A scoping review of the clinical application of machine learning in data-driven population segmentation analysis

Study design

The scoping review was conducted according to PRISMA for Scoping Reviews guidelines (PRISMA-ScR).¹² Scoping studies, a popular research review method, involve mapping evidence to summarize a field’s breadth and depth, clarifying complex concepts, and refining future research. They do not assess study quality and require analytical interpretation compared with systematic review and narrative reviews separately. Relevant to emerging evidence, they could include published and gray literature, complementing clinical trial results.¹³

Search strategies

We used literature databases to obtain a comprehensive list of possibly relevant papers for this scoping review (Figure 1). From January 1, 2000, through October 1, 2022, the MEDLINE, Embase, Web of Science, and Scopus databases were searched. The publication year range was specified to not only study the use and development of ML in population segmentation since 2000 when unsupervised ML methods become widespread but also supplement the existing reviews when ML became popular in recent years. Our search strategy included the key search terms stratif* OR segment* OR categor* OR “cluster analysis” OR cluster* OR pattern* OR profil* OR phenotyp* OR partition* OR subgroups OR subtypes OR subpopulations. The search strategy for each database is presented in the Supplementary Material. We also conducted the ancestral search.

Eligibility criteria

Full-text, original, peer-reviewed, English-language papers that employed data-driven population segmentation analysis on structured data were included. The segmentation methods had to be applied to real-world datasets, not theoretical, hypothetical, or simulated ones. The meta-analysis, case series, case reports, conference papers, gray literature, and reviews were excluded. Articles that were not on human subjects, not in English, not applying ML methods (statistical methods such as latent class analysis would not be included), not using unsupervised segmentation methods (using unlabeled data as inputs), not applying in clinical settings, and only relying on experts’ opinions were also excluded. Clustering not on population, for example, on the country as a unit of analysis levels, was excluded as well.

Selection of studies

EndNote, version X9 was used to remove duplicate publications. Two independent researchers (PL and ZW) reviewed the abstracts of retrieved papers for inclusion and discussed any discrepancies. They then independently examined and assessed the full-text papers for eligibility.

As illustrated in Figure 1, there were 3398 articles found in databases searching after 2679 duplicates were removed. After screening the abstracts, 3233 papers that did not meet the inclusion criteria were eliminated, leaving 165 articles. For the 165 recognized articles, 79 papers were ultimately included for full-text review. The percentage agreement between PL and ZW achieved 90%. Disputes in the selection procedure were resolved by discussion and consensus with 2 other researchers (MAP and NL).

Data extraction

Once an article was determined to be included, 2 researchers (PL and ZW) independently extracted data from eligible papers. In addition to the title and year of publication, we mainly extracted information from selected publications from 3 perspectives: populations studied (including target population, population size, country of the study, health conditions, data sources, data type, and study settings), segmentation (including segmentation objectives, number of variables, type of variables used, statistical methods, newly designed/existed, number of segments, and names of segments), and validation (comparison with other methods, criteria to determine the number of segments, internal validation, external validation, interpretability, and methods for missingness).

Data synthesis

In the included publications, we summarized target group characteristics, such as inclusion criteria, sample size, and geographic region, illustrating the populations subjected to data-driven segmentation analysis. We detailed data sources (primary or secondary) and contexts (community-based or healthcare institutions) for future research. Moreover, we established segmentation objective topics, assessed articles' utilization of segments, and highlighted limitations of statistical methods to guide method selection. We examined method comparisons for conclusion validity and proposed pilot plans for consistency evaluation across different approaches.

Population studied

As exhibited in Table 1, the population can be broadly classified as either the general population or the disease/condition-specific population. The included studies took a sample of the population. The segmentation targets of most articles (59/79; 74.7%) were the population with specific diseases/conditions. Individuals with cardiometabolic (7/79; 8.9%), diabetes (8/79; 10.1%), epilepsy (2/79; 2.5%), Covid-19 (4/79; 5.1%), ICU admissions (2/79; 2.5%), emergency (3/79; 3.8%), oral health (2/79; 2.5%), nervous system diseases (3/79; 3.8%), mental health (8/79; 10.1%), hospitalization (3/79; 3.8%), respiratory diseases (2/79; 2.5%), infectious diseases (3/79; 3.8%), multimorbidity (4/79; 5.1%), health-services usage (2/79; 2.5%), pregnancy (2/79; 2.5%), and other conditions (4/79; 5.1%) were included in these articles. Other papers (20/79; 25.3%) targeted the general population.¹⁴ However, in this review, there are still some unincluded health conditions where segmentation analysis has been merely not applied, such as unintentional injuries, stroke, kidney, and ophthalmology, which could be the future direction to improve the delivery of health plans using segmentation methods.

For the sample size in 85 studies (see Figure 2), the smallest number is 51 by Galvez-Goicurla et al, who classified a population of migraine patients into 4 clusters of pain types with distinct morphological features using a Logistic Model Tree algorithm.¹⁵ The most extensive study involved 1 608 741 patients from a diverse hospital cohort.¹⁶ As for study locations (see Figure 2), the vast majority were carried out in North America and Europe, with the United States (40/85; 47.1%) having the most studies.

Segmentation variables, data sources, and settings

Depending on the purpose of segmentation, variables used for segmentation analysis in these studies were heterogeneous, including sociodemographic, vital signs, pharmacy, care utilization, psychosocial factors, behavioral factors, and lab tests. Supplementary Table S3 summarized the details of segmentation variables. Based on the study design (cross-sectional or longitudinal), the clustering could be longitudinal or static. For example, Wong et al used trajectories of hospitalization probability consisting of demographics, clinical, vital signs, utilization, pharmacy, and lab values to identify latent subgroups of high-risk patients for hospitalization.⁶⁴ In addition to EHR, Benis et al also analyzed documented access to various communication channels to determine which channels of communication, within the health organization, were most preferred by the patients and could be used to deliver patient-specific messages.⁸⁵ Sociodemographic data were nearly all utilized to identify the disease’s clinical subtypes.^53,64,71,86 Diagnosis codes and procedural patterns were also used to assess patient representations and subtypes.^21,23,87

As shown in Figure 3, most studies (69/85; 81.2%) utilized secondary data from earlier studies or third-party sources, such as the National Health and Nutrition Examination Survey, which was designed to evaluate the general health and nutritional status of the US people and various demographic groups, including medical, dental, and physiological measurements, as well as laboratory tests administered by highly trained medical personnel. Besides, Bej et al used the National Family Health Survey-4 dataset from India to examine the type-2 diabetes subpopulation.⁸⁸ Conversely, only 14 studies (14/85; 16.5%) employed primary data for population segmentation. For instance, researchers obtained real-time pain evolution data through a mobile application to analyze pain episodes in chronic diseases.¹⁵

This review classified the study settings into 2 categories: healthcare institutions and communities. The majority (55/85; 64.7%) of the study settings were healthcare institutions (see Figure 3). To illustrate, Rodríguez et al analyzed data from adult patients with confirmed SARS-CoV-2 infection admitted in 63 ICUs across Spain.⁵⁴ Some studies were conducted in communities. For example, Hu used data from the Chinese Longitudinal Health Longevity Survey, which collected gathered data on aging and caregiving from a representative sample of the oldest old in 23 Chinese provinces.¹⁸

Objectives of segmentation

Included studies consistently identified 3 primary population segmentation goals: (1) clinical prediction, (2) health grouping/profiling, and (3) delivery of healthcare interventions. There is some overlap, and they are not mutually exclusive themes. All included articles (n = 79) centered on Health Grouping/Profiling. Furthermore, most (n = 52) aimed to deliver healthcare interventions specific to each demographic group. Twenty-five studies aimed to conduct clinical prediction to identify related risk factors. In conclusion, the objective is to categorize individuals by care type and frequency, enabling tailored clinical care, health interventions, and policies based on unique health factors, rather than predicting risk scores for specific outcomes like hospitalization or monthly costs, according to the Johns Hopkins Adjusted Clinical Groups system.⁸⁹

Segmentation methods

Based on the goal of population segmentation, we only focused on unsupervised classifications which are used to analyze and cluster unlabeled datasets. Figure 4 shows that among 117 methods, k-means cluster analysis (47/117; 40.2%) was the most widely used unsupervised method. For example, Flores et al applied the k-means clustering algorithm for the automated detection of coronary artery disease subgroups.³⁴ Some clustering analyses were also performed using k-medoids, a variant of k-means that works well with ordinal scale variables.^23,34,37,57 Some segmentation analyses used a mix of techniques, as they were not mutually exclusive. To illustrate, Benis et al used k-means clustering for discretization and patient profile construction and finally employed hierarchical clustering to show the various communication profiles with the health care organization.⁸⁵ In order to cope with more sophisticated situations, deep learning was also applied in classification. Thomas et al proposed a new real-time tool for visualizing these high-dimensional data to expose previously undisclosed relationships and clusters with deep autoencoders.⁷³

Other clustering algorithms included deep autoencoder,⁷³ clustering large applications,⁶¹ the mixture of multivariate generalized linear mixed models,⁸² uniform and manifold approximation and projection (UMAP),⁶⁷ logistic model tree,¹⁵ latent trajectories analysis,¹⁸ Gaussian mixture models,⁶³ 1-layer convolutional neural networks,¹⁶ hypergraph network analysis with eigenvector centralities,⁷⁵ composite mixture model (CMM),⁶⁵ and collaborative modeling.⁴⁶ Dimension reduction methods were sometimes used together with clustering. For example, self-organizing mapping (SOM) was used to reduce a high-dimensional dataset to a lower-dimensional one while maintaining its topological structure. Dipnall et al considered SOM after addressing the imbalanced nature of the data, which is the primary rationale for undertaking SOM.⁴¹ UMAP was another method found in this review for nonlinear dimensionality reduction. Hurley et al applied UMAP first for dimensionality reduction, which can capture both local and global structures in high-dimensional data, and then used Gaussian mixture models for clustering data.⁶³

Segmentation outcome

Each study’s segmentation outcome was evaluated and summarized in Supplementary Table S4, considering the following items: internal validity (the degree to which the clustering solution accurately reflects the underlying structure of the data,⁹⁰ through measures such as cluster stability), external validity (the extent to which the results can be generalized to other populations or contexts, by comparing the results to existing theory or assessing consistency across different datasets), and interpretability (ensuring segments are easily recognized and understood³). Figure 5 summarizes the number of articles on each evaluation criterion. All papers (n = 79) performed internal validation for the segmentation analysis. For example, Hurley et al employed the adjusted Rand index to establish stability in detecting these clusters found by Gaussian mixture models.⁶³ Nevertheless, only a few papers incorporated external validation (11/79; 13.9%). Faghri et al used a replication cohort to validate the clusters externally.⁶⁷ In addition, all papers (n = 79) interpreted the segmentation outcomes based on related analysis (eg, regression or Student’s t tests) or clinical experts’ viewpoints. For the number of segments (see Figure 4), we only counted and presented the one with the best performance result in each study. Most studies (79/85; 92.9%) produced less than 10 segments.²² Various methods were employed to decide the number of segments, including the Calinski-Harabasz criterion, elbow method, C index, Silhouette coefficient, clustering cohesion index, dendrogram, Akaike information criterion, Bayesian information criterion, cross-validation analysis, the Dunn index, stability index, Jaccard score, Jump method, Davis-Bouldin index, Wald’s minimum variance method, and expert viewpoints.

Methods for missing values

A total of 48 of 79 articles (48/79; 60.1%) did not detail how to deal with missing values. Moreover, 14 articles (14/79; 17.7%) excluded all missingness and used complete cases for analysis. The rest of those papers used imputation mainly for missing values, including imputation by filling the subgroup-level mean values,⁶⁴ imputation with the regularized iterative principal component analysis (PCA) algorithm¹⁷ for variables with less than 33% missingness, k-nearest neighbor imputation,⁶⁷ generalized low-rank models,³⁴ imputation with mean values,^91,92 sensitive analysis to replace with positive values,⁷⁵ genotype imputation,⁴² CMM,⁶⁵ creating a separate variable indexing the count of missing variables out of the final variable set,⁶⁶ replacement with the median value,³⁵ and MissForest imputation method.²³ Based on previous illustrated methods and various unlisted ones, future studies should be cautious when dealing with missingness according to data types and structure to keep as much information as possible.

Proof of the validity of methods

Nearly all papers concluded that the ML algorithms they used performed well in segmentation and provided valuable value in personalized treatment plans. However, less than 30% compared different methods. In order to discover underlying patterns in early childhood dental care, Peng et al⁷⁶ compared k-means clustering with traditional segmentation analysis by calculating per member per year dental cost by gender. Results generated by k-means clustering performed better than random clustering and provided more details than traditional segmentation. Some studies compared several unsupervised methods. For example, Bose and Radhakrishnan³⁶ compared hierarchical clustering, k-means clustering, and partitioning around medoids to identify subgroups among home health patients with heart failure.

Since unsupervised learning worked on the unlabeled dataset, this would make the results less accurate. Therefore, some studies also incorporated supervised or semisupervised learning and compared the results. Faghri et al applied unsupervised UMAP modeling and semisupervised (neural network UMAP) modeling to identify amyotrophic lateral sclerosis clinical subtypes.⁶⁷ Although both modelings defined the same subgroups, they found that semisupervised ML produced the optimum results based on visual inspection. However, not all studies found that supervised methods outperformed unsupervised methods. Goodman et al used both unsupervised and supervised clustering algorithms to identify emergency department frequent user subgroups; however, they found that supervised algorithms could not group patients into clusters, while unsupervised modeling was able to form distinct clusters.⁶²

Future studies should provide more evidence to show the strengths of applying ML to segmentation, especially compared with traditional segmentation analysis or some golden standards. Additionally, a lack of rigorous evaluation of segmentation methods was observed, with most studies merely employing standard metrics to assess internal validity. In light of this, we emphasize the necessity for further examination and exploration of segmentation models, encompassing both external and internal validation.

