Machine learning for predicting institutionalization and mortality risks among older home care recipients with routinely collected need assessment data: explainable AI for long-term care Authors: Hongsoo Kim, Jimin Yoon, Hyerim Noh, Sunghun Yun, Seungyeon Chun, Woojoo Lee Published: 2026 ABSTRACT BackgroundIdentifying older home care recipients at risk of institutionalization in advance is crucial for providing preventive services. Supporting the decision-making of care workers in prioritizing their limited resources would be useful for promoting an aging-in-place policy.ObjectiveThis study aims to develop and interpret explainable machine learning (ML) methods for predicting institutionalization and mortality as competing events among older adults newly entering the Long-Term Care Insurance (LTCI) home care system.Materials and methodsWe explored and evaluated ML algorithms using routinely collected standardized functional assessment data to examine the care needs of 36,526 older beneficiaries enrolled in a nationwide public LTCI program. Individuals were classified according to whether institutionalization or death occurred first during follow-up. Accordingly, we employed multinomial logistic regression (MLR), least absolute shrinkage and selection operator (LASSO), extreme gradient boosting (XGBoost), and random forest (RF) algorithms to predict institutionalization and death in a multinomial approach. The performance of the model was assessed via the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and expected calibration error (ECE). Youden’s J statistic was used to determine the cutoff point. ResultsAmong newly enrolled home care beneficiaries, 10.4% were institutionalized in long-term care facilities within six months, increasing to 18.8% within 12 months after the initiation of long-term care services. The LASSO model achieved the best performance, with the most predictive features of institutionalization including eating nonfood items which marked the advanced dementia, need for suctioning (aspiration care) and dementia diagnosis. The most predictive features of mortality included need for oxygen therapy, cancer pain management and diabetic foot care.ConclusionThe newly developed algorithm may be implemented to examine the care needs of older home care users without additional risk assessment and for early intervention to reduce institutionalization risks.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12911-026-03641-8. Introduction Aging in place is a shared policy goal of countries with rapidly aging populations and limited human and financial resources to deliver extensive formal care services [1–4]. Older adults are also known to maintain their independence and privacy while staying in their own homes in familiar neighborhoods for as long as they can, even with formal home care support. Notably, a significant number of older adults with long-term care (LTC) needs are likely to be institutionalized, after which they find it difficult to return to their communities with home care support [2, 5–7]. In addition, a significant proportion of patients experience early mortality despite receiving home-based services [8, 9]. From a policy perspective, both outcomes carry heavy social and fiscal costs: institutionalization increases public expenditure. Conversely, premature death signals unmet health or social needs and raises ethical concerns regarding the quality of care. Therefore, health systems must identify which home care users are at imminent risk of institutionalization or death so that limited resources can be prioritized for timely, preventive interventions. Various policies and services exist to prevent these outcomes [10, 11]. However, because the conditions of older home care users are quite complex, it is challenging for care providers to assess their institutional and mortality risks with sufficient time to intervene in fast-paced, real-world practical situations with limited resources. Studies have reported factors related to institutionalization risks for home care recipients and interventions that effectively prevent and delay institutionalization and death [5, 12–21]. Such evidence should be integrated into the day-to-day direct care setting and should be able to guide and support care workers with a wide range of knowledge, experience, and expertise, decreasing unnecessary variations and advancing person-centered care. Notably, compared with acute hospitalizations of relatively healthy community-dwelling older populations, only a few studies have focused on developing prediction models of the institutionalization or mortality risk of older adults with home care services [20–22]. The lack of such studies is partly due to the lack of relevant, standardized qualitative and quantitative data for developing such models in high-income countries with advanced LTC systems, as well as the lack of such formal home care services in lower- and middle-income countries. In 2008, the Republic of Korea (hereafter Korea), a country with one of the most rapidly aging populations, implemented a nationwide, tax- and contribution-financed public Long-Term Care Insurance (LTCI) program operated by the National Health Insurance Service (NHIS). Eligibility to apply is universal to all residents aged 65 years or older, as well as to residents under 65 with designated geriatric conditions (dementia, cerebrovascular disease, Parkinson’s disease, and related disorders); applications are self- or family-initiated without prior clinical referral or means testing. LTCI benefits are delivered mostly as in-kind services. The home care benefit package comprises five service types: (a) home visit care—assistance with activities of daily living, instrumental activities of daily living, and related personal support delivered by certified personal carers, which is the dominant service