coclusters are denoted as Co-0, Co-1, …, Co-k, while row biclusters are denoted as Bi-0, Bi-1, …, Bi-k. This convention is applied uniformly throughout the paper in figures, tables, and text, facilitating direct comparison between spectral coclustering and biclustering solutions. Clustering Definitions. The cluster configuration refers to the specific combination of the dataset, the selected clustering algorithm, the number of clusters (n_clusters), and other relevant algorithm parameters. The cluster solution is generated by applying the cluster configuration to the data and consists of the assigned cluster labels for each row and each column in the dataset. Selection of clustering algorithm. The choice of clustering algorithm depends on several factors, such as the data type (numerical, categorical, or mixed), dimensionality, noise levels, and domain-specific knowledge, such as the expected number of clusters [26–29]. Algorithms suited for genomic data may not perform as well for phenomic data [16]. Since we aimed to cluster both rows (cases) and columns (signs and symptoms), we selected eight biclustering algorithms for initial tests to choose two algorithms for further studies with FES. The eight methods are the Bi-correlation Clustering Algorithm (BCCA), Cheng Church Algorithm (CCA), Fabia, Order Preserving Sub Matrix (OPSM), Plaid, Qualitative Biclustering (QUBIC), Spectral Biclustering (SpecBi), and Spectral co-clustering (SpecCo) [30–37]. The open implementations of BCCA, CCA, OPSM, QUBIC, Spectral Biclustering and Spectral Coclustering in Python were obtained from scikit-learn and biclustlib and open implementation of Fabia was obtained from R Fabia package. We found the recent biarchetype analysis (biAA) subsequent to our submission of this research [38]. BiAA is based on archetype analysis which identifies archetypes, points that lie on the boundary of the convex hull of the data, and represents data as combination of archetypes [39]. Since archetypes explain structure of data more qualitatively than methods that look at average prototype, biAA can be used in place of conventional biclustering where a smooth prototype mixture model is desired. We believe [38] and this article both belong in a portfolio of explainable AI methods in biomedical applications. Cluster validity indices (CVIs) are used to assess the quality of the clusters identified by a clustering algorithm on presented data. Some CVIs only require ground truth and partitioning produced by a clustering algorithm to evaluate the clustering quality. Such methods are called external CVIs as they do not assess the compactness of clusters. Internal CVIs, on the other hand, use the data and partitioning determined by a clustering algorithm to evaluate the compactness and separation of the identified clusters. We used various external CVIs, the adjusted Rand index, mutual information, homogeneity score, completeness score, and V-measure score, and internal CVIs, Calinski-Harabasz index, Davies-Bouldin index and silhouette score, as implemented in scikit-learn, to determine the quality of row clusters produced during biclustering. These CVIs were chosen based on the availability of ground truth pertinent to row clusters and consideration for cluster compactness; note that column clusters are not considered for initial algorithm selection. For CVIs adjusted Rand index, Calinski-Harabasz index, mutual information, homogeneity score, completeness score, V-measure score, and silhouette score, higher scores correspond to better partitioning. For the Davies-Bouldin index, lower scores correspond to better partitioning. The ranges for CVIs used for this study are provided in Table S2 of Supplementary material. During experimentation, we observed that some biclustering methods identified biclusters with overlapping rows and columns. Biologically, an overlap in rows corresponds to observations sharing phenotypes common to two different disease subgroups, thus a potential for misdiagnosis and prompting further expert investigation. To compute CVI values using row ground truth labels, the overlapping rows of biclusters were considered to form a cluster different from the remaining rows. For example, if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R_{B_{j}} = \{r_{i}^{B_{j}} | r_{i}^{B_{j}} \in B_{j} \} $$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ R_{B_{k}} = \{r_{i}^{B_{k}} | r_{i}^{B_{k}} \in B_{k} \}$$\end{document} represent rows of biclusters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$B_{j} $$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$B_{k}$$\end{document} respectively, then the row clusters are redefined as the following. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} C_{1} &= R_{B_{j}} \setminus R_{B_{j}} \cap R_{B_{k}}, \nonumber \\ C_{2} &= R_{B_{k}} \setminus R_{B_{j}} \cap R_{B_{k}}, \quad \mathrm{and} \nonumber \\ C_{3} &= R_{B_{j}} \cap R_{B_{k}} \nonumber\end{aligned}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\setminus$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\cap$$\end{document} represent set difference and set intersection. Further, some biclustering methods do not assign each row to a bicluster. In such cases, rows that do not have membership in any bicluster are considered to form a cluster. The hyperparameter choice for each biclustering algorithm and the observed best CVI values for each method are reported respectively in Tables S1 and S3 of Supplementary material. Throughout the experiments, whenever an algorithm requires the number of clusters, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ k $$\end{document}, as a hyperparameter, the value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document} was varied between [2, 12]. Any partitioning with partitions greater than 12 was considered biologically implausible. Since the reported best CVI values were observed for different parameter combination for each algorithm, for an interested reader to review, the observed CVI values for different parameter combinations are provided at https://github.com/ACIL-Group/formal_explanation_space.git. From Table S3 of Supplementary material, it can be noted that none of the algorithms identified partitions similar to the ground truth, which can be observed from low ARI values. Algorithm Plaid was observed to partition rows better, resulting in the best observed CVI scores. Despite producing better CVI scores, algorithms other than BCCA, Spectral Biclustering, and Spectral Coclustering did not form biclusters with all the dataset rows, thus leaving some unassigned rows. Moreover, BCCA formed many singleton clusters, thus resulting in the lowest observed ARI value. Due to these reasons, we selected the spectral coclustering and spectral biclustering algorithms to investigate further by constructing FES. The results corresponding to the experimentation are provided in https://github.com/ACIL-Group/formal_explanation_space.git, where an interested reader can review the hyperparameter combinations and corresponding observed CVI values. Spectral methods have demonstrated strong performance in identifying biclusters, as shown in a comparison of 16 state-of-the-art methods on 35 gene expression datasets [40]. Key advantages of spectral methods include their ability to handle both numerical and categorical data, intrinsic dimension reduction, flexibility in distance metrics, and suitability for processing nonlinear data structures. Spectral coclustering and spectral biclustering differ in their assumptions and in how they assign cluster numbers. Spectral coclustering treats the data matrix as bipartite such that rows and columns are clustered simultaneously and the resulting row and column cluster numbers are aligned. This produces a block-diagonal structure, where Co-0 for rows corresponds directly to Co-0 for columns. In contrast, spectral biclustering assigns cluster memberships to rows and columns independently, constrained only by the same \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document} value. As a result, row biclusters (Bi-0, Bi-1, …) and column biclusters are numbered separately and do not correspond directly. Biclustering typically produces a checkerboard pattern of enriched submatrices, whereas coclustering emphasizes a diagonal block structure. The complementary nature of these methods justifies examining both sets of solutions within the FES for comparison. Selection of hyperparameters and computational efficiency of Spectral methods. The hyperparameters used for spectral coclustering were n_clusters = k, init =’k-means++’, svd_method =’randomized’, n_svd_vecs = None, n_init = 10, random_state = 0. For spectral biclustering, the hyperparameters were n_clusters = (k, k), method =’bistochastic’, n_components = 6, init =’k-means++’, svd_method =’randomized’, n_init = 10, n_best = 3, random_state = 0. Both methods rely on eigenvalue decomposition, with computational complexity roughly estimated as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O(\min(m^2 \cdot n, m \cdot n^2))$$\end{document} for a data matrix of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ m \times n $$\end{document}. Given the small size of our dataset (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$87 \times 15$$\end{document}), computational complexity was not an issue for either algorithm. Selection of number of clusters for FES construction. Spectral coclustering requires specifying the parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document} (n_clusters). Although spectral biclustering allows for separate specifications of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{rows}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{columns}$$\end{document}, we simplified this by setting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{rows} = k_{columns}$$\end{document}. By setting the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k_{rows} = k_{columns}$$\end{document} for spectral biclustering, the algorithm is constrained to finding block diagonal structure within the data, where each column cluster is exclusively associated with one row cluster, instead of finding a checkerboard structure, where each column cluster can be associated with any identified row cluster and vice versa. This reduction, while allowing for comparison with other algorithms, facilitates better explaining observed partitioning. The identified partitionings