Cluster analysis of elderly emergency department patients in Taiwan: identifying clinical profiles and healthcare utilization patterns

Abstract

Background

To identify distinct clinical phenotypes among older adults presenting to the emergency department (ED) using an unsupervised machine learning approach and to characterize their clinical characteristics, healthcare utilization patterns, and discharge outcomes.

Methods

The retrospective study was based on secondary analysis of data from the “Program for the Taiwan Triage and Acuity Scale (TTAS) for Elderly in Emergency Department”. A total of 1,617 non-traumatic older patients (aged ≥ 65 years) presenting to the emergency department (ED) were included. K-prototypes cluster analysis, an unsupervised machine learning approach, was used to identify distinct clinical phenotypes.

Results

The K-prototypes clustering algorithm identified three distinct phenotypes: Cluster 1 (frail and highly comorbid; 31.6%), Cluster 2 (younger and functionally preserved; 32.2%), and Cluster 3 (metabolically burdened; 36.2%). Cluster 1 was characterized by the highest median ED costs (4,191.5 points) and the longest LOS (median 8 days). Although 77.1% of patients in Cluster 2 were discharged from the ED, those requiring hospitalization had the highest proportion of in-hospital death or against advice discharge (AAD) (17.6%). Cluster 3 demonstrated obesity and metabolic risks but maintained relatively intact functional status.

Conclusions

Machine learning clustering identified distinct clinical phenotypes among older adults presenting to the ED. These phenotypes may provide additional insights into the heterogeneity of this population and could help inform future phenotype-oriented risk stratification and individualized emergency care.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12877-026-08052-y.

Introduction

The global population is undergoing an unprecedented demographic transition toward an older age structure, presenting substantial challenges to healthcare systems worldwide. In Taiwan, this demographic shift is occurring at a remarkably rapid pace; the region is projected to officially transition into a “super-aged” society by 2030, a threshold at which older adults (aged ≥ 65 years) comprise over 24% of the total population [1]. Unsurprisingly, this aging demographic accounts for a disproportionate and accelerating volume of ED visits [2]. Compared with younger cohorts, elderly patients presenting to the ED exhibit considerable clinical complexity. They frequently present with significant multimorbidity, extensive polypharmacy, baseline cognitive impairment, and varying degrees of diminished physiological reserve, commonly recognized as frailty [3–5] Furthermore, acute illnesses in older adults often manifest with atypical or nonspecific presentations—such as generalized weakness, delirium, falls, or functional decline—rather than specific physiological derangements or classic localized symptoms [6].

This inherent clinical complexity complicates acute diagnostic and therapeutic workflows in the emergency setting. Traditional medical education often emphasizes disease-specific symptomology, which may be absent or obscured in frail older adults due to immunosenescence or the blunting effects of chronic medications [3]. Consequently, elderly patients generally require more extensive diagnostic testing, undergo longer ED lengths of stay, and possess a markedly higher likelihood of hospital admission, early readmission, and short-term mortality [3]. As EDs increasingly serve as a key point of access and diagnostic hub for acute geriatric care, the growing number of complex older patients contributes to ED crowding—a systemic crisis that compromises patient safety and severely strains limited healthcare resources [7, 8].

Given the substantial heterogeneity among elderly ED patients, accurate clinical risk stratification remains challenging. Traditional ED triage systems—such as the Emergency Severity Index (ESI) or the Taiwan Triage and Acuity Scale (TTAS)—were primarily developed and validated for the general adult population [9–11]. These systems rely largely on physiological parameters (e.g., vital signs) and presenting complaints to rapidly identify patients with immediate life-threatening conditions. However, they may not fully capture the multidimensional vulnerabilities inherent to older adults, such as underlying frailty, subtle metabolic disturbances, declining functional status, and complex psychosocial needs [12, 13]. As a result, some older patients may be assigned lower acuity levels despite having clinically significant conditions [14]. This potential under-recognition of illness severity may delay appropriate evaluation and management in certain cases [13].