in terms of utilization; (b) home visit bathing—specialized bathing services at the recipient’s home; (c) home visit nursing—registered-nurse care (e.g., wound care, tube management, medication administration) provided on a physician’s order; (d) day and night care—center-based day programming that includes supervision and light rehabilitation; and (e) short-term respite care—temporary institutional stays up to nine days per month. Among older adults newly admitted to LTCI home care in Korea, the risk of transitioning to institutional care is substantial, making risk stratification at the LTCI entry point a central operational need for case managers and care planners. A nationwide, standardized functional assessment system was also implemented to assess the care needs of its citizens, with approximately 54 items routinely collected by trained assessors at the applicant’s home at the time of application. On the basis of this assessment together with a physician’s medical opinion on the applicant’s functional status, a Grading Committee determines each applicant’s care grade (grades 1–5, with 1 indicating the highest need for care, or cognitive support grade) or denies eligibility; only those receiving a care grade may use LTCI-financed services. For those granted a care grade, recertification assessments are conducted periodically—typically every one to two years depending on the assigned grade—and ad-hoc reassessments may be triggered by a substantial change in functional status. Assessment data are collected and automatically stored in the centralized information system of the NHIS, the single public insurer of the public LTCI in Korea [23]. By taking advantage of such routinely collected data, this study aims to develop and interpret explainable machine learning models for predicting institutionalization and death as competing events among older adults newly entering the LTCI home care system. The prediction models developed can be used to design decision-support tools for administrators and clinicians, enabling them to intervene within sufficient time to prevent the institutionalization of older home care beneficiaries under the public LTCI by providing personalized home care packages on the basis of the identified risks of each individual. Materials and methods Study design and data This observational study leverages a secondary analysis of the deidentified administrative data of the public LTCI offered by NHIS for research purposes in Korea. In this study, we used the LTCI administrative database [23, 24] merged with three variables from the NHIS database (medical history of cardiovascular disease, dementia, and hemiplegia) as predictor variables for the years 2015–2022 [25–27]. The LTCI administrative databases include various databases, including the care needs assessment (CNA) database, which contains standardized CNA items routinely collected from older people who take the public LTCI eligibility test. The CNA database includes data on the following six categories: functional impairments of basic activities of daily living (ADL) and instrumental ADL (IADL), cognitive impairments, problem behaviors, nursing care, and rehabilitation [28–32]. In terms of basic and instrumental activities of daily living, the CNA assesses whether the applicant required partial or full assistance with 23 tasks (13 basic and 10 instrumental) in the past few months. The CNA also assesses the cognitive impairments of the applicants by checking for the presence of 10 symptoms (e.g., disorientation to time or place, memory loss, and impaired judgment) in the past month. In the problem behavior section, 22 items are used to determine whether the applicant has shown specific symptoms in the past month. The nursing care section evaluates whether the applicant needed any of 10 nursing care activities in the past two weeks. Finally, the rehabilitation section considers impairments or restrictions in 10 areas for the applicant. In addition, the CNA database includes information on informal caregiving, such as primary caregivers and the living arrangements of the subject. We included all functional and social needs-related information as predictor variables in our analysis. Given that it would be useful for public care planners of the NHIS if we identified meaningful but parsimonious prediction models from a minimal set of data at the first encounter of the public LTCI system, we attempted to restrict the use of NHIS claims-based medical records as much as possible. The inclusion and exclusion criteria for the subjects (cases) are subsequently outlined. We identified older adults who recently received CNA and who began using home care services under the LTCI program within three months after the assessment in 2015 (no prior history of using LTCI benefits). For each individual, predictor variables were derived exclusively from the initial (first-time) LTCI eligibility assessment. We selected the initial assessment because (i) it is the operational decision point at which NHIS case managers would apply a prospective risk-stratification tool—the intended decision-support context of the model—and (ii) it provides a baseline uncontaminated by prior service exposure, which itself modifies subsequent functional trajectories. For follow-up, we extracted their profile data and CNA data and linked the data with their health and LTC utilization data for up to one year for each patient. Figure 1 illustrates the selection process for the study sample. In 2015, 177,473 older LTCI beneficiaries were registered, 94,254 of whom were included in the analysis. These individuals had no prior history of nursing home admission, received home care services under the LTCI, or underwent a CNA within three months prior to utilization. These inclusion criteria were used to identify more precise risk features in a homogeneous