may, nevertheless, not be optimal. To determine the optimal number of clusters, we evaluated cluster quality indices over a range of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document} values from 2 to 10. A value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document} greater than 10 was considered unlikely due to the size of the data set and the observed reduction in most CVI scores for both methods when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k > 10$$\end{document} as shown in Figure S1 of Supplementary material observed during initial tests. Moreover, even though both values of normalization hyperparameter method, ’log’ and ’bistochastic’, optimized different CVI scores during the preliminary tests, hyperparameter was set to ’bistochastic’ for later evaluation due the sparseness of the data. The remaining hyperparameters of both methods were set to the values reported in Table S1 of Supplementary material, the same as the initial tests. To assess the quality of clustering, previously mentioned CVIs were used. While these indices guided our initial settings, the final number of clusters was determined by additional factors, including the presence of low-membership clusters, the configuration of clusters in dimension reduction scatter plots, an analysis of cluster composition by features, and the domain expert’s assessment of the biological plausibility and interest of the cluster solution. Explanation space The explanation space was assembled from multiple complementary tools, including heat maps, word clouds, t-SNE plots, and SHAP plots. Cluster visualization. Cluster solutions were visualized with t-SNE scatter plots [41]. These plots color data points by cluster or ground-truth labels, projecting high-dimensional phenotypes into two dimensions while preserving local and global topological structure. t-SNE is a nonlinear dimensionality reduction method that emphasizes the preservation of local relationships but may not preserve the global structure, governed by hyperparameter - perplexity. Moreover, because of nonlinearity, t-SNE can expand dense regions while shrinking less populated regions of a high-dimensional space, thus distorting the size and density of the lower-dimensional projections. Therefore, determining an optimal value of perplexity by observing the generated plots is imperative. The t-SNE hyperparameters were tuned iteratively by varying perplexity between 5 and 87, while visually assessing the produced 2D point clouds to ensure well distributed cohesive clusters that are consistent with the current knowledge of the data. The hyperparameters were set to n_components = 2, perplexity = 22, n_iters = 3000, learning_rate = 2000, and random_state = 40. Class centroids were computed and displayed as previously described [42]. Although t-SNE supports explainability, well-separated far-apart clusters in a t-SNE plot may be proximal in the original high-dimensional space and vice versa; thus necessitating additional evaluation. Feature importance. Influential features driving cluster membership were identified using SHAP. Each cluster solution was fitted to an XGBClassifier from xgboost [43], and SHAP values were computed with shap.Explainer. SHAP provides local explanations that can be used to compute global feature importance for a model’s predictions. The local absolute SHAP values from shap.Explainer were averaged to compute global feature importance and are presented as bar graphs for each classification, with bar length corresponding to the average absolute SHAP value. Cluster composition. Feature means for each cluster were standardized to [0,1] and visualized as heat maps using Plotly 5.14.1 [44]. Word clouds were also generated to highlight dominant features. Additional heat maps and word clouds based on ground-truth label counts provided a complementary view of cluster composition. Biological plausibility. No universally accepted metrics exist for biological plausibility of machine learning outputs [45]. We defined biologically interesting as results or observations that provoke curiosity and have the potential to provide insights into biological mechanisms. For example, did the symptoms grouped together make sense? Did the grouped diseases make sense? Did the diseases that were grouped with symptoms make sense? The cluster solution was considered biologically plausible if these questions could be answered affirmatively. Each cluster solution provided implicit estimates of distances between diseases as row clusters, distances between symptoms as column clusters, and distances between diseases grouped with symptoms as biclusters; these implicit distances had to be plausible based on current neurological knowledge. Here, results were considered plausible if they aligned with established neurological knowledge [46, 47]. A neurology domain expert qualitatively evaluated each cluster solution for both plausibility and biological interest.Fig. 1Measures of cluster quality. Cluster validity metrics for spectral coclustering (a) and spectral biclustering (b) applied to the dementia dataset, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document} varied between 2 and 10. Metrics include adjusted