To address the limitations of conventional triage approaches, medical researchers have increasingly turned to artificial intelligence and machine learning (ML) to augment clinical decision-making [15]. While supervised ML models have demonstrated success in predicting predefined outcomes, they inherently require labels and typically evaluate patients based on specific endpoints. This approach may not fully capture the underlying heterogeneity and phenotypic diversity of complex clinical populations. In contrast, unsupervised machine learning—particularly cluster analysis—provides a data-driven framework to identify latent structures and define clinically meaningful phenotypes within heterogeneous populations without prespecified outcomes [16, 17]. Identifying such phenotypic groupings in the geriatric ED population may offer insights into how different constellations of clinical features relate to patterns of healthcare utilization and outcomes. However, clustering clinical ED data presents methodological challenges, as these datasets commonly include both continuous variables (e.g., vital signs) and categorical variables (e.g., comorbidities and diagnoses). Traditional clustering techniques, such as the well-known K-means algorithm, rely exclusively on Euclidean distance metrics, which are mathematically invalid and conceptually meaningless for categorical variables [18].

To address this challenge the K-prototypes was applied, as it is specifically designed for clustering mixed-type data by integrating distance measures for both continuous and categorical variables [19]. This approach allows the incorporation of diverse clinical features while preserving the multidimensional structure of the data. Using this method, the present study aims to identify distinct clinical phenotypes within the elderly ED population and to explore their clinical characteristics and associated patterns of healthcare utilization and outcomes.

The primary objective of this study was to identify distinct clinical phenotypes among older adults presenting to the emergency department (ED) using an unsupervised machine learning approach. We further aimed to characterize the clinical and functional features of these phenotypes and describe their healthcare utilization patterns and short-term outcomes.

Materials and methods

Study design and setting

This study was a retrospective secondary analysis of prospectively collected utilized an unsupervised machine learning approach to characterize the clinical phenotypes and healthcare utilization patterns of older adults to the emergency department (ED). Data were retrieved from the “Program for the Taiwan Triage and Acuity Scale (TTAS) for Elderly in Emergency Department”, a research program conducted at Linkou Chang Gung Memorial Hospital (LCGMH), the largest single-site tertiary academic medical center in Taiwan, with approximately 3,700 beds and a high volume of acute ED visits. The study population included non-trauma adult patients aged 65 years and older who presented to the LCGMH Emergency Department (ED) between March 2024 and February 2025. The study protocol adhered to the ethical guidelines of the Declaration of Helsinki. It was reviewed and approved by the Institutional Review Board of the Chang Gung Medical Foundation (Approval No. 202400144B0D0001). To strictly maintain patient privacy and data confidentiality, all institutional identification numbers were cryptographically scrambled prior to database access. Due to the retrospective and anonymized nature of the dataset, the requirement for informed consent was formally waived by the approving institutional review board.

Selection of participants

Non-traumatic Patients aged ≥ 65 years presenting to the ED were eligible for inclusion. The patient selection and data allocation workflow is illustrated in Fig. 1. Cases with systematic missing data—defined as those completely lacking anthropometric measurements or vital sign records, or cases with an ED disposition indicating admission to the ward or intensive care unit (ICU) but without any corresponding hospitalization records—were excluded from the primary cluster analysis to ensure data integrity. Furthermore, records with formatting errors were excluded. The remaining eligible complete cases (n = 1519, 91.2% of the study cohort) were used to develop the primary clustering model. A small subset of cases exhibiting sporadic missing values (n = 94, with a low overall proportion of missingness, < 5% per variable) were regarded as missing at random (MAR) based on clinical documentation workflows and the absence of systemic missing patterns. These cases were excluded from the primary cluster derivation and were included only in a subsequent secondary exploratory analysis.

Fig. 1.

Flow chart of patient selection for machine learning cluster analysis

Data collection and variables

To ensure data accuracy and minimize the inherent biases of manual chart review, an independent clinical data analyst used a standardized, committee-authorized query to extract data from the hospital’s centralized electronic health records for older adult patients (aged ≥ 65 years) presenting to the emergency department (ED). A comprehensive set of variables was retrieved for the cluster analysis. Continuous variables included age, anthropometric measurements (e.g., body mass index, waist circumference, and hip circumference), initial triage vital signs (e.g., heart rate, blood pressure, and oxygen saturation), Glasgow Coma Scale (GCS), Edmonton Symptom Assessment Scale, Clinical Frailty Scale (CFS), length of hospital stay (days), and ED costs, defined as the National Health Insurance point value incurred during the emergency department visit only (excluding subsequent hospitalization costs) [4, 13, 20]. Upon arrival at the ED, the GCS, Edmonton Symptom Assessment Scale and CFS were evaluated and recorded by senior nurses who had undergone standardized geriatric assessment training into the dataset. Categorical variables encompassed sex, Eastern Cooperative Oncology Group (ECOG) performance status, triage level, longitudinal ED disposition, and final discharge status. Additionally, all underlying medical comorbidities (e.g., cardiovascular disease, diabetes mellitus, and cancer) and primary acute ED diagnoses (e.g., pneumonia, urinary tract infection, and heart failure) were systematically categorized using the International Classification of Diseases, Tenth Revision (ICD-10) coding framework. Discharge status was categorized as alive or unfavorable outcome. An unfavorable outcome comprised in-hospital death and discharge against advice (AAD, discharge in extremis), which refers to terminal discharge at the request of the patient or family before in-hospital death. Because AAD commonly represents end-of-life discharge in Taiwan, death and AAD were considered together as a composite unfavorable outcome in the cluster analysis.