population and to minimize misclassification errors. Among the 94,254 participants, we restricted the analytic sample to the 40,405 first-time users (excluding 53,849 renewal applicants), and we excluded 3,144 participants who were not living at home because home care services are typically provided to individuals residing in their own homes. Finally, we excluded 735 participants with missing values for the CNA items. This increased the size of the dataset to 36,526 older adults. This study was approved by the institutional review board (IRB No. E2201/001–001) for human subject research at the institution to which the authors are affiliated. Fig. 1Study flow chart Study flow chart Outcome variables and potential features Our primary focus was to derive prediction models for assessing the risk of institutionalization and death among older beneficiaries of the public LTCI in Korea. We primarily defined our outcomes on the basis of whether institutionalization in a long-term care facility (LTCF) and all causes of death that occurred within six months of the LTCI eligibility assessment. To account for the competing risk relationship between institutionalization and death, in which the occurrence of death precludes subsequent observation of institutionalization, we employed a multinomial approach. In this model, outcomes were categorized into three classes: no event (coded as 0), institutionalization (coded as 1), and death (coded as 2). All study participants were followed up continuously, starting from the date of their first home care service application in 2015. We selected a 6-month primary prediction horizon for two reasons. First, the risk of facility transition among older adults newly admitted to LTCI home care in Korea is substantial and concentrated in the first year of service use—in our cohort, 10.4% of first-time home care users were institutionalized within 6 months of service initiation and 18.8% within 12 months. Given that the LTCI pathway from application through in-home assessment, Grading Committee decision, and home care service initiation is typically completed within one to two months, early identification of high-risk individuals at the LTCI entry point—rather than after extended observation—is the operationally decision-relevant need, and a 6-month horizon captures this period during which entry-point risk stratification can inform preventive intervention, transitional care planning, and in-home service reinforcement. Second, the 6-month institutionalization event rate was sufficient for stable model training at our available sample size (n = 36,526) while remaining proximal enough to capture acute, decision-relevant risk. In addition to the primary 6-month horizon, we examined 3- and 12-month outcome windows, which span clinically distinct planning contexts—short-term transitional-care planning and longer-term program-level resource allocation. Because the outcome horizon defines the prediction target rather than being an incidental modelling choice, comparing models across horizons serves two complementary purposes: it tests whether a common core set of predictors is identified consistently across windows, and it characterizes how the relative prominence of these predictors shifts as the horizon lengthens. The potential features in this study were primarily limited to the CNA items, with the addition of demographics and medical history for cardiovascular disease, dementia, and hemiplegia using the NHIS claims dataset. The age variable was standardized via min‒max scaling to range between 0 and 1. We considered the medical history of the individuals up to one year before the date of application for home care services. Additional details of the timeline for the study participants are provided in Supplementary Fig. 1. Furthermore, based on the Charlson Comorbidity Index (CCI), 17 comorbidities (myocardial infarction, congestive heart failure, peripheral vascular disease, cerebrovascular disease, dementia, chronic pulmonary disease, rheumatic disease, peptic ulcer disease, mild liver disease, diabetes without chronic complication, diabetes with chronic complication, hemiplegia or paraplegia, renal disease, any malignancy, moderate or severe liver disease, metastatic solid tumor, and AIDS/HIV) were categorized into a scale of ‘0’ to ‘3 or more’ according to the individual comorbidity profiles and included in the analysis as predictor variables [33]. In total, 83 predictive features capturing the needs of the participants were extracted and included in the model. Model development and evaluation Figure 2 illustrates the model-building process. Before model development, the dataset was randomly divided into training (80%) and test (20%) datasets. We performed 5-fold cross-validation (CV) to select the optimal hyperparameters for each machine learning (ML) method in the training dataset. The test dataset was utilized only for model evaluation. This procedure guarantees that the results are not susceptible to overfitting [34]. Fig. 2Illustration of the model building process. Note: AUC, area under the receiver operating characteristic curve; ACC, accuracy; SEN, sensitivity; SPE, specificity; PPV, positive predictive value; NPV, negative predictive value; LASSO, least absolute shrinkage and selection operator; XGBoost, extreme gradient boosting Illustration of the model building process. Note: AUC, area under the receiver operating characteristic curve; ACC, accuracy; SEN, sensitivity; SPE, specificity; PPV, positive predictive value; NPV, negative predictive value; LASSO, least absolute shrinkage and selection operator; XGBoost, extreme gradient boosting In this study, we employed ML methods in multinomial approach such as multinomial logistic regression (MLR), least absolute shrinkage and selection operation (LASSO), extreme gradient boosting (XGBoost), and random