Rand index (ARI), V-measure (V), completeness score (C), homogeneity score (H), mutual information (MI), Calinski–harabasz score (CH), Davies–Bouldin score (DB), and silhouette score (Sil). We plot the low membership cluster (LMC) measure, defined as the number of clusters with fewer than two members. The LMC curve indicates that solutions with more than five clusters produces unstable partitions with singleton clusters. Due to varying ranges of each metric, provided in Table S2 of supplementary material, the observed values of ARI, V, C, H and Sil are multiplied by 10 Measures of cluster quality. Cluster validity metrics for spectral coclustering (a) and spectral biclustering (b) applied to the dementia dataset, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document} varied between 2 and 10. Metrics include adjusted Rand index (ARI), V-measure (V), completeness score (C), homogeneity score (H), mutual information (MI), Calinski–harabasz score (CH), Davies–Bouldin score (DB), and silhouette score (Sil). We plot the low membership cluster (LMC) measure, defined as the number of clusters with fewer than two members. The LMC curve indicates that solutions with more than five clusters produces unstable partitions with singleton clusters. Due to varying ranges of each metric, provided in Table S2 of supplementary material, the observed values of ARI, V, C, H and Sil are multiplied by 10 Results We developed a Formal Explanation Space (FES) for simultaneous clustering of dementia cases, where rows represent cases (each with a diagnosis) and columns represent features (signs and symptoms). Unlike traditional clustering, which assigns membership to rows only, simultaneous clustering assigns membership to both rows and columns, producing a reordered data matrix that we term a cluster solution. Two methods—spectral coclustering and spectral biclustering—were applied to a dataset of 87 dementia cases (rows), 15 features (columns), and six ground truth diagnoses. Spectral biclustering and coclustering were chosen for further study due to their inclusion of all rows and columns while forming simultaneous partitions and the observed effectiveness compared to other considered methods during the initial study. The number of clusters (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k$$\end{document}) was evaluated from 2 to 10 using cluster validity indices, which consistently indicated \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k=5$$\end{document} (Fig. 1). This choice produced strong validity scores, minimized small clusters, and distributed diagnoses across groups in biologically plausible ways (Fig. 2). Both methods identified five clusters that corresponded to distinct dementia phenotypes with characteristic symptom profiles (Table 1).Fig. 2Cluster composition. Spectral coclustering and spectral biclustering distributed the ground truth labels (diseases) across row clusters in a biologically plausible way, although the cluster configuration differed by method. Each method produced clusters that were enriched for certain dementia diagnosesTable 1Enriched features and diagnoses for row coclusters (Co-0–Co-4) and row biclusters (Bi-0–Bi-4)Spectral coclusteringRow coclusterEnriched featuresEnriched diseasesCo-0EOMS, cranial nerve, hyperkinesia, visionCJD, HDCo-1gait, hyperreflexia, hypertonia, hypokinesiaNPH, LBD, HDCo-2behavior, cognitive, speechFTD, ALZCo-3ataxia, tremor, sleepLBDCo-4sphincterNPHSpectral biclusteringRow biclusterEnriched featuresEnriched diseasesBi-0hyperkinesia, visionHD, CJDBi-1cognitive, speech, behaviorFTD, ALZ, HDBi-2hyperkinesia, cognitive, behaviorHD, CJD, ALZBi-3hypertonia, sleep, hypokinesiaLBDBi-4gait, sphincterNPHBiclustering yields independent row/column partitions; coclustering aligns row/column IDs. The coclustering and biclustering identified row clusters that are mixtures of diagnoses as shown in the Fig.2). Note similarities between Bi-4 and Co-4, between Co-3 and Bi-3, and between Co-0 and Bi-0 Cluster composition. Spectral coclustering and spectral biclustering distributed the ground truth labels (diseases) across row clusters in a biologically plausible way, although the cluster configuration differed by method. Each method produced clusters that were enriched for certain dementia diagnoses Table 1Enriched features and diagnoses for row coclusters (Co-0–Co-4) and row biclusters (Bi-0–Bi-4)Spectral coclusteringRow coclusterEnriched featuresEnriched diseasesCo-0EOMS, cranial nerve, hyperkinesia, visionCJD, HDCo-1gait, hyperreflexia, hypertonia, hypokinesiaNPH, LBD, HDCo-2behavior, cognitive, speechFTD, ALZCo-3ataxia, tremor, sleepLBDCo-4sphincterNPHSpectral biclusteringRow biclusterEnriched featuresEnriched diseasesBi-0hyperkinesia, visionHD, CJDBi-1cognitive, speech, behaviorFTD, ALZ, HDBi-2hyperkinesia, cognitive, behaviorHD, CJD, ALZBi-3hypertonia, sleep, hypokinesiaLBDBi-4gait, sphincterNPHBiclustering yields independent row/column partitions; coclustering aligns row/column IDs. The coclustering and biclustering identified row clusters that are mixtures of diagnoses as shown in the Fig.2). Note similarities between Bi-4 and Co-4, between Co-3 and Bi-3, and between Co-0 and Bi-0 Enriched features and