The SUPREME Index for 30-day all-cause readmission: development and internal validation of a machine learning-based risk stratification index in community-dwelling adults Authors: Yeong Jun Ju, Soon Young Lee Published: 2026 ABSTRACT BackgroundHospital readmissions are a key indicator of healthcare quality and have substantial implications for patient outcomes and system-level costs. Identifying individuals at high risk of experiencing 30-day all-cause readmission events is essential for implementing preventive strategies and improving resource allocation. This study aimed to develop and internally validate a simple, interpretable risk stratification index for identifying individuals at elevated risk of experiencing at least one 30-day all-cause readmission event in a community-dwelling general population using machine learning (ML)-based feature selection and logistic regression (LR)-based risk scoring.MethodsWe analyzed data on adults residing in Suwon City from the 2016–2019 South Korean National Health Insurance Service–National Sample Cohort. The outcome was defined at the person level as experiencing at least one all-cause readmission within 30 days of discharge from any hospitalization during the study period. Elastic net (EN) regularization was used to select candidate predictors, and LR was used to derive a point-based scoring index. Model performance was evaluated in an internal validation dataset using the area under the receiver operating characteristic curve (AUC), calibration curves, and Brier score.ResultsAmong 3,357 adults, 357 (10.6%) experienced at least one 30-day readmission event during the study period. Four of seven features selected by EN (top-quartile number of hospital admissions, length of stay, total healthcare costs, and any malignancy) were retained to construct the SUwon Population-based Readmission risk Estimation ModEl (SUPREME) Index, with scores ranging from 0 to 14. Internal validation showed good performance (AUC = 0.861; Brier score = 0.067) with adequate calibration. A cutoff score of 8 provided a balanced operating point (sensitivity 0.742; specificity 0.867) for classifying individuals as high risk.ConclusionsThe SUPREME Index, developed using EN and LR, provides a practical and interpretable tool for person-level risk stratification using administrative claims data. Because predictors were derived from routinely available claims and screening data and were not anchored to a single uniform baseline time point, the index should be interpreted as a stratification tool rather than an event-level prediction model anchored to a specific index hospitalization. External validation and evaluation in diverse populations are warranted.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12911-026-03536-8. Background Hospital readmissions are a substantial contributor to healthcare costs, often reflecting gaps in quality and continuity of care, which may lead to adverse patient outcomes [1]. Consequently, minimizing the incidence of readmission to reduce expenditure without compromising quality is a healthcare system priority worldwide [2, 3]. To this end, research has focused on developing validated tools for identifying patients at high risk of readmission [4, 5]. To stratify the risk of hospital readmission, multiple tools have been developed using a range of approaches, from conventional regression methods to machine-learning (ML) techniques, each of which has advantages and limitations [6, 7]. Risk scoring and classification approaches based on conventional regression methods, such as logistic regression (LR) and Cox regression, have advantages in terms of ease of implementation, interpretability, and transparency, but their performance and applicability may be limited when used with high-dimensional healthcare data unless complex functional forms and interactions are explicitly specified [8]. In contrast, more flexible algorithmic approaches, particularly nonparametric methods, can model nonlinear relationships and higher-order interactions with less manual specification, but have traditionally been viewed as less transparent in explaining predictor contributions, which may limit acceptance and clinical application [9]. Recent advances in explainable artificial intelligence (XAI) techniques, such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), can provide post-hoc explanations of feature contributions for complex nonparametric models, partially mitigating these interpretability concerns [10, 11]. At the same time, approaches that emphasize model simplification and inherent interpretability to facilitate practical implementation have gained attention. Developing readmission risk stratification tools with high discriminative performance remains a challenge [5, 12, 13]. Recent efforts have therefore focused on hybrid strategies that combine complementary modeling components to improve risk classification performance [14–16]. For example, candidate predictors may be selected using regularized regression techniques, such as LASSO or elastic net (EN), which can accommodate high-dimensional feature spaces while reducing complexity through shrinkage and variable selection [17]. The selected predictors may then be incorporated into a conventional regression model to derive a