Limitations of methods

In the reviewed studies, researchers selected the most suitable segmentation methods based on data characteristics and variables, but no method is perfect. About half of these studies addressed the limitations of their chosen methods, including potential informative missingness from the k-means clustering algorithm,⁶⁴ the conceptual costliness of the SOM ML algorithm with increasing variables and grid size, and the unsuitability of SOM for diagnostic modeling.²² Furthermore, incomplete data and outliers posed challenges for SOM maps,⁴¹ while fuzzy c-means⁵⁰ and k-median⁷¹ clustering were found to be less susceptible to outliers.

The temporal stability of identified clusters emerged as another concern, particularly in studies using unsupervised ML approaches like the mixture of multivariate generalized linear mixed models.⁸² Future research could involve phenotyping investigations that assess chronotype stability associated with each phenotype.

Lastly, some newly designed methods, such as the CMM proposed by Mayhew et al, faced issues like computational intensity and the need to manually specify model templates or univariate distribution lists in high-dimensional feature spaces.⁶⁵ Addressing these limitations and refining existing algorithms remain trending subjects in ongoing research.

Postsegmentation analysis

Only one-third of the papers in this review conducted postsegmentation analysis. The rest of them only discussed the profiles/characterization of segments and made suggestions for the individualized delivery of interventions but without real-life clinical applications. This is another serious issue to be considered for future research that how to make valuable use of those segments.

For those with postsegmentation analysis articles, most of them aimed to compare characteristics between clusters and characterize subgroups. Dipnall et al⁴¹ applied Multiple Additive Regression Tree boosted ML algorithm to identify the important medical symptoms for each key cluster and used binary logistic regression to identify the key clusters with higher levels of depression. Another common way is to use those clusters as labels for the later prediction by applying supervised ML. Fighri et al⁶⁷ used supervised ML to build predictor models to classify individual patients. Segmentation results were also used to develop a web-based tool for clinical use.⁶⁷ Landi et al¹⁶ evaluated disease subtypes as labels for training supervised models that can predict stratified patient risk scores. In addition, segments were also able to form patients’ migration patterns across 3 periods and classify disease trajectories.³⁵ We advocate for the dissemination of current methods and methodological advancements in “postsegmentation” inferential and predictive modeling research.

Sensitive analysis for segmentation

Only a few studies in this review conducted a sensitivity analysis for segmentation to make results more confident. For missing values, Larvin et al⁷⁵ examined the impact of absent survey responses and osteoporosis data by comparing the results with this cycle eliminated. For the quality, the silhouette score and the Calinski-Harabasz score were used to compare the clustering's quality to that of a random clustering.⁷⁶ Sangkaew et al⁵⁷ explored the effect that using the Gower distance in place of the Euclidean distance. In addition to this, Violán et al⁵⁰ test with different prevalence cut-off points for chronic diseases.

Future direction

As shown in this review, using different segmentation algorithms on the same population may result in different clusters. Although some papers have checked the similarity of clusters derived by different methods, only 4 papers have researched the consistency of belonged clusters for the same individual using various methods in our review. One study checked the consistency for each individual with original clusters by drawing scatterplots.⁶⁸ Landi et al¹⁶ compared type 2 diabetes and Parkinson’s disease subgroups’ results to related studies based on ad hoc cohorts. And Stafford et al⁴⁹ checked the coincidence with the literature. Besides, Vidal et al⁵⁸ conducted laboratory testing to check the concordance. In the future, the research could focus more on checking the consistency of resulting clusters for the same person using different methods, which is a major concern among clinicians, as different clusters would result in specific intervention plans or treatments. Another reason is that various clustering algorithms and threshold values could produce distinct clusters and migration patterns.³⁵ However, when an individual is classified into different clusters under different models, how to calibrate the clusters of such fuzzy individuals is also a meaningful issue that can be paid attention to in the future. It is essential to be more cautious when it comes to interpreting and naming the derived segments. Therefore, it is also necessary to well-establish criteria for optimal segmentation.

Strengths and limitations

To the best of the authors’ knowledge, this is the first study to conduct a scoping review of data-driven population segmentation analysis using ML. We provided a summary of the various segmentation methods that are commonly used, the evaluations of outcomes, methods for coping with missing values, and the various clinical settings to which the segmentation analysis was applied. It is also the first study to summarize the subsequent use of the segments, which can provide some insights to guide future research. Nonetheless, our study has limitations, as we only reviewed literature written in English, excluded clustering not based on population levels and unstructured data, disregarded gray literature, and omitted papers published before January 2000.