diagnoses for row coclusters (Co-0–Co-4) and row biclusters (Bi-0–Bi-4) Biclustering yields independent row/column partitions; coclustering aligns row/column IDs. The coclustering and biclustering identified row clusters that are mixtures of diagnoses as shown in the Fig.2). Note similarities between Bi-4 and Co-4, between Co-3 and Bi-3, and between Co-0 and Bi-0 A juxtaposition of the data matrix before and after applying spectral methods, sorted according to the cluster memberships, was visualized using heatmaps to elucidate the identified clusters (Fig. 3). Further, cluster solutions were visualized with t-SNE scatter plots (Fig. 4). A perplexity of 22 gave stable separation of clusters, and both methods produced five clusters with good internal cohesion. The key differences resided in their differing emphasis on cognitive versus motor phenotypes. Spectral biclustering generated a large cognitive cluster (Bi-1) that combined Alzheimer’s disease (ALZ) and frontotemporal dementia (FTD), also drawing in HD and CJD patients with prominent dementia features. Spectral coclustering emphasized motor symptoms, producing a large movement cluster (Co-0) that grouped HD and CJD, with motor-dominant ALZ and FTD cases reassigned there. Both methods consistently separated Lewy body dementia (LBD) and normal pressure hydrocephalus (NPH) into distinct clusters.Fig. 3Heat maps of the dementia dataset before and after applying spectral methods. For both spectral methods, rows and columns have been reordered by cluster membership. Spectral coclustering produced the characteristic diagonal pattern running from the lower left to the upper right of the heat map, whereas spectral biclustering yielded the checkerboard pattern typical of independent row–column partitionsFig. 4t-SNE plots for dementia dataset. (a): t-SNE plot of dementia patients colored by ground truth labels (diagnoses). (b): t-SNE plot of dementia patients colored by spectral cocluster membership. (c): t-SNE plot of dementia patients colored by spectral bicluster membership. (d): centroids (mean x,y coordinates) for diagnoses (green circles), spectral biclusters (blue diamonds), and spectral coclusters (red squares) Heat maps of the dementia dataset before and after applying spectral methods. For both spectral methods, rows and columns have been reordered by cluster membership. Spectral coclustering produced the characteristic diagonal pattern running from the lower left to the upper right of the heat map, whereas spectral biclustering yielded the checkerboard pattern typical of independent row–column partitions t-SNE plots for dementia dataset. (a): t-SNE plot of dementia patients colored by ground truth labels (diagnoses). (b): t-SNE plot of dementia patients colored by spectral cocluster membership. (c): t-SNE plot of dementia patients colored by spectral bicluster membership. (d): centroids (mean x,y coordinates) for diagnoses (green circles), spectral biclusters (blue diamonds), and spectral coclusters (red squares) Because t-SNE plots are crowded, we overlaid centroids of ground-truth diagnoses and algorithmic clusters (Fig. 4d). ALZ and FTD centroids lay close together, as did HD and CJD, while LBD and NPH remained distinct. Coclustering positioned Co-0 near the HD–CJD pair, whereas biclustering placed Bi-1 closer to the ALZ–FTD pair. These differences illustrate how the two methods provide complementary perspectives: one emphasizing cognitive symptoms, the other motor symptoms. While t-SNE plots illustrate cluster configuration, they do not describe cluster composition. To address this, the FES incorporated heat maps (Fig. 5) showing feature means and diagnosis counts by cluster. For example, Co-0 showed enrichment of FTD and ALZ diagnoses along with behavior, cognitive, and speech features. Additional word clouds and SHAP plots (Figs 6, 7) highlighted enriched and influential features for row coclusters and row biclusters. Finally, detailed algorithmic descriptions are provided in the Supplementary Materials.Fig. 5(a) Ground truth label counts by cocluster (b) Feature centroids by cocluster (c) Ground truth label counts by bicluster (d) Feature centroids by biclusterFig. 6Explanations for cluster from spectral coclustering. The first column is a word cloud of enriched features in the cluster. The second column is a word cloud of enriched ground truth labels (diagnoses) in the cocluster. The third column is a SHAP bar chart of features that predict cluster membershipFig. 