parsimonious, point-based risk stratification tool, thereby combining data-driven feature selection with transparent effect estimation [15, 18]. Such hybrid approaches have been reported to achieve improved classification performance compared with models built using a single component alone [14–16]. In this study, we combined ML techniques, including EN, with LR to develop a risk stratification index for identifying individuals in the community-dwelling general population with elevated risk of experiencing at least one all-cause 30-day readmission event during the study period. To explore the potential utility of the index in the local healthcare setting, we intentionally selected a community-based general population from the region in which the investigators reside. By adopting an all-cause approach rather than a disease-specific focus, we aimed to enhance the pragmatic utility of the index for population-level risk stratification and local healthcare planning. Methods Study design We aimed to develop an all-cause 30-day readmission risk stratification scoring index in a community-based general population, with the goal of identifying individuals at elevated risk. The study was conducted in three steps: Predictor variables were selected using EN regularization, an ML-based penalization method for optimal feature selection;The selected predictors were incorporated into an LR model to derive odds ratios, which were then used to construct a scoring index; andThreshold values were determined to stratify individuals into risk groups for potential targeting of preventive interventions. Predictor variables were selected using EN regularization, an ML-based penalization method for optimal feature selection; Predictor variables were selected using EN regularization, an ML-based penalization method for optimal feature selection; The selected predictors were incorporated into an LR model to derive odds ratios, which were then used to construct a scoring index; and The selected predictors were incorporated into an LR model to derive odds ratios, which were then used to construct a scoring index; and Threshold values were determined to stratify individuals into risk groups for potential targeting of preventive interventions. Threshold values were determined to stratify individuals into risk groups for potential targeting of preventive interventions. The overall model development procedure is illustrated in Fig. 1. Fig. 1Development process of the SUPREME Index Development process of the SUPREME Index Data source and study population We used the 2016–2019 data from the National Health Insurance Service–National Sample Cohort (NHIS-NSC; Sample Cohort DB) in South Korea, a de-identified population-based database comprising approximately one million National Health Insurance and Medical Aid beneficiaries (about 2% of eligible individuals) selected via stratified sampling by sex, age, residential area, insurance premium level, and type of subscription. The NHIS-NSC provides longitudinally linkable eligibility records (including sociodemographic characteristics and death information), nationwide inpatient and outpatient claims, and national health screening records for participants who underwent NHIS health screenings, along with healthcare provider information. Because reimbursed services are billed to and adjudicated within Korea’s single-insurer national health insurance system, the claims data capture covered healthcare utilization across healthcare facilities nationwide, enabling ascertainment of inpatient admissions and subsequent readmissions irrespective of hospital location. Claims data reflect reimbursed care within the benefit package and therefore do not include non-covered services. The study population consisted of individuals ≥ 19 years who had at least one hospitalization record during the study period. Among them, residents of Suwon City were identified using the regional code corresponding to each individual’s registered residential address; hospitalizations occurring at any facility nationwide were included. We excluded individuals who died during the study period (n = 42) to ensure complete outcome ascertainment within the observation window. After these exclusions, a total of 3,591 individuals were initially identified. We then excluded those with missing data on admission type (n = 177) and insurance premium-based income level (n = 57). As a result, the final analytic sample comprised 3,357 individuals. Outcomes The outcome was defined as the occurrence of any all-cause hospital readmission within 30 days of discharge from any hospitalization during the study period. To comprehensively reflect the diverse patterns of healthcare utilization in a community-based population with relatively low admission rates, all hospitalization records were included in the analysis. The data were then aggregated at the individual level, and individuals were classified as having experienced a 30-day readmission if they had at least one qualifying readmission event. Accordingly, individuals with one event and those with multiple events were categorized identically for the primary outcome definition. This approach aligns with previous studies that have used varying definitions of readmission based on the population characteristics and study