forest (RF) classifiers to predict three outcomes: no event, institutionalization and death. Multinomial logistic regression is a traditional regression model, whereas LASSO is a penalized logistic regression that performs model fitting and variable selection simultaneously [35]. XGBoost and RF are ensemble methods that combine multiple weak learners to improve predictive performance. XGBoost and RF are more effective than linear regression models in the presence of complex interactions between variables. The details of the hyperparameter settings for each ML method are provided in Supplementary Table 1. We assessed the prediction performance in terms of the area under the receiver operating curve (AUC), accuracy, sensitivity, and specificity. The latter two quantities were calculated, leveraging Youden’s J statistic to determine the optimal cutoff point [36]. The model with the highest AUC value in the test dataset was primarily considered the best. Additionally, the positive predictive value (PPV) and negative predictive value (NPV) were calculated to evaluate the models’ specific performance, and the expected calibration error (ECE) was included to assess calibration reliability. Finally, because the multinomial approach utilizes a unified model to predict the three outcomes concurrently with shared hyperparameters, the performance metrics were calculated via a one-vs-rest approach. We interpreted the black-box model like XGBoost by employing SHapley Additive exPlanations (SHAP) values. Rooted in cooperative game theory, SHAP determines the contribution of each feature by calculating its Shapley value, which measures the average change in the expected prediction when a specific feature is included versus when it is excluded across all possible feature combinations. While the standard variable importance metric based on the permutation method yields only an overall ranking of feature importance, SHAP values enable the interpretation of each feature’s contribution to individual predictions [37]. We identified the 15 most important features with the highest average absolute SHAP values. All calculations were performed via R version 4.0.3 software (R Foundation for Statistical Computing, Vienna, Austria). We used the R packages ‘multinom,’ ‘glmnet,’ ‘ranger,’ and ‘XGBoost’ for model development, ‘cutpointr’ and ‘pROC’ for model evaluation, and ‘shapviz’ for model interpretation. Results Characteristics of the study population The descriptive profiles of the study population at the point of applying for LTC eligibility are summarized in Table 1. The total sample size was 36,526 individuals. The mean age was 81.0 years, and 69.8% of the participants were women. Regarding family caregiver, which reflects unpaid support, 5.0% of the total sample did not receive any family-based care, 29.1% received care from their spouse, and 65.9% were cared for by their children. Furthermore, 24.5% lived alone, and 75.5% lived with others. In the year prior to the initiation of home care services, 40.0% of the population had been diagnosed with cardiovascular disease, 49.9% had been diagnosed with dementia, and 6.0% had been diagnosed with hemiplegia. The baseline comorbidity status showed that 12.8% of the study population had no comorbidities (CCI = 0) and the proportions of those with CCI scores of 1, 2, and 3 or more were 16.6%, 17.5%, and 53.1%, respectively. For the primary categories of the outcome variables, the rates of institutionalization under LTCF were 5.7% for 3 months, 10.4% for 6 months, and 18.8% for 12 months. The rates of death were 4.6% at 3 months, 7.2% at 6 months, and 10.8% at 12 months. The descriptive statistics of CNA for our sample are presented in Supplementary Table 2. Table 1General characteristics of the study populationVariable n (%)Total36,526(100%) Explanatory Variables Demographics n (%) SexMale11,023(30.2%)Female25,503(69.8%)AgeMean (sd)81.0(6.9)LivingsituationWith others27,584(75.5%)Alone8,942(24.5%)FamilycaregiverNone1,817(5.0%)Spouse10,629(29.1%)Daughter or Son24,080(65.9%)Medical historyCardiovascular disease14,611(40%)Dementia18,221(49.9%)Hemiplegia2,201(6.0%)Charlson ComorbidityIndex (CCI)04,654(12.8%)16,062(16.6%)26,403(17.5%)3 or more19,407(53.1%) Outcome Variables n (%) No eventThree months32,775(89.7%)Six months30,104(82.4%)Twelve months25,741(70.4%)Institutionalization(LTCF admission)Three months2,072(5.7%)Six months3,801(10.4%)Twelve months6,851(18.8%)DeathThree months1,679(4.6%)Six months2,621(7.2%)Twelve months3,934(10.8%)Note: Age is presented as the mean and standard deviation (SD) Table 1General characteristics of the study populationVariable n (%)Total36,526(100%) Explanatory Variables Demographics n (%) SexMale11,023(30.2%)Female25,503(69.8%)AgeMean (sd)81.0(6.9)LivingsituationWith others27,584(75.5%)Alone8,942(24.5%)FamilycaregiverNone1,817(5.0%)Spouse10,629(29.1%)Daughter or Son24,080(65.9%)Medical historyCardiovascular disease14,611(40%)Dementia18,221(49.9%)Hemiplegia2,201(6.0%)Charlson ComorbidityIndex (CCI)04,654(12.8%)16,062(16.6%)26,403(17.5%)3 or more19,407(53.1%) Outcome Variables n (%) No eventThree months32,775(89.7%)Six months30,104(82.4%)Twelve months25,741(70.4%)Institutionalization(LTCF admission)Three months2,072(5.7%)Six months3,801(10.4%)Twelve months6,851(18.8%)DeathThree months1,679(4.6%)Six months2,621(7.2%)Twelve months3,934(10.8%)Note: Age is presented as the mean and standard deviation (SD) General characteristics of the study population Living situation Family caregiver Charlson Comorbidity Index (CCI) Institutionalization (LTCF admission) Note: Age is presented as the mean and standard deviation (SD) Comparison of model performance We primarily assessed the prediction performance of the models in terms of the AUC. Table 2 outlines the performances of the four ML methods for the three categories: no event (class 