7Explanations for clusters from spectral biclustering. Each row explains a bicluster. The first column is a word cloud of enriched features in the cluster. The second column is a word cloud of enriched labels (diagnoses) for each cluster. The third column is a SHAP bar plot of predictive features cluster membership. Note that enriched features (column 1) may be different than predictive features (column 3) (a) Ground truth label counts by cocluster (b) Feature centroids by cocluster (c) Ground truth label counts by bicluster (d) Feature centroids by bicluster Explanations for cluster from spectral coclustering. The first column is a word cloud of enriched features in the cluster. The second column is a word cloud of enriched ground truth labels (diagnoses) in the cocluster. The third column is a SHAP bar chart of features that predict cluster membership Explanations for clusters from spectral biclustering. Each row explains a bicluster. The first column is a word cloud of enriched features in the cluster. The second column is a word cloud of enriched labels (diagnoses) for each cluster. The third column is a SHAP bar plot of predictive features cluster membership. Note that enriched features (column 1) may be different than predictive features (column 3) Discussion The application of a Formal Explanation Space (FES) to the simultaneous clustering of dementia phenotypes provided a structured framework that addressed two major challenges: the interpretability of cluster outputs and the presence of multiple competing, yet plausible, solutions. By visualizing and comparing results from spectral coclustering and spectral biclustering, FES yielded new insights into both the algorithms and the data. Several takeaways emerged:Cluster validity indices may not always definitively determine the optimal number of clusters, but they provide useful guidance. In our analysis, they consistently suggested \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k=5$$\end{document} as the preferred number of clusters (Fig. 1).When ground truth labels are available, simple bar charts are effective for checking cluster composition (Fig. 2).Heat maps capture the structure of the dataset, with rows as cases and columns as features (Fig. 3a). After clustering, they reveal how rows and columns are reordered into clusters (Figs 3b and 3c).Dimension reduction methods such as t-SNE provide a cluster-independent view of the data. When labeled by cluster, t-SNE plots highlight cohesion and separation and allow side-by-side comparison of cluster solutions in the same embedded space (Fig. 4b and c).Replacing individual cases on t-SNE plots with cluster centroids further clarifies proximity and relationships across solutions (compare Fig. 4d). As shown, when ground truth labels are available, cluster centroids can be compared to label centroids on the t-SNE plots.Heat maps also visualize composition by labels (diagnoses) (Figs 5a and 5c) and column labels (features) (Figs 5b and 5d), illustrating the dual nature of simultaneous clustering.Word clouds provide an alternative visualization of cluster composition (Figs 6 and 7).Feature importance can be assessed with SHAP plots (Figs 6 and 7). Unlike enrichment analysis, which highlights features overrepresented within a cluster, SHAP identifies the features most influential in separating one cluster from the others. These discriminative features may overlap with, but are not always identical to, the most enriched features. Cluster validity indices may not always definitively determine the optimal number of clusters, but they provide useful guidance. In our analysis, they consistently suggested \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k=5$$\end{document} as the preferred number of clusters (Fig. 1). Cluster validity indices may not always definitively determine the optimal number of clusters, but they provide useful guidance. In our analysis, they consistently suggested \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k=5$$\end{document} as the preferred number of clusters (Fig. 1). When ground truth labels are available, simple bar charts are effective for checking cluster composition (Fig. 2). When ground truth labels are available, simple bar charts are effective for checking cluster composition (Fig. 2). Heat maps capture the structure of the dataset, with rows as cases and columns as features (Fig. 3a). After clustering, they reveal how rows and columns are reordered into clusters (Figs 3b and 3c). Heat maps capture the structure of the dataset, with rows as cases and columns as features (Fig. 3a). After clustering, they reveal how rows and columns are reordered into clusters (Figs 3b and 3c). Dimension reduction methods such as t-SNE provide a cluster-independent view of the data. When labeled by cluster, t-SNE plots highlight cohesion and separation and allow side-by-side comparison of cluster solutions in the same embedded space (Fig. 4b and c). Dimension reduction methods such as t-SNE provide a cluster-independent view of the data. When labeled by cluster, t-SNE plots highlight cohesion and separation and allow side-by-side comparison of cluster solutions in the same embedded space (Fig. 4b and c). Replacing individual cases on t-SNE plots with cluster centroids further clarifies proximity and relationships across solutions (compare Fig. 4d). As shown, when ground truth labels are available, cluster centroids can be compared to label centroids on the t-SNE plots. Replacing individual cases on t-SNE plots with cluster centroids further clarifies proximity and relationships across solutions (compare Fig. 4d). As shown, when ground truth labels are available, cluster centroids can be compared to label centroids on the t-SNE plots. Heat maps also visualize composition by labels (diagnoses) (Figs 5a and 5c) and column labels (features) (Figs 5b and 5d), illustrating the dual nature of simultaneous clustering. Heat maps also visualize composition by labels (diagnoses) (Figs 5a and 5c) and column labels (features) (Figs 5b and 5d), illustrating the dual nature of simultaneous clustering. Word clouds provide an alternative visualization of cluster composition (Figs 6 and 7). Word clouds provide an alternative visualization of cluster composition (Figs 6 and 7). Feature importance can be assessed with SHAP plots (Figs 6 and 7). Unlike enrichment analysis, which highlights features overrepresented within a cluster, SHAP identifies the features most influential in separating one cluster from the others. These discriminative features may overlap with, but are not always identical to, the most enriched features. Feature importance can be assessed with SHAP plots (Figs 6 and 7). Unlike enrichment analysis, which highlights features overrepresented within a cluster, SHAP identifies the features most influential in separating one cluster from the others. These discriminative features may overlap with, but are not always identical to, the most enriched features. Increased interpretability through FES A central contribution of this work is the development of FES to enhance interpretability. Simultaneously clustering rows (patients) and columns (signs and symptoms) produces complex biclusters that are difficult to interpret without structured tools. FES integrates t-SNE, heat maps, word clouds, and SHAP plots, enabling interpretable representations of cluster structure, quality, and feature influence. This allowed us to translate cluster outputs into biologically meaningful insights. The biological plausibility of the cluster solutions We define biological plausibility as alignment between algorithmic clusters and established neurological knowledge [45, 48–50]. While the concept remains underexplored in machine learning evaluation [51], it provides a clinically grounded lens for interpreting clustering results. Both spectral coclustering and spectral biclustering produced solutions that domain experts judged biologically plausible (Table 1). When the dementia patients were simultaneously clustered in five subgroups (k = 5) several biologically plausible associations and observations emerged:HD and CJD motor–visual clusters (Bi-0, Co-0, and Bi-2). Both spectral methods produced clusters enriched with hyperkinetic and visual symptoms, including abnormal eye movements (EOMS), visual deficits, and cranial nerve abnormalities. These features co-occurred with Huntington’s disease (HD) and Creutzfeld–Jakob disease (CJD), consistent with their shared presentation of movement disorders and oculomotor abnormalities. The t-SNE plots reinforced this relationship, placing HD and CJD close to each other and near the centroids of Bi-0 and Co-0 (Fig. 4d). Notably, Bi-2 deserves special mention: although it resembled Bi-0 in being enriched for HD and CJD, it also showed more cognitive impairment. This suggests underlying heterogeneity within HD and CJD cases, aligning with clinical knowledge that both disorders can manifest with variable contributions of motor and cognitive symptoms.AD and FTD cognitive-speech-behavior clusters (Co-2 and Bi-1). Both spectral methods showed clusters enriched for cognitive, speech, and behavioral abnormalities and were also enriched with Alzheimer’s disease (AD) and frontotemporal dementia (FTD) diagnoses. This reflects the well-known overlap of signs and symptoms between these two diseases (Table 1). This proximity was confirmed by the t-SNE plot showing proximity of the centroids for FTD and AD (Fig. 4 d).LBD-hypertonia, sleep, hypokinesia tremor clusters (Bi-3 and Co-3). Lewy body dementia (LBD) consistently formed its own cluster, driven by hypokinesia, hypertonia, and sleep disturbance. These features align closely with the hypokinetic and sleep hallmarks of the disorder. The grouping of hypokinesia, sleep disorders, hypertonia, and tremor was biologically plausible and reflected known associations between these signs and symptoms.NPH-sphincter-gait clusters (Bi-4 and Co-4). Both spectral methods produced a cluster enriched for normal pressure hydrocephalus (NPH) and characterized by sphincter and gait disturbances, consistent with the known characteristics of this disorder. In the t-SNE space, NPH emerged as a unique entity isolated from the other dementia diagnoses (Fig. 4 d). HD and CJD motor–visual clusters (Bi-0, Co-0, and Bi-2). Both spectral methods produced clusters enriched with hyperkinetic and visual symptoms, including abnormal eye movements (EOMS), visual deficits, and cranial nerve abnormalities. These features co-occurred with Huntington’s disease (HD) and Creutzfeld–Jakob disease (CJD), consistent with their shared presentation of movement disorders and oculomotor abnormalities. The t-SNE plots reinforced this relationship, placing HD and CJD close to each other and near the centroids of Bi-0 and Co-0 (Fig. 4d). Notably, Bi-2 deserves special mention: although it resembled Bi-0 in being enriched for HD and CJD, it also showed more cognitive impairment. This suggests underlying heterogeneity within HD and CJD cases, aligning with clinical knowledge that both disorders can manifest with variable contributions of motor and cognitive symptoms. HD and CJD