Data preprocessing

Continuous variables were standardized prior to clustering, whereas categorical variables were included without additional preprocessing. Cases with systematic missing data, such as those lacking all anthropometric measurements or vital sign records, were excluded from the cluster analysis. Among the remaining cases, only sporadic missing values were observed, with a low overall proportion of missingness (< 5% per variable). These missing values were therefore considered MAR based on the clinical documentation workflow and the absence of systematic missing patterns. Clustering was performed using the complete-case dataset. After the cluster centroids were derived, the excluded cases were subsequently assigned to the nearest cluster based on their available features.

Cluster analysis

Cluster analysis was performed using the K-prototypes algorithm, a machine learning method that accommodates mixed-type data consisting of both continuous and categorical variables. The algorithm calculates a composite distance between observations by combining the Euclidean distance for continuous variables and a simple matching dissimilarity for categorical variables. Formally, the distance between observation and cluster prototype is defined as: , where C denotes the set of continuous variables, D denotes the set of categorical variables, if the categorical variable matches and 1 Otherwise, and is a weighting parameter that balances the contribution of categorical variables relative to continuous variables, which was set according to the default implementation in the clustMixType package [21]. Each observation is assigned to the cluster whose prototype minimizes this distance, and cluster prototypes are iteratively updated (centroid for continuous variables, mode for categorical variables) until convergence.

The optimal number of clusters (K) was determined using a combination of the Elbow method and the Gap statistic [19, 21, 22]. The Gap statistic compares the within-cluster sum of squares (WCSS) of the observed data with that expected under a reference null distribution. Formally, it is defined as: , where is the WCSS for the observed data, is the WCSS of the b-th simulated reference dataset, and  B is the number of simulations. The optimal K is chosen as the smallest K such that: , where is the standard error of Gap(K + 1). This provides a robust criterion for cluster selection.

Post-cluster analysis

After the the primary K-prototypes clusters were established, the quality of the identified clusters was further examined using visualization techniques. Factor Analysis for Mixed Data (FAMD) and t-Distributed Stochastic Neighbor Embedding (t-SNE) were applied to project the mixed-type data into a two-dimensional space to facilitate visual assessment of cluster separation and internal cohesion. These visualizations were intended solely for descriptive purposes and were not used as quantitative criteria for determining the optimal number of clusters.

In addition, patients with sporadic missing values who had been excluded from the primary clustering analysis were assigned post hoc to the nearest pre-defined K-prototypes cluster centroid using only the available variables. Outcome variables were excluded from the distance calculation. This secondary exploratory analysis was conducted solely to illustrate how incomplete cases would be classified within the establishing cluster structure and was not intended as missing-data imputation or external validation of the clustering solution.

Statistical Analysis

After cluster identification, the characteristics of each cluster were summarized to examine differences in demographic, clinical, and healthcare utilization patterns. Continuous variables were described using means and Standard deviation (SD), while categorical variables were summarized as proportions. Inter-cluster differences were assessed using analysis of variance (ANOVA) for continuous variables and Chi-square or Fisher’s exact tests for categorical variables. All statistical tests were two-sided, and a p-value < 0.05 was considered statistically significant. To facilitate interpretation of the identified phenotypes, various graphical methods were employed, including stacked bar plots, box plots, radar plots, and heatmaps, to visually highlight patterns and differences across clusters.

All analyses were performed using R (version 4.5.0, R Foundation for Statistical Computing, Vienna, Austria). The main R packages used included “clustMixType”, “klaR”, “ggplot2”, “fmsb”, “dplyr”, and “tidyr”.