0), institutionalization (class 1), and death (class 2), considering competing risks. All performances were calculated on the test dataset via the four ML methods, with the optimal hyperparameters selected via 5-fold CV. Among the models predicting events within six months, the results indicated that the LASSO model was the best for all three classes, with an AUC of 0.671 for no event, 0.603 for institutionalization, and 0.766 for death. Multinomial logistic regression and XGBoost performed slightly worse than LASSO did, with AUCs of 0.669 in both models for class 0 (no event), 0.599 and 0.602 for class 1 (institutionalization), and 0.761 and 0.765 for class 2 (death), respectively. We also considered other evaluation metrics, including accuracy, sensitivity, specificity, PPV, NPV, and ECE. Therefore, comprehensively, for the class 0 (no event), LASSO achieved the highest specificity, and PPV, while MLR showed superior performance in accuracy, sensitivity, and NPV. For class 1 (institutionalization), LASSO achieved the highest performance in PPV. In other metrics, MLR showed the best sensitivity and NPV, while RF outperformed the other models in accuracy and specificity. Regarding class 2 (death), LASSO showed the best sensitivity and NPV, whereas MLR yielded the highest accuracy, specificity, and PPV. In terms of calibration, RF achieved the smallest ECE in all classes; however, the gap between the models was marginal. Therefore, for events occurring within six months, the LASSO model was selected as the final model. Table 2Comparison of the prediction performances on the test dataset for six-month outcomesModelSelected HyperparametersClassCutoff pointAUCACCSENSPEPPVNPVECELTCF admission or death within six monthsMLR00.7920.6690.7410.8070.4330.8690.3250.18810.0880.5990.4330.7510.3960.1280.9310.06620.1010.7610.8110.5690.8290.2040.9620.072LASSOlambda=0.000700.8330.6710.6480.6630.5790.8800.2690.18910.0960.6030.4990.6720.4790.1320.9250.06720.0800.7660.7560.6320.7650.1720.9640.074XGBoostnrounds = 200,colsample_bytree = 1,max_depth = 3,eta = 0.100.8260.6690.6740.7040.5360.8760.2800.19110.0970.6020.5070.6560.4890.1310.9240.06720.0880.7650.7770.6110.7890.1820.9630.073RFnum.trees = 25000.8050.6460.6700.7010.5250.8730.2740.18310.0680.5490.6230.4350.6450.1260.9070.06620.1210.6180.6350.5480.6420.1050.9490.064Note: Due to their small magnitudes, lambda values are rounded to 4 decimal places; remaining values are rounded to 3. AUC: area under the receiver operating characteristic curve; ACC: accuracy; SEN: sensitivity; SPE: specificity; PPV: positive predictive value; NPV: negative predictive value; ECE: expected calibration error; LTCF: long-term care facility; MLR: multinomial logistic regression; LASSO: least absolute shrinkage and selection operator; XGBoost: extreme gradient boosting; RF: random forest Table 2Comparison of the prediction performances on the test dataset for six-month outcomesModelSelected HyperparametersClassCutoff pointAUCACCSENSPEPPVNPVECELTCF admission or death within six monthsMLR00.7920.6690.7410.8070.4330.8690.3250.18810.0880.5990.4330.7510.3960.1280.9310.06620.1010.7610.8110.5690.8290.2040.9620.072LASSOlambda=0.000700.8330.6710.6480.6630.5790.8800.2690.18910.0960.6030.4990.6720.4790.1320.9250.06720.0800.7660.7560.6320.7650.1720.9640.074XGBoostnrounds = 200,colsample_bytree = 1,max_depth = 3,eta = 0.100.8260.6690.6740.7040.5360.8760.2800.19110.0970.6020.5070.6560.4890.1310.9240.06720.0880.7650.7770.6110.7890.1820.9630.073RFnum.trees = 25000.8050.6460.6700.7010.5250.8730.2740.18310.0680.5490.6230.4350.6450.1260.9070.06620.1210.6180.6350.5480.6420.1050.9490.064Note: Due to their small magnitudes, lambda values are rounded to 4 decimal places; remaining values are rounded to 3. AUC: area under the receiver operating characteristic curve; ACC: accuracy; SEN: sensitivity; SPE: specificity; PPV: positive predictive value; NPV: negative predictive value; ECE: expected calibration error; LTCF: long-term care facility; MLR: multinomial logistic regression; LASSO: least absolute shrinkage and selection operator; XGBoost: extreme gradient boosting; RF: random forest Comparison of the prediction performances on the test dataset for six-month outcomes lambda= 0.0007 nrounds = 200, colsample_bytree = 1, max_depth = 3, eta = 0.1 Note: Due to their small magnitudes, lambda values are rounded to 4 decimal places; remaining values are rounded to 3. AUC: area under the receiver operating characteristic curve; ACC: accuracy; SEN: sensitivity; SPE: specificity; PPV: positive predictive value; NPV: negative predictive value; ECE: expected calibration error; LTCF: long-term care facility; MLR: multinomial logistic regression; LASSO: least absolute shrinkage and selection operator; XGBoost: extreme gradient boosting; RF: random forest The prediction performances for events within three and twelve months are presented in Supplementary Table 3. For the additional prediction task, regarding class 0 (no event), the XGBoost model achieved the highest AUC for the three-month prediction (0.697), while the LASSO model performed best for the twelve-month prediction (0.666). For class 1 (institutionalization), the highest AUC was achieved by the XGBoost model at three months (0.624) and by the LASSO model at twelve months (0.598). Finally, for class 2 (death), the LASSO model showed the best performance for the three-month prediction (0.794), whereas the XGBoost model showed the highest AUC for the twelve-month prediction (0.766). Overall, however, the predictive performances of the three models—MLR, LASSO, and XGBoost—were largely comparable, with only marginal differences observed. Interpretation of models LASSO, the final model, provided a straightforward interpretation to examine coefficients. Figure 3 presents the 15 variables with the largest absolute LASSO coefficients for the 6-month outcome, shown separately for each outcome category (no event, institutionalization, and death); the corresponding plots for the 3- and 12-month outcomes are provided in Supplementary Figs. 2 and 3. Each coefficient represents the contribution of a feature, with positive values indicating an increased probability and negative values indicating a decreased probability of each specific event. Fig. 3Top 15 features selected by the optimal LASSO algorithm for the events within six months according to the absolute magnitude of coefficient. LASSO: least absolute shrinkage and selection operator Top 15 features selected by the optimal LASSO algorithm for the events within six months according to the absolute magnitude of coefficient. LASSO: least absolute shrinkage and selection operator For institutionalization at the 6-month horizon, the largest positive coefficient was that of the behavioral-change item “eating nonfood items,” a marker of advanced dementia. Dementia diagnosis and a cluster of dementia-related behavioral and cognitive items—delusion, wandering/restlessness (being fidgety and nervous), shouting, and time disorientation—were also selected with positive coefficients, as was the need for suctioning (aspiration care). Conversely, several skilled nursing-care needs, such as dialysis and diabetic foot care, carried negative coefficients, indicating a lower relative probability of institutionalization - plausibly because medically complex cases of this kind more often follow a mortality trajectory than transitioning to a facility. The 3- and 12-month models (Supplementary Figs. 2 and 3) identified the same core set of institutionalization predictors—“eating nonfood items,” dementia, the dementia-related behavioral and cognitive cluster, and the need for suctioning (aspiration care) - with consistent sign and comparable magnitude, indicating that this set is identified independently of the outcome window rather than being specific to the 6-month horizon. The relative prominence of these predictors, however, shifted systematically with horizon length: at 3 months the behavioral and cognitive markers dominated and the need for suctioning (aspiration care) contributed only a minor coefficient, whereas at 6 and 12 months the aspiration care coefficient rose to become the second largest, after “eating nonfood items.” Because each horizon defines a distinct prediction target, this gradient is clinically coherent rather than inconsistent: imminent (3-month) institutionalization was most strongly associated with acute behavioral and cognitive disturbance, while markers of advanced physical deterioration gained relative weight as the horizon lengthened. For mortality, the predictor structure was distinct from that of institutionalization, yet likewise stable across all three horizons. Skilled-nursing-care needs were dominant: the need for oxygen therapy carried the largest coefficient at every horizon, followed by cancer pain nursing care, with the needs for dialysis and pressure ulcer care also selected. Female sex and dementia diagnosis were associated with a lower relative probability of death. This pattern is consistent with mortality being driven primarily by serious, medically complex conditions captured through skilled-nursing-care needs. The SHAP results for XGBoost, which demonstrated the second-highest performance, are presented in Supplementary Fig. 4. Additionally, odds ratios were calculated from the LASSO model coefficients and are reported by each category in Supplementary Tables 4, 5, and 6. Discussion This study aimed to develop an ML-based predictive model to identify older adults at high risk of institutionalization and death during their initial eligibility assessment for Korea’s public LTCI. By utilizing routinely collected, standardized CNA data, we sought to identify a model that is implementable within existing administrative systems and interpretable by frontline care planners and policymakers. Among several tested models, LASSO demonstrated the best performance in predicting institutionalization and mortality risk, achieving an AUC of 0.603 and 0.766, respectively. For institutionalization, the most influential predictors were dementia diagnosis and a set of advanced-dementia behavioral and cognitive markers—most prominently the assessment item “eating nonfood items,” together with delusion, wandering/restlessness (being fidgety and nervous), shouting, and time disorientation—as well as the need for suctioning (aspiration care). These factors align with established clinical knowledge and prior research. Consistent with prior studies - including Gaugler et al. [5] and Han et al. [17] - dementia diagnosis and dementia-related behavioral and cognitive symptoms emerged as dominant predictors of institutionalization in our model. The advanced-dementia behavioral phenotype was particularly prominent: the Korean LTCI behavioral-change item “eating nonfood items,” which operationalizes late-stage dementia behavior, carried by far the largest coefficient, alongside delusion, wandering/restlessness (being fidgety and nervous), shouting, and time disorientation. This pattern is consistent with the established view that dementia-related behavioral and psychological symptoms, particularly when severe, exceed caregiver coping capacity and precipitate facility placement [12, 15]. A finding less commonly emphasized in prior nursing home admission literature is the prominence of the need for suctioning (aspiration care), a marker of advanced physical and swallowing deterioration, whose contribution strengthens at the 6- and 12-month horizons. Taken together, these results indicate that, within Korea’s LTCI applicant population, the combination of an advanced-dementia behavioral phenotype and progressive physical-functional decline constitutes the dominant risk profile for institutionalization. The 6-month horizon served as our primary analysis, with the 3- and 12-month models examined to characterize institutionalization risk across clinically distinct planning contexts - short-term transitional-care planning and longer-term program-level resource allocation. Because the outcome horizon defines the prediction target rather than being an incidental modelling choice, the multi-horizon comparison should be read on two dimensions. First, the same core set of predictors for institutionalization - dementia, the advanced-dementia behavioral and cognitive cluster, and the need for suctioning (aspiration care) - was selected with consistent sign and comparable magnitude at all three horizons (Fig. 3(b); Supplementary Fig. 2(b) and 3(b)), indicating that the core predictor set is a property of the applicant population rather than an artifact of any single window, and supporting its use for early risk stratification at the point of LTCI entry. Second, the relative prominence of these predictors shifted systematically with horizon length: acute behavioral and cognitive disturbances were the dominant correlates of imminent (3-month) institutionalization, whereas the need for suctioning (aspiration care) - a marker of progressive physical and swallowing deterioration - rose to become the second-largest coefficient at 6 and 12 months. Rather than signaling instability, this gradient is itself clinically informative, suggesting that the most actionable intervention targets differ by planning horizon: behavioral and caregiver-support needs for imminent-risk cases, and management of progressive functional and medical decline for medium- to longer-term risk. Our study has several distinct strengths apart from existing research, particularly in terms of its contextual, methodological, and practical implementation perspectives. First, previous studies have predominantly focused on Western contexts, such as Nordic countries or Finland (e.g., Sørbye et al. [38]; Nuutinen et al. [39]), their findings may have limited generalizability to East Asian populations due to profound cultural, familial, and systemic institutional discrepancies. By focusing specifically on the Korean geriatric population, this study addresses a crucial geographic and socio-cultural gap, examining context-specific characteristics. Second, our work provides substantial methodological advantages over prior studies. While they either relied on traditional linear statistical methods or were constrained by relatively small sample sizes, our study leverages a large, highly representative dataset sourced from the NHIS, which encompasses approximately 97% of the entire Korean population. Furthermore, unlike previous models that often simplified the outcome into a binary event, our approach integrates a multinomial approach that explicitly accounts for competing risks combined with machine learning algorithms like XGBoost and RF. This large-scale data environment is exceptionally well-suited for machine learning algorithms, ensuring robust predictive power, minimal overfitting, and superior generalizability. Finally, the clinical and administrative value of our model is highly synchronized with the specific implementation context of Korea’s LTCI system. Currently, the uniform nationwide assessment issued at the LTCI entry point often fails to consistently differentiate care goals by grade or fully encapsulate individuals’ heterogeneous functional needs [40], leading to substantial friction between recommended plans and actual service utilization [41]. Against this documented gap, our model introduces an individual-level, multi-horizon risk stratification tool anchored right at the system entry point. This allows for care-plan tailoring tied to specific risk drivers, mitigates unwarranted variations in service-intensity recommendations, and prioritizes proactive interventions within the concentrated early-transition window. The implications of this study are threefold from the health system and policy perspectives. First, the model demonstrates the feasibility of risk stratification by existing data infrastructure without requiring new data collection efforts. Given the rapidly increasing number of LTC applicants in aging societies such as Korea, scalable and low-cost decision-support tools are essential to assist care managers in allocating limited home care resources more effectively. Our cohort findings underscore why entry-point stratification is policy-relevant: among first-time LTCI home care users, 10.4% were institutionalized within 6 months of service initiation, rising to 18.8% within 12 months—that is, the rate of facility transitions were concentrated in the first half, with the cumulative rate nearly doubling by 12 months. This concentrated high-transition window places the applicants identified as high-risk by our model squarely within the priority subgroup for targeted preventive support, transitional care planning, and in-home service reinforcement at the point of initial LTCI certification. Second, because our model is anchored in routinely collected data at the initial assessment stage, it can serve as a transparent and explainable complement to the clinical judgments of caseworkers, facilitating the reduction of unwarranted variations in care decisions across regions and provider types. Third, such tools could support central or local governments in identifying high-risk subpopulations for targeted interventions—such as enhanced care coordination, caregiver support programs, and transitional care planning—to prevent avoidable institutionalization and promote aging in place. The operational value of risk stratification is particularly salient given documented limitations in Korea’s current care-planning process. The Standard Long-Term Care Utilization Plan issued by NHIS at the conclusion of the eligibility assessment is intended to translate each beneficiary’s care needs into recommended service types and intensity. However, prior research has shown that the current plan does not consistently differentiate care goals by care grade and does not fully reflect the individual beneficiary’s assessed functional needs, and that substantial disagreement exists between the insurer’s recommendation and beneficiaries’ actual service choice. Against this backdrop, the individual-level risk score produced by our model—together with the SHAP-identified features contributing to that risk—can contribute to improving the care-planning process at the LTCI entry point in three ways. First, it enables individualized tailoring of the care plan to each applicant’s specific risk drivers: a beneficiary whose elevated institutionalization risk is driven primarily by dementia and behavioral symptoms would be directed toward different service-mix recommendations (e.g., dementia-specific day programming, family caregiver respite) than a beneficiary whose elevated risk is driven by functional impairment and living-alone status (e.g., intensified home visit care, transitional care planning). The current Standard Care Plan does not systematically make such distinctions. Second, surfacing transparent risk drivers alongside the plan provides a standardized reference against which individual plan decisions can be discussed and benchmarked, reducing unwarranted variation in post-assessment service intensity across regions, providers, and caseworker judgment for similarly risk-profiled applicants. Third, by identifying applicants within the concentrated early-transition window (10.4% at 6 months, 18.8% at 12 months described above) as priority cases for proactive planning, the model can direct limited case-management attention toward the subgroup most likely to benefit from early intervention. Despite these advantages, several limitations of this study must be acknowledged. A key methodological decision was to restrict the analytical sample to newly certified beneficiaries who had not previously used home care services. This decision was based on both empirical and interpretive considerations. Exploratory analyses revealed substantial heterogeneity between new and renewing users, particularly regarding the influence of prior LTC utilization history on future institutionalization risk. For renewal users, variables capturing previous institutional care experience (e.g., nursing home or hospital stays) tended to dominate the prediction model, thereby reducing the explanatory power and clinical meaningfulness of other features, such as functional impairments or cognitive symptoms. Moreover, although ML techniques enable flexible modeling across heterogeneous populations, our objective was not only to maximize predictive accuracy but also to ensure that the model outputs remain interpretable and actionable. The inclusion of both groups in a unified model yielded SHAP-based explanations that were less intuitive and more challenging to translate into policy or care planning decisions. In contrast, the current model, which focuses exclusively on new users, provides clearer insights into how current needs and conditions translate into short- to medium-term risks of institutionalization. This is particularly relevant for public insurers and local governments, who must develop individualized care plans within a limited timeframe after the initial assessment. Notably, our model did not account for heterogeneity in post-assessment care trajectories. Even among individuals with similar assessed needs, actual service use patterns can differ significantly, depending on caregiver support, personal preferences, and regional service supply. Renewal users may reflect more complex dynamics, such as long-term caregiver burnout or fluctuating care goals, which could lead to bias in risk estimates if pooled with first-time users. Therefore, we concluded that a separate, tailored model for new beneficiaries would yield more policy-relevant insights. Future work should explore the development of a complementary model for renewal users, possibly incorporating variables such as prior service trajectories, changes in care need scores, or caregiver continuity, as these may yield distinct policy levers for managing chronic care transitions. A further scope-related consideration concerns the target population to which the models apply. Although LTCI’s application pathway is universal to Koreans aged 65 and older (and to those under 65 with designated geriatric conditions), actual applicants constitute a self-selected subset with at least some perceived care need. Our findings therefore do not generalize beyond this applicant population and should not be used to infer institutionalization or mortality risk among community-dwelling older adults who have not entered the LTCI system, nor among older adults who rely on private or alternative long-term care arrangements outside the LTCI track. Conclusion Despite these limitations, this study demonstrates the potential of integrating explainable ML tools into the core assessment process of national LTC systems. By leveraging existing data and focusing on first-time home care users, the model provides actionable insights that can support early interventions, prioritize scarce resources, and promote the broader policy goal of aging in place. As aging societies face growing care demands and workforce shortages, such tools can serve as key components in building proactive, data-driven long-term care systems. Future research should expand this work by validating the model in other cohorts, piloting its implementation in real-world settings, and developing tailored tools for other key subgroups, including renewing users and high-cost care recipients. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1