objectives [19, 20]. Potential predictors Candidate predictors were selected based on factors previously reported to be associated with all-cause 30-day readmission [12], including admission characteristics and comorbidities that had been validated in various clinical settings, including community-dwelling populations [21–26], subject to their availability within the dataset. To avoid unnecessary model complexity, individual comorbid conditions defined using Charlson Comorbidity Index (CCI) disease categories were screened using chi-squared tests, and only those significantly associated with 30-day readmission were retained in the candidate predictor set. Twenty-seven potential factors were extracted: (1) socio-demographic characteristics: sex, age, and income; (2) health behaviors: smoking, drinking, physical activity, and body mass index; (3) comorbidities and chronic conditions: diabetes mellitus and hypertension (identified from health screening records), congestive heart failure, peripheral vascular disease, cerebrovascular disease, chronic pulmonary disease, mild liver disease, and any malignancy (identified based on CCI categories), along with the CCI score and hospital frailty risk index group; (4) medical utilization: emergency department (ED) visit, ED visit 6 months before the first hospitalization, total healthcare costs (top quartile), total length of hospital stay (top quartile), total number of hospital admissions (top quartile), total outpatient days (top quartile), length of stay for the first hospitalization (top quartile), admission type for the first hospitalization, surgery during first hospitalization, and surgery during ED visits. The hospital frailty risk index was calculated using algorithms based on ICD-10 codes, validated by Gilbert et al. [27]. Because the primary analysis was conducted at the individual level, predictors were operationalized using the information available in the database, including (i) variables defined at the time of the first hospitalization during the study period (e.g., admission type, length of stay, and in-hospital procedures) and (ii) variables summarized over the study period (e.g., total costs and counts of healthcare utilization). Comorbidities and chronic conditions derived from claims data were identified using ICD-10 code–based Charlson Comorbidity Index categories and were defined as the presence of each condition at any time during the study period (2016–2019). Health behaviors and screening-based clinical measures were obtained from the most recent available health screening record within the database. For healthcare utilization measures, including total healthcare costs, total length of hospital stay, total number of hospital admissions, and total outpatient days, cumulative values were calculated over the entire study period (2016–2019). These variables were then log-transformed and categorized into quartiles based on their respective distributions, and binary indicators for the upper quartile were used in model development. ED visit during the study period and ED visit within 6 months before the first hospitalization were coded as binary indicators (yes or no). A single uniform baseline date was not assigned across all predictors; rather, the ascertainment time point depended on the operational definition of each variable. Accordingly, predictor ascertainment did not correspond to a single uniform baseline time point across all variables. Several healthcare utilization variables, including admission type, length of stay, and surgical procedure during the first hospitalization, as well as surgery during ED visits, were included as proxy indicators of the clinical acuity and severity of the initial hospitalization episode [5]. The use of upper-quartile binary indicators for utilization-related variables was intended to identify high-utilization subgroups while preserving structural simplicity and statistical stability in the EN-based risk scoring framework [28]. Prior evidence has also suggested that patients with high healthcare utilization are more likely to experience readmission within 30 days [29]. Feature coding and pre-processing All predictors were categorical at the modeling stage. Continuous healthcare utilization and cost variables were pre-processed as described in the preceding section, including log-transformation and upper-quartile binary encoding. Predictors were encoded as dummy variables for elastic net fitting, and no additional scaling or normalization was applied. Data processing and statistical analysis The entire dataset was randomly allocated into a training set (n = 2,686; 80%) and an internal validation set (n = 671; 20%) using stratified sampling to preserve the distribution of 30-day readmission. The training set was used for feature selection and index derivation, and the internal validation set was used to evaluate the operating characteristics of the index. Descriptive statistics were performed to examine differences in patient characteristics according to 30-day readmission status. Categorical variables were expressed as frequencies and percentages and groups were compared using the chi-squared test. All analyses were performed using SAS (version 9.4; SAS Institute, Cary, NC, USA) or R (version 4.4.1; The R Foundation for