motor–visual clusters (Bi-0, Co-0, and Bi-2). Both spectral methods produced clusters enriched with hyperkinetic and visual symptoms, including abnormal eye movements (EOMS), visual deficits, and cranial nerve abnormalities. These features co-occurred with Huntington’s disease (HD) and Creutzfeld–Jakob disease (CJD), consistent with their shared presentation of movement disorders and oculomotor abnormalities. The t-SNE plots reinforced this relationship, placing HD and CJD close to each other and near the centroids of Bi-0 and Co-0 (Fig. 4d). Notably, Bi-2 deserves special mention: although it resembled Bi-0 in being enriched for HD and CJD, it also showed more cognitive impairment. This suggests underlying heterogeneity within HD and CJD cases, aligning with clinical knowledge that both disorders can manifest with variable contributions of motor and cognitive symptoms. AD and FTD cognitive-speech-behavior clusters (Co-2 and Bi-1). Both spectral methods showed clusters enriched for cognitive, speech, and behavioral abnormalities and were also enriched with Alzheimer’s disease (AD) and frontotemporal dementia (FTD) diagnoses. This reflects the well-known overlap of signs and symptoms between these two diseases (Table 1). This proximity was confirmed by the t-SNE plot showing proximity of the centroids for FTD and AD (Fig. 4 d). AD and FTD cognitive-speech-behavior clusters (Co-2 and Bi-1). Both spectral methods showed clusters enriched for cognitive, speech, and behavioral abnormalities and were also enriched with Alzheimer’s disease (AD) and frontotemporal dementia (FTD) diagnoses. This reflects the well-known overlap of signs and symptoms between these two diseases (Table 1). This proximity was confirmed by the t-SNE plot showing proximity of the centroids for FTD and AD (Fig. 4 d). LBD-hypertonia, sleep, hypokinesia tremor clusters (Bi-3 and Co-3). Lewy body dementia (LBD) consistently formed its own cluster, driven by hypokinesia, hypertonia, and sleep disturbance. These features align closely with the hypokinetic and sleep hallmarks of the disorder. The grouping of hypokinesia, sleep disorders, hypertonia, and tremor was biologically plausible and reflected known associations between these signs and symptoms. LBD-hypertonia, sleep, hypokinesia tremor clusters (Bi-3 and Co-3). Lewy body dementia (LBD) consistently formed its own cluster, driven by hypokinesia, hypertonia, and sleep disturbance. These features align closely with the hypokinetic and sleep hallmarks of the disorder. The grouping of hypokinesia, sleep disorders, hypertonia, and tremor was biologically plausible and reflected known associations between these signs and symptoms. NPH-sphincter-gait clusters (Bi-4 and Co-4). Both spectral methods produced a cluster enriched for normal pressure hydrocephalus (NPH) and characterized by sphincter and gait disturbances, consistent with the known characteristics of this disorder. In the t-SNE space, NPH emerged as a unique entity isolated from the other dementia diagnoses (Fig. 4 d). NPH-sphincter-gait clusters (Bi-4 and Co-4). Both spectral methods produced a cluster enriched for normal pressure hydrocephalus (NPH) and characterized by sphincter and gait disturbances, consistent with the known characteristics of this disorder. In the t-SNE space, NPH emerged as a unique entity isolated from the other dementia diagnoses (Fig. 4 d). Taken together, these findings directly answer the questions posed at the outset. We identified biologically consistent disease clusters (AD with FTD; HD with CJD; LBD and NPH remaining distinct) as well as symptom clusters (cognitive–speech–behavior; motor–visual; hypokinesia–sleep–tremor; gait–sphincter). Both the disease clusters and the symptom clusters aligned with established neurological knowledge, strengthening confidence in the biological plausibility of our solutions. Limitations and future work Although FES improved interpretability, several limitations remain. First, our analysis focused on spectral methods; applying additional algorithms would test generalizability. Second, the dataset was small, and scalability to larger cohorts remains to be evaluated. Third, our use of ground truth labels aided interpretation, but such labels are often unavailable in real-world data. Fourth, cocluster interpretation is more challenging when algorithms cluster only subsets of rows or columns, a scenario not addressed here. Finally, extending FES to other biomedical domains will be essential to establish its broader utility. Conclusion The Formal Explanation Space (FES) provides a structured framework for interpreting clustering results. By enabling comparison of multiple solutions, FES revealed biologically plausible and clinically meaningful dementia subgroups. More broadly, FES demonstrates how explanation frameworks can help domain experts engage with unsupervised machine learning methods that often produce divergent but valid solutions. For complex biomedical data, pairing clustering with an explanation space can substantially enhance interpretability, support critical checks of biological plausibility, and open new avenues for insight across diverse datasets. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1