Results

Figure 1 illustrates the study enrollment and data processing flow. Out of the initial elderly ED cohort (n = 1665), 52 cases were excluded due to missing vital signs (n = 6), lack of admission records despite ED disposition (n = 40), or formatting errors (n = 6). The remaining 1613 eligible cases were divided into 1519 complete cases for K-prototypes clustering analysis and 94 cases with partially missing data, which were later assigned to the nearest cluster centroid. A detailed breakdown of the reasons for case exclusion is provided in Supplementary Supplementary Table S1.

Cluster K determination

The optimal number of clusters (k) for K-prototypes clustering was determined by evaluating both the Gap Statistic and the within-cluster sum of squares (WSS) across k values ranging from 2 to 10 (Supplementary Fig. S1; Supplementary Table S4). The WSS decreased gradually with increasing k, reflecting improved cluster compactness, while the Gap Statistic remained relatively stable.

Considering both cluster separation and interpretability, k = 3 was selected as the optimal number of clusters. The distribution of the three clusters was further explored using FAMD and t-SNE (Supplementary Fig. S2). As shown in Fig. S2, the two-dimensional projections demonstrated partial overlap between clusters, reflecting the continuous and heterogeneous nature of clinical phenotypes. Nevertheless, distinct density patterns and local grouping tendencies were observed, suggesting meaningful underlying structure. These visualizations provide an intuitive representation of the clustering landscape and complement the quantitative clustering results.

Cluster characterization

Table 1 summarizes selected baseline characteristics of the three clusters identified by K-prototypes clustering. Cluster-specific patterns were further illustrated in Fig. 2, showing (A) ECOG performance status distribution and (B) triage-level distribution as stacked bar plots, and presenting heatmaps of comorbidities and ED diagnoses across clusters. Cluster names, which were assigned based on their characteristic profiles are as follows.

Table 1.

Cluster characterization (post–K-prototypes clustering)

Variables	Post-K-prototypes clustering
	Cluster 1	Cluster 2	Cluster 3
Sample size N=	480	489	550
Age (years)	77.4 ± 7.82	75.7 ± 7.62	74.9 ± 7.35
Sex
Female	55.4%	64.6%	27.5%
Male	44.6%	35.4%	72.5%
Body measurement
BMI (kg/m2)	21.3 ± 3.43	22.3 ± 3.66	26.5 ± 4.76
Upper arm circumferences (cm)	25.3 ± 2.97	27.0 ± 2.81	29.3 ± 3.19
Thigh circumferences (cm)	39.3 ± 4.88	42.2 ± 4.84	46.0 ± 5.92
Calf circumferences (cm)	30.7 ± 3.83	31.8 ± 4.11	33.2 ± 4.23
Chest circumferences (cm)	89.9 ± 9.14	92.4 ± 8.76	95.8 ± 9.52
Hip circumferences (cm)	93.6 ± 9.37	95.2 ± 8.57	100.0 ± 9.82
Waist circumferences (cm)	88.1 ± 9.73	91.3 ± 10.16	99.9 ± 10.86
Clinical Frailty scale	5.0 ± 1.63	3.3 ± 1.71	3.4 ± 1.64
Edmonton Symptom Assessment scale	5.5 ± 2.20	3.0 ± 2.37	3.4 ± 2.11
Vital sign
Heart Rate (bpm)	89.4 ± 18.59	80.0 ± 22.34	87.8 ± 18.92
Respiratory Rate (breaths/min)	17.8 ± 2.74	18.0 ± 4.63	18.0 ± 2.57
Systolic Blood Pressure (mmHg)	142.2 ± 31.42	138.1 ± 32.14	153.7 ± 31.62
Diastolic Blood Pressure (mmHg)	77.9 ± 16.19	74.7 ± 17.45	83.8 ± 17.23
Body Temperature (°C)	36.7 ± 0.96	36.5 ± 0.87	36.7 ± 0.99
SpO2 (%)	96.1 ± 2.39	95.8 ± 4.28	95.7 ± 4.46

Table 1 shows a subset of variables from Supplementary Table 3. See Supplementary Table S3 for full details

Abbreviation BMI Body mass index

Fig. 2.

Stacked bar of (a) ECOG performance status (b) Triage level and Heatmaps of (c) comorbidities and (d) diagnoses across the identified clusters

Cluster 1: frail older adults with high comorbidity

Cluster 1 comprised the oldest patients (mean age 77.4 ± 7.8 years) with the lowest body mass index and circumferential measurements across all body regions, indicating a lean body habitus (Fig. 3). Females accounted for a slight majority (55.4%). This cluster carried a high burden of chronic comorbidities, including diabetes mellitus (39.4%), hypertension (52.1%), cardiovascular disease (26.9%), renal disease (23.5%), neurologic disease (24.6%), and a notably high prevalence of cancer history (47.7%). Functional impairment and frailty were prominent, with the highest CFS (5.0 ± 1.6) and Edmonton Symptom Assessment Scale, (5.5 ± 2.2), as well as a high proportion of patients with ECOG performance status ≥ 3. Clinically, this group demonstrated higher heart rates, a greater proportion of impaired consciousness (GCS < 15).

Fig. 3.

Radar plots for the anthropometric measurements, in terms of standardized Z scores across three clusters

Cluster 2: younger, functionally preserved patients

Cluster 2 patients were relatively younger (75.7 ± 7.6 years) and predominantly female (64.6%), with intermediate body measurements. This group had the lowest prevalence of chronic diseases, including diabetes mellitus (20.9%), hypertension (32.9%), cardiovascular disease (17.8%), and neurologic disease (11.9%). Functional status was the best among the three clusters, reflected by the lowest CFS (3.3 ± 1.7), Edmonton Symptom Assessment Scale (3.0 ± 2.4), and the highest proportion of ECOG performance status 0–1 (63.0%). Most patients were fully alert (GCS 15 in 93.9%) and lower acuity at emergency department triage.

Cluster 3: metabolically burdened, overweight adults

Cluster 3 was predominantly male (72.5%) and characterized by the highest body mass index (26.5 ± 4.8 kg/m²) and largest circumferential measurements, consistent with overweight or obese body composition. This cluster demonstrated the highest prevalence of metabolic and cardiovascular conditions, including diabetes mellitus (43.5%) and hypertension (78.2%), as well as a relatively high burden of cardiovascular disease (26.7%), while the prevalence of cancer was comparatively lower. Functional status and CFS were intermediate between Clusters 1 and 2, with the highest proportion of patients maintaining full consciousness (GCS 15, 96.0%). Vital signs revealed higher systolic and diastolic blood pressure levels.

Healthcare utilization and discharge outcomes across clusters

Post-K-prototypes clustering identified three distinct patient clusters with differing patterns of healthcare utilization (Figs. 4 and 5, Supplementary Tables S3A and S3B). Cluster 1 had the highest median ED cost [4191.5 (IQR 3000–7512.5)] and a high proportion of patients hospitalized to the ward (85.8%), with a median hospital stay of 8 days (IQR 4–12). Most patients in this cluster were ultimately discharged (91.4%), while 8.6% either died or were discharged against medical advice.

Fig. 4.

Distribution of ED cost and hospital stay across three clusters. A ED cost (NTD, log scale). B Hospital stay (days, log scale). C Scatter plot of ED cost and hospital stay by clusters

Fig. 5.

Overall final discharge status by clusters. A Distribution of discharge status. B ED disposition. C Final discharge status for hospitalized patients

Cluster 2 exhibited the lowest ED cost [2580 (IQR 925–4080)] and the highest proportion of patients discharged directly from the emergency department (77.1%). Only 19.4% required hospitalization, and the median hospital stay was minimal at 0.5 days (IQR 0.5–1). The death/AAD rate was highest in this cluster (17.6%), likely reflecting younger but potentially acutely ill patients presenting to the ED. Cluster 3 represented an intermediate-to-high resource utilization group, with a median ED cost of 3780 (IQR 2380–7030) and 67.3% hospitalized to the ward. Median hospital stay was 5 days (IQR 0.5–9), and 88.6% of patients were ultimately discharged, while 10.9% died or experienced AAD. Disposition patterns, hospital stay, and costs were all statistically significant across clusters in pairwise comparisons (Supplementary Tables S3B).

Assignment of incomplete cases to pre-defined clusters

When applying post-hoc clustering to incomplete cases using only available variables (excluding outcome variables from distance calculation), all 94 cases were assigned to the pre-defined Cluster 3 (Supplementary Table S6). This indicates that, under the current cluster structure, the available information from these incomplete cases was most similar to the centroid of Cluster 3.

Furthermore, comparisons between these assigned cases and the predefined Cluster 3 centroid (using variables without missing data) showed minimal deviations. The mean differences were − 181.85 for ED cost and − 3.05 days for hospital stay, and discharge status concordance reached 73%.

Discussion

In this study, we applied an unsupervised K-prototypes machine learning algorithm to a cohort of older adults presenting to the emergency department (ED). By analyzing mixed-type data—including continuous physiological parameters, anthropometric measurements, and categorical clinical variables—we identified three distinct, clinically meaningful patient phenotypes: Frail Older Adults with High Comorbidity (Cluster 1), Younger, Functionally Preserved Patients (Cluster 2), and Metabolically Burdened, Overweight Adults (Cluster 3). These clusters exhibited differences in baseline characteristics, healthcare resource utilization, and short-term clinical outcomes. Our findings suggest that traditional ED triage systems, which primarily rely on physiological parameters and presenting complaints, may not fully capture the multidimensional characteristics of older adults. In this context, a multidimensional phenotyping approach may provide additional insights into the heterogeneity of geriatric patients in the acute care setting.

Cluster 1 appears to represent a clinically vulnerable geriatric phenotype. These patients were the oldest, with a lower body mass index and smaller anthropometric measurements, and had a high burden of chronic medical conditions, particularly previous malignancies and neurological diseases. This pattern is consistent with a lean body habitus, which may suggest strongly underlying sarcopenia, malnutrition or sarcopenic vulnerability. Accordingly, this group presented with the poorest baseline functional status, reflecting the highest frailty and Edmonton Symptom Assessment Scale. Consequently, patients in Cluster 1 had a higher ED costs, with high ward admission rate (85.8%) and intensive care unit (ICU) admission rate (3.5%). Also, cluster 1 experienced the longest median length of hospital stay. The association between physical frailty, multimorbidity, prolonged hospitalization, and heightened healthcare expenditure has been well documented in geriatric emergency medicine [23, 24]. For this deeply complex phenotype, standard, disease-specific emergency diagnostic pathways are frequently insufficient and may expose frail patients to iatrogenic harm and functional decline. Our findings advocate for the targeted integration of comprehensive geriatric assessment (CGA) and multidisciplinary interventions, such as early palliative care screening and transitional care nursing, directly at the ED front door [20, 25, 26].

In contrast, Cluster 2 presented a paradoxical clinical profile. Comprising the youngest subset of our elderly cohort, these patients were predominantly female, had the lowest prevalence of chronic diseases, and exhibited the highest level of functional independence (ECOG 0–1). Unsurprisingly, a large proportion of these patients (77.1%) were discharged directly from the ED, contributing to the lowest overall median ED costs and shorter hospital stays. However, a critical trend emerged within the small subset (19.4%) that required admission: their mortality or AAD rate was disproportionately the highest across all clusters (17.6%). This bimodal outcome distribution—where patients experience either rapid discharge or catastrophic deterioration—reflects a critical phenomenon. Functionally independent older adults generally avoid the ED unless struck by sudden, high-acuity, and often devastating pathologies, such as massive hemorrhagic strokes, large-territory ischemia strokes, or severe sepsis [20, 27]. Because their robust baseline physiological function visually masks their acute vulnerability, these patients may be at severe risk of experiencing the “triage illusion,” wherein standardized triage algorithms may underestimate their true illness severity based on a lack of obvious red flags or chronic comorbidities [28]. Emergency clinicians must therefore maintain a heightened index of suspicion when previously healthy older adults present with acute complaints, as their physiological decompensation can be remarkably rapid and highly lethal.

Cluster 3 outlines a distinct phenotype primarily characterized by metabolic vulnerability. Predominantly male, this group exhibited the highest body mass index and largest circumferential body measurements, coupled with a high prevalence of hypertension (78.2%) and diabetes mellitus (43.5%). Despite these cardiometabolic risk factors, their baseline functional status and frailty scores were notably better preserved than those of Cluster 1. This potentially reflects the well-documented “obesity paradox” often observed in geriatric populations, where a higher body mass index may paradoxically offer a protective energy reserve against acute wasting illnesses and severe physiological stress [29]. Nevertheless, their hospital admission rate remained substantial (67.3%), likely driven by acute exacerbations of their underlying chronic cardiometabolic conditions. Emergency care pathways for this metabolically burdened cluster should prioritize rigorous hemodynamic monitoring, aggressive metabolic stabilization, and targeted transitions to outpatient secondary prevention programs to effectively reduce the incidence of future cardiovascular readmissions [30].

From a methodological perspective, the K-prototypes algorithm is well suited for clinical datasets that include both continuous and categorical variables. Traditional machine learning models often focus on predefined outcomes or require transformation of categorical variables, which may not fully capture the multidimensional nature of heterogeneous geriatric populations. Selection of the optimal number of clusters requires consideration of both statistical criteria and clinical interpretability. Although several candidate solutions demonstrated similar Gap statistic values, the three-cluster solution provided relatively balanced cluster sizes and yielded clinically distinct phenotypes that were straightforward to interpret. Increasing the number of clusters resulted in smaller and less clinically meaningful subgroups, reducing the practical applicability of the clustering results.

Limitations

Several limitations to our study should be acknowledged. First, as a retrospective, single-center study conducted in Taiwan using data from a research cohort that enrolled only non-traumatic older adults presenting to the emergency department, the newly identified phenotypes may heavily reflect local institutional admission practices, regional demographic characteristics, and specific healthcare-seeking behaviors. Therefore, the findings may not be generalizable to trauma-related presentations (e.g., falls) or to other settings. Second, patients with sporadic missing values were assigned to the nearest predefined cluster centroids using available variables in a secondary exploratory analysis. Because incomplete information was used for the assignment, the ability to distinguish among clusters may have been limited. Therefore, this procedure was intended solely to explore how incomplete cases aligned with the existing cluster structure and should not be interpreted as formal external validation of the clustering solution. Future studies are warranted to evaluate more robust approaches for handling incomplete data in clinical phenotyping. Third, because healthcare utilization variables contributed to the clustering procedure, subsequent comparisons of these variables were intended solely to descriptively characterize the identified phenotypes and should not be interpreted as independent inferential findings. Fourth, the use of National Health Insurance point values as a proxy for ED costs may not fully capture the indirect economic burden and out-of-pocket expenditures. Finally, the outcomes were strictly restricted to short-term ED disposition and immediate hospital discharge status. Further longitudinal studies evaluating outcomes such as 30-day readmission rates, long-term functional decline, and sustained survival are required to fully validate the prognostic value and clinical utility of these specific phenotypes.

Conclusion

Unsupervised K-prototypes clustering successfully identified three distinct clinical phenotypes among elderly ED patients: frail/comorbid, functionally preserved and metabolically burdened. These findings highlight the heterogeneity of geriatric patients in the ED and suggest that multidimensional phenotyping may provide additional insights beyond traditional triage approaches.

Abbreviations

AAD

Against advice discharge

ANOVA

Analysis of variance

BMI

Body mass index

CGA

Comprehensive geriatric assessment

ECOG

Eastern Cooperative Oncology Group

ED

Emergency department

ESI

Emergency Severity Index

FAMD

Factor Analysis for Mixed Data

GCS

Glasgow Coma Scale

ICD-10

International Classification of Diseases, Tenth Revision

ICU

Intensive care unit

IQR

Interquartile range

LCGMH

Linkou Chang Gung Memorial Hospital

LOS

Length of stay

MAR

Missing at random

ML

Machine learning

SD

Standard deviation

t-SNE

t-Distributed Stochastic Neighbor Embedding

TTAS

Taiwan Triage and Acuity Scale

WCSS

Within-cluster sum of squares

Authors’ contributions

SCH and YCS conceptualized and designed the study. CYC and HJT contributed to data acquisition, data curation, and methodology. SCH and CYC performed the formal data analysis and visualization. SCH drafted the original manuscript. YCS supervised the project and critically revised the manuscript for important intellectual content. All authors read and approved the final manuscript.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Data availability

The data sets generated or analyzed during this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The study protocol adhered to the ethical guidelines of the Declaration of Helsinki. It was reviewed and approved by the Institutional Review Board of the Chang Gung Medical Foundation (Approval No. 202400144B0). To strictly maintain patient privacy and data confidentiality, all institutional identification numbers were cryptographically scrambled prior to database access. Due to the retrospective and anonymized nature of the dataset, the requirement for informed consent was formally waived by the approving institutional review board.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.