Statistical Computing, Vienna, Austria) software. Feature selection and model development EN is a regularization technique that combines L1 (LASSO) and L2 (Ridge) penalties to address multicollinearity and facilitate variable selection in regression models [17, 28]. The L1 penalty drives less informative coefficients to exactly zero, enabling automatic feature selection, whereas the L2 penalty shrinks correlated predictors jointly without eliminating them, thereby stabilizing estimates [17, 28]. In this study, EN was applied for feature selection among 27 potential predictors. Hyperparameter optimization involved a two-phase grid search: initially, a broad exploration of α (the mixing ratio of L1 and L2 penalties) and λ (overall penalty strength) across extensive log-scaled ranges using 20-fold cross-validation to identify promising regions with high AUCs. Focused searches were then conducted within these regions, using the same cross-validation protocol, to determine the optimal hyperparameter values. These tuning procedures were performed exclusively within the training set, and the internal validation set was held out and not used at any stage of hyperparameter selection. Based on the optimal hyperparameters, LR was applied to the set of predictors selected by EN. Among these, statistically significant variables were retained, and their corresponding odds ratios were rounded to the nearest integer value to assign individual scores. These scores were then summed to construct the SUwon Population-based Readmission risk Estimation ModEl (SUPREME) Index, a parsimonious risk stratification tool designed to classify individuals according to their likelihood of experiencing at least one all-cause 30-day readmission event during the study period. Index evaluation The performance of the SUPREME Index was evaluated in the internal validation dataset using both calibration and receiver operating characteristic (ROC) curves. Calibration performance was investigated through the calibration curve and the Brier score. Calibration curves were constructed using 1,000 bootstrap resamples to adjust for potential overfitting and estimate bias-corrected predicted probabilities. Risk stratification The optimal cutoff value for stratifying high-risk patients using the SUPREME Index was identified through a two-step process. First, we screened candidate thresholds across the score range using sensitivity, specificity, and Youden’s Index. We then selected three representative cutoffs (6, 8, and 10) to reflect sensitivity-oriented, balanced, and specificity-oriented operating points, respectively, and compared their operating characteristics. For each candidate threshold, a confusion matrix was constructed to calculate performance metrics, including sensitivity, specificity, positive and negative predictive values, F1 score, and overall classification performance. The final cutoff value was determined by comparing these metrics across the candidate thresholds, with comparable emphasis placed on sensitivity and specificity to identify a pragmatic balanced operating point. Results Participant characteristics Among the 3,357 study participants, 357 (10.6%) experienced at least one 30-day readmission event during the study period. Readmission was more likely among individuals with specific comorbidities and healthcare utilization characteristics, specifically those with any malignancy, elevated CCI scores, frequent ED visits, and those in the top quartile for healthcare utilization, as measured by total healthcare costs, length of stay, and number of hospital admissions (Table 1). In addition, 42 individuals were excluded because they died during the study period. This excluded group was older and clinically more vulnerable than the analytic cohort: 59.5% were aged ≥ 65 years, 66.7% had a Charlson Comorbidity Index score ≥ 3, and 16.7% were classified as high frailty risk (Supplementary Table S1). Table 1General characteristics of the study participantsVariablesTotalAll-cause 30-day readmissionp-valueNoYes N % N % N % Sex 0.789 Male1,45443.3129789.215710.8 Female1,90356.71,70389.520010.5 Age < 0.001 19 ~ 441,15734.51,07893.2796.8 45 ~ 641,61948.21,43588.618411.4 65+58117.348783.89416.2 Income 0.556 Q11,25737.41,13190.012610.0 Q21,13333.81,00488.612911.4 Q396728.886589.510210.6 Smoking 0.956 No2,44973.02,18989.426010.6 Yes90827.181189.39710.7 Drinking 0.067 No1,11533.298188.013412.0 Yes2,24266.82,01990.122310.0 Physical Activity 0.195 No1,12933.699888.413111.6 Yes2,22866.42,00289.922610.1 Body Mass Index 0.329 Normal (BMI < 25.0)2,03660.71,82889.820810.2 Obesity (BMI ≥ 25.0)1,32139.41,17288.714911.3 Diabetes mellitus 0.005 No2,80083.42,52190.027910.0 Yes55716.647986.07814.0 Hypertension < 0.001 No2,21265.92,01291.02009.0 Yes1,14534.198886.315713.7 Congestive heart failure 0.023 No3,26197.12,92189.634010.4 Yes962.97982.31717.7 Peripheral vascular disease 0.001 No3,03890.52,73289.930610.1 Yes3199.526884.05116.0 Cerebrovascular disease < 0.001 No2,98288.82,69090.22929.8 Yes37511.231082.76517.3 Chronic pulmonary disease 0.009 No2,02960.41,83690.51939.5 Yes1,32839.61,16487.716412.4 Mild liver disease 0.062 No2,60477.62,34189.926310.1 Yes75322.465987.59412.5 Any malignancy < 0.001 No3,07191.52,78490.72879.4 Yes2868.521675.57024.5 CCI score group < 0.001 