Sage Journals: Discover world-class research

Abstract

Objective: With the growing burden of type 2 diabetes and its associated healthcare costs, the factors influencing future expenditures, particularly among long-term care insurance (LTCI) users, must be identified. Few studies have addressed the prediction of multiple cost domains, including medical, LTC, and dental expenditures. This study predicted medical, dental, and LTC costs in the following year for patients with type 2 diabetes and identified key predictors based on health information from the previous year. Methods: We applied three machine learning models—random forest, boosted trees, and neural networks—to LTCI users’ data in Japan and incorporated prior-year healthcare costs, service usage patterns, and diabetes status. Results: In the 2019 medical cost model, boosted trees showed the best performance for those aged 74 or younger (R² = 0.46, RMSE = 151,804 JPY). LTC costs were influenced by prior LTC spending (∼40%) and facility service use (30–50%), while dental costs were predicted by prior dental expenditures. Conclusions: Prior-year medical costs strongly influenced later medical expenditures, while LTC costs reflected prior LTC spending and facility use. These quantified relationships provide insights for healthcare cost optimization and support policymakers in designing preventive strategies and care systems for aging populations with chronic diseases.

Keywords

medical costs dental care costs long-term care costs machine learning diabetes

Background

In 1961, a universal health insurance system was adopted in Japan, and since then, national healthcare spending has continued to rise.¹ The same is true for long-term care (LTC) expenses. Japan’s cumulative LTC costs in 2023 were 11 trillion and 500 billion yen, up from 320 billion yen in the previous year (a growth rate of 2.9%). The LTC insurance (LTCI) system was launched in 2000, and the total value of LTC costs has continued to grow since.² Lifestyle diseases have been a major driver of increased medical spending in recent years. Medical costs, especially for cardiovascular diseases and diabetes, have been rising partly because of increasing longevity. Lifestyle diseases account for approximately 30%–40% of Japan’s national healthcare expenses. Of Japan’s national healthcare costs of around 42 trillion yen in 2020, approximately 12–15 trillion yen were spent on lifestyle diseases.^3–5 Patients with type 2 diabetes often require extensive medical, dental, and LTC services, particularly when comorbidities accumulate later in life.

Predicting future medical costs is essential for extending healthy life expectancy, curbing healthcare spending, adjusting insurance premiