A preliminary study of a lifetime long-term care costs estimation model based on changes in care level: implications for sustainable long-term care in Japan

Abstract Background Japan faces an unprecedented demographic shift, characterized by the world’s most rapidly aging population and a projected surge in annual deaths, leading to a “frequent death society.” This trend places substantial fiscal pressure on national healthcare and long-term care (LTC) systems, with expenditures already representing a significant share of gross domestic product (GDP) and continuing to rise. To support sustainability, accurate and proactive cost-prediction models are needed for resource allocation and policy planning. Japan’s Long-Term Care Insurance (LTCI) system, established in 2000, provides services based on a seven-level care needs certification, which directly determines monthly benefit limits and strongly influences overall LTC expenditures. Ongoing revisions to the certification system underscore the need to understand how changes in care levels relate to future costs. Traditional cost-prediction models often rely on static, short-term aggregates and may miss dynamic spending patterns. In contrast, data-driven approaches (e.g., trajectory-based methods and machine learning) can identify evolving patterns over longer periods and leverage routinely collected data to enable earlier risk stratification and targeted interventions. Objectives This preliminary study addresses a specific research gap by uniquely focusing on estimating lifetime LTC costs based on "changes in care levels," utilizing only initial (first three months) service utilization data and associated costs, without requiring extensive patient background information. Although we refer to “lifetime cost estimation,” the present analysis is based on observed service utilization and expenditures over a 12–36-month observation window; therefore, findings should be interpreted as an estimation of longer-term cost trajectories rather than directly observed lifetime costs. Methods We analyzed data from 5,925 LTC users who initiated services at one of 91 home care service centers operated nationwide by a single company in Japan in June 2015 or later, continued service use for 12–36 months, and were certified at Care Levels 1–4. The provider is a privately held (non-listed) corporation; therefore, publicly available audited financial statements and dividend policies are limited. As supplementary context, we referenced publicly available Official Gazette (Kanpo)-derived corporate information (Kanpo-derived database; CATR) [1]. The outcome was monthly average LTC service cost. Predictors included initial care level, first-month costs, binary indicators for seven LTC service types used in the first month, binary indicators for changes in costs for each service type during the first three months, and interruption of LTC service use during the first three months. We constructed prediction models using random forest and multivariable linear regression, with an 80/20 split for training/validation. For cost comparisons, users were categorized into a Maintenance/Improvement Group (final care level unchanged or improved from baseline) and a Deterioration Group (final care level worsened from baseline). Results The Deterioration Group showed significantly higher costs from the first month, particularly among users with higher independence (Care Level 1 or 2), which may reflect early anticipation of deterioration by care providers. Predictive performance was high for both random forest (R2 = 0.677 in the preliminary study) and linear regression models. The linear regression model performed best primarily in the stable Care Level 1 Maintenance Group, whereas the random forest model performed better across most other cohorts, particularly at higher Care Levels (3 and 4). High predictive accuracy was achieved without requiring basic patient attributes (e.g., age, sex) or underlying disease information. In contrast, predictive performance was relatively low in the Care Level 1 Deterioration Group, suggesting greater heterogeneity in cost trajectories among users who are mild at baseline but subsequently deteriorate. Conclusions This preliminary study demonstrates the feasibility of estimating longer-term LTC cost trajectories based on early service utilization patterns, highlighting the potential role of care managers in shaping future cost trajectories. These findings may inform efforts to enhance the fiscal sustainability and quality of Japan’s LTCI system.