083925.078293.2576.8 1 ~ 21,64449.01,51892.31267.7 3+87426.070080.117419.9 Hospital Frailty Risk index group < 0.001 Low risk (< 5)1,83754.71,69792.41407.6 Intermediate risk (5–15)1,44243.01,25887.218412.8 High risk(> 15)782.34557.73342.3 Any ED visit < 0.001 No2,37570.82,19192.31847.8 Yes98229.380982.417317.6 Any ED visit within 6 months before the first hospitalization 0.332 No3,22095.92,88189.533910.5 Yes1374.111986.91813.1 Top-quartile total healthcare costs < 0.001 No2,51875.02,40595.51134.5 Yes83925.059570.924429.1 Top-quartile total length of hospital stay < 0.001 No2,51875.02,43796.8813.2 Yes83925.056367.127632.9 Top-quartile total number of hospital admissions < 0.001 No2,51875.02,44597.1732.9 Yes83925.055566.228433.9 Top-quartile total outpatient days < 0.001 No2,51875.02,31391.92058.1 Yes83925.068781.915218.1 Top-quartile LOS for the first hospitalization 0.001 No2,51875.02,27590.42439.7 Yes83925.072586.411413.6 Admission type for the first hospitalization 0.109 ED73521.964587.89012.2 Outpatient2,62278.12,35589.826710.2 Surgery during the first hospitalization < 0.001 No1,63348.61,40185.823214.2 Yes1,72451.41,59992.81257.3 Surgery during ED visits < 0.001 No3,17194.52,85289.931910.1 Yes1865.514879.63820.4 Total 3,357100.03,00089.435710.6Notes: Unless otherwise specified, “total” variables were calculated by aggregating claims over the study period at the person level. “First hospitalization” refers to the first hospitalization during the study period Table 1General characteristics of the study participantsVariablesTotalAll-cause 30-day readmissionp-valueNoYes N % N % N % Sex 0.789 Male1,45443.3129789.215710.8 Female1,90356.71,70389.520010.5 Age < 0.001 19 ~ 441,15734.51,07893.2796.8 45 ~ 641,61948.21,43588.618411.4 65+58117.348783.89416.2 Income 0.556 Q11,25737.41,13190.012610.0 Q21,13333.81,00488.612911.4 Q396728.886589.510210.6 Smoking 0.956 No2,44973.02,18989.426010.6 Yes90827.181189.39710.7 Drinking 0.067 No1,11533.298188.013412.0 Yes2,24266.82,01990.122310.0 Physical Activity 0.195 No1,12933.699888.413111.6 Yes2,22866.42,00289.922610.1 Body Mass Index 0.329 Normal (BMI < 25.0)2,03660.71,82889.820810.2 Obesity (BMI ≥ 25.0)1,32139.41,17288.714911.3 Diabetes mellitus 0.005 No2,80083.42,52190.027910.0 Yes55716.647986.07814.0 Hypertension < 0.001 No2,21265.92,01291.02009.0 Yes1,14534.198886.315713.7 Congestive heart failure 0.023 No3,26197.12,92189.634010.4 Yes962.97982.31717.7 Peripheral vascular disease 0.001 No3,03890.52,73289.930610.1 Yes3199.526884.05116.0 Cerebrovascular disease < 0.001 No2,98288.82,69090.22929.8 Yes37511.231082.76517.3 Chronic pulmonary disease 0.009 No2,02960.41,83690.51939.5 Yes1,32839.61,16487.716412.4 Mild liver disease 0.062 No2,60477.62,34189.926310.1 Yes75322.465987.59412.5 Any malignancy < 0.001 No3,07191.52,78490.72879.4 Yes2868.521675.57024.5 CCI score group < 0.001 083925.078293.2576.8 1 ~ 21,64449.01,51892.31267.7 3+87426.070080.117419.9 Hospital Frailty Risk index group < 0.001 Low risk (< 5)1,83754.71,69792.41407.6 Intermediate risk (5–15)1,44243.01,25887.218412.8 High risk(> 15)782.34557.73342.3 Any ED visit < 0.001 No2,37570.82,19192.31847.8 Yes98229.380982.417317.6 Any ED visit within 6 months before the first hospitalization 0.332 No3,22095.92,88189.533910.5 Yes1374.111986.91813.1 Top-quartile total healthcare costs < 0.001 No2,51875.02,40595.51134.5 Yes83925.059570.924429.1 Top-quartile total length of hospital stay < 0.001 No2,51875.02,43796.8813.2 Yes83925.056367.127632.9 Top-quartile total number of hospital admissions < 0.001 No2,51875.02,44597.1732.9 Yes83925.055566.228433.9 Top-quartile total outpatient days < 0.001 No2,51875.02,31391.92058.1 Yes83925.068781.915218.1 Top-quartile LOS for the first hospitalization 0.001 No2,51875.02,27590.42439.7 Yes83925.072586.411413.6 Admission type for the first hospitalization 0.109 ED73521.964587.89012.2 Outpatient2,62278.12,35589.826710.2 Surgery during the first hospitalization < 0.001 No1,63348.61,40185.823214.2 Yes1,72451.41,59992.81257.3 Surgery during ED visits < 0.001 No3,17194.52,85289.931910.1 Yes1865.514879.63820.4 Total 3,357100.03,00089.435710.6Notes: Unless otherwise specified, “total” variables were calculated by aggregating claims over the study period at the person level. “First hospitalization” refers to the first hospitalization during the study period General characteristics of the study participants Any ED visit within 6 months before the first hospitalization Notes: Unless otherwise specified, “total” variables were calculated by aggregating claims over the study period at the person level. “First hospitalization” refers to the first hospitalization during the study period SUPREME Index development and validation Hyperparameter tuning via grid search identified α = 0.445 and λ = 0.0203 as the optimal combination (Supplementary Fig. S1). At this λ value, EN regularization retained seven variables with non-zero coefficients from the 27 candidates: total number of hospital admissions (top quartile), total length of hospital stay (top quartile), total healthcare costs (top quartile), hospital frailty risk index group, any malignancy, CCI group, and ED visit. The coefficient path plot illustrating the variable selection process is presented in Supplementary Fig. S2, and the relative importance of the selected variables is shown in Supplementary Fig. S3. Of these seven variables, four showed statistically significant associations in multivariable logistic regression and were retained for the final SUPREME Index: top-quartile total number of hospital admissions, top-quartile total length of hospital stay, top-quartile total healthcare costs, and any malignancy. The total score ranged from 0 to 14 (Table 2). Table 2Variables included in the SUPREME Index based on multivariable logistic regression in the development datasetVariablesAll-cause 30-day readmissionAdjusted-OR (95% CI)*p-valueRisk Score Top-quartile total number of hospital admissions No1.006 Yes6.34 (4.43, 9.10)< 0.001 Top-quartile total length of hospital stay No1.004 Yes4.49 (3.14, 6.44)< 0.001 Top-quartile total healthcare costs No1.002 Yes1.58 (1.11, 2.26)0.010 Any malignancy No1.002 Yes1.56 (1.01, 2.44)0.048*Adjusted for top-quartile total number of hospital admissions, top-quartile total length of hospital stay, top-quartile total healthcare costs, hospital frailty risk index group, any malignancy, CCI score group, and ED visit Table 2Variables included in the SUPREME Index based on multivariable logistic regression in the development datasetVariablesAll-cause 30-day readmissionAdjusted-OR (95% CI)*p-valueRisk Score Top-quartile total number of hospital admissions No1.006 Yes6.34 (4.43, 9.10)< 0.001 Top-quartile total length of hospital stay No1.004 Yes4.49 (3.14, 6.44)< 0.001 Top-quartile total healthcare costs No1.002 Yes1.58 (1.11, 2.26)0.010 Any malignancy No1.002 Yes1.56 (1.01, 2.44)0.048*Adjusted for top-quartile total number of hospital admissions, top-quartile total length of hospital stay, top-quartile total healthcare costs, hospital frailty risk index group, any malignancy, CCI score group, and ED visit Variables included in the SUPREME Index based on multivariable logistic regression in the development dataset *Adjusted for top-quartile total number of hospital admissions, top-quartile total length of hospital stay, top-quartile total healthcare costs, hospital frailty risk index group, any malignancy, CCI score group, and ED visit The model performance was evaluated based on calibration and ROC curves using the internal validation set. The bias-corrected calibration curve, obtained through bootstrap resampling, closely followed the ideal 45-degree line across the full range of predicted probabilities, indicating adequate calibration. The readmission probabilities were slightly overestimated in the 0.1 to 0.2 range; however, the bias-corrected curve remained consistently aligned with the ideal line in other regions. The mean absolute error of 0.013 and Brier score of 0.067 demonstrated that the model was adequately calibrated (Fig. 2A), and the AUC of 0.861 demonstrated good discrimination for identifying individuals who experienced at least one 30-day readmission event during the study period (Fig. 2B). Fig. 2Model performance in calibration and discrimination of the SUPREME Index in the internal validation set, as assessed by (A) the calibration curve and (B) the ROC curve. ROC: receiver operating characteristic; AUC: area under the curve Model performance in calibration and discrimination of the SUPREME Index in the internal validation set, as assessed by (A) the calibration curve and (B) the ROC curve. ROC: receiver operating characteristic; AUC: area under the curve Risk stratification Risk stratification was performed using SUPREME Index thresholds of 6, 8, and 10. Candidate thresholds were screened across the score range using sensitivity, specificity, and Youden’s Index, and three representative cutoffs (6, 8, and 10) were selected to reflect sensitivity-oriented, balanced, and specificity-oriented operating points. The overall differences in performance metrics across the three thresholds were modest. A cutoff of 6 showed the highest sensitivity (0.840) but comparatively lower specificity (0.772) and F1 score (0.450), whereas a cutoff of 10 showed higher specificity (0.900) and lower sensitivity (0.704). A cutoff of 8 provided the most balanced performance across sensitivity (0.742), specificity (0.867), and F1 score (0.521), and was therefore selected as the threshold for identifying individuals at high risk (Fig. 3). To examine potential heterogeneity by age at the selected operating point, we additionally evaluated operating characteristics at a cutoff of 8 within age strata. Sensitivity and specificity were 0.548 and 0.924 in the 19–44 group, 0.792 and 0.858 in the 45–64 group, and 0.810 and 0.768 in the ≥ 65 group, respectively (Supplementary Table S2). Fig. 3Comparison of performance metrics for risk stratification using SUPREME Index thresholds of 6 (solid blue line), 8 (dotted red line), and 10 (solid green line). Threshold 6 showed the highest sensitivity, whereas threshold 10 demonstrated the highest specificity. Threshold 8 provided the most balanced overall performance and was selected as the pragmatic cutoff. PPV: positive predictive value; NPV: negative predictive value Comparison of performance metrics for risk stratification using SUPREME Index thresholds of 6 (solid blue line), 8 (dotted red line), and 10 (solid green line). Threshold 6 showed the highest sensitivity, whereas threshold 10 demonstrated the highest specificity. Threshold 8 provided the most balanced overall performance and was selected as the pragmatic cutoff. PPV: positive predictive value; NPV: negative predictive value Discussion By combining elastic net (EN) regularization with logistic regression (LR), this study developed a pragmatic risk-stratification index (the SUwon Population-based Readmission risk Estimation ModEl [SUPREME] Index) to classify community-dwelling adults according to their risk of experiencing at least one all-cause 30-day readmission event during the study period. EN regularization was applied to candidate variables derived from administrative claims and screening records, and the predictors selected by EN were subsequently modeled using LR to derive a point-based scoring index based on adjusted odds ratios. The final index comprised four variables (top-quartile total number of hospital admissions, top-quartile total length of hospital stay, top-quartile total healthcare costs, and any malignancy) and showed good discrimination and calibration in internal validation, as assessed by AUC and calibration measures. Candidate cutoffs were screened using sensitivity, specificity, and Youden’s Index, and three representative thresholds (6, 8, and 10) were compared using confusion matrix–derived metrics; a cutoff of 8 was selected as a balanced operating point for risk stratification. Importantly, the SUPREME Index is intended for population-level risk stratification rather than time-anchored prediction from a uniform baseline (e.g., an index hospitalization). All four predictor variables included in the SUPREME Index have been identified in previous studies as risk factors for readmission within 30 days. A higher number of previous hospitalizations [1, 30], a longer length of stay [25, 26, 31], and a higher per capita healthcare cost, which is associated with higher inpatient service use [32, 33], have been associated with a higher likelihood of readmission within 30 days. However, the association between length of stay and 30-day readmission is context-dependent and may differ across cohorts and care settings. Malignancy has been identified as a disease category associated with high readmission rates owing to complex treatment pathways and a high risk of complications [31, 34]. The inclusion of both healthcare utilization and clinical condition variables in the SUPREME Index is consistent with factors reported in prior readmission literature [12], supporting the clinical plausibility of the selected predictors; however, external validation is required to evaluate transportability to other populations and time periods. In this context, the SUPREME Index leverages routinely available utilization and diagnostic information that is commonly captured in claims-based settings, supporting its potential feasibility for implementation where such data are readily accessible. The SUPREME Index also has practical strengths. Many existing readmission tools have been developed in hospital-specific, condition-specific, or older-adult cohorts, which can limit generalizability across settings and populations [5, 6, 12, 35, 36]. In contrast, this study focused on a community-based general population, which may broaden applicability for population health management initiatives [37, 38]. Moreover, the index requires only four variables, supporting operational simplicity and potential scalability in administrative data environments. However, this study has important limitations that should be considered when interpreting the findings and potential uses of the index. First, the analysis was restricted to residents of a single region, which may limit generalizability to other areas with different demographic or healthcare-utilization patterns. In addition, area-level characteristics that may influence readmission risk (e.g., proximity to acute-care or tertiary hospitals, local primary care supply, public transportation accessibility, and regional healthcare infrastructure) were not available for inclusion in the present claims-based model. Future studies should evaluate whether incorporating such contextual measures improves transportability and calibration across regions and supports more locally tailored risk stratification. Second, the index was internally validated but not externally validated in an independent dataset; therefore, its performance in other regions or time periods remains uncertain. Future studies should evaluate the SUPREME Index in external cohorts and consider recalibration as needed. Third, and most importantly, the predictors were not ascertained at a single uniform baseline time point. Because the primary analysis was conducted at the individual level, predictors were operationalized using information available in the database, including (i) variables defined at the time of the first hospitalization in the study period (e.g., admission type, in-hospital procedures, and length of stay for the first hospitalization) and (ii) variables summarized over the entire study period (e.g., total costs and counts of healthcare utilization). Health behaviors and screening-based clinical measures were obtained from the most recent available health screening record. Accordingly, predictor ascertainment did not correspond to a single baseline date for all variables, and some predictors may have been measured after the occurrence of a 30-day readmission event for certain individuals. This temporal heterogeneity limits a strict interpretation of the index as a forward-looking prediction model and supports framing the tool as a pragmatic stratification index derived from available administrative information. Future work should develop and validate time-anchored models using clearly defined baseline dates and predictor windows preceding an index event. Fourth, the outcome definition and aggregation strategy also merit consideration. The outcome was defined as experiencing at least one 30-day readmission event following any hospitalization discharge during the study period, and analyses were conducted at the person level. Therefore, individuals with multiple 30-day readmissions were treated equivalently to those with a single event, and the index was calibrated to a person-level outcome defined by the presence of any 30-day readmission rather than to an event-level outcome anchored to a specific index hospitalization. Although this approach was intended to reflect heterogeneous utilization patterns in a community-based population with relatively low admission rates, it targets a different estimand than conventional index-hospitalization–based 30-day readmission models. In addition, operating characteristics at the selected cutoff varied across age strata, suggesting that the preferred operating point may differ depending on the age distribution and the intended use case of the target population. Fifth, the sample size and the relatively small number of readmission cases may have contributed to outcome imbalance, which can affect model stability. Several utilization variables were dichotomized using the top quartile to maintain parsimony and reduce complexity; this choice may have reduced information compared with continuous or multi-level representations. In addition, certain conditions relevant to readmission, particularly psychiatric disorders, could not be fully assessed because some psychiatric diagnostic codes are masked in the claims data. Sixth, the reliance on utilization history (previous hospitalizations, length of stay, and costs) as primary predictors warrants caution regarding potential algorithmic bias. As highlighted in recent literature [39, 40], algorithms trained on healthcare utilization or cost data may systematically underestimate the health needs of vulnerable populations who face barriers to access, because lower observed utilization does not necessarily indicate lower morbidity. Consequently, if the SUPREME Index is used for risk stratification, it may preferentially identify individuals with more frequent prior contact with the healthcare system and may be less sensitive for high-need individuals whose utilization is low or incompletely captured in claims. Although the universal health coverage system in the study setting may mitigate such disparities compared with other healthcare environments, gradients in healthcare utilization can still persist. Therefore, the SUPREME Index should be interpreted as reflecting observed utilization and expenditure patterns in addition to underlying health status, and future studies should evaluate calibration and error patterns across relevant subgroups and consider incorporating additional measures of clinical severity and social determinants of health. Finally, the study used data from 2016 to 2019. Temporal changes in clinical practice, coding patterns, and healthcare delivery may affect the performance of the index in more recent cohorts. Where the tool is applied in later periods, periodic re-evaluation and recalibration may be required. In summary, the SUPREME Index provides a claims-based risk stratification tool for identifying community-dwelling adults at elevated risk of experiencing at least one all-cause 30-day readmission event during the study period. External validation in independent populations and evaluation of operational utility in routine care and population health management are warranted before broader implementation. Conclusions his study developed and internally validated the SUPREME Index, a parsimonious risk stratification index to identify individuals in a community-dwelling general population at increased risk of experiencing at least one all-cause 30-day readmission event during the study period. The index was derived by combining EN regularization for feature selection with LR to construct a point-based scoring index using four routinely available claims-based predictors. The SUPREME Index demonstrated good discrimination and calibration in internal validation, supporting its potential as a practical tool for risk classification using administrative claims data. Given the person-level outcome definition and the non-uniform timing of predictor ascertainment, the index should be interpreted as a risk stratification tool rather than an index-hospitalization–anchored event-level prediction model. Further research is warranted to externally validate the index in diverse populations and settings and to evaluate its utility for population health management and policy-relevant risk targeting. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2: Peer review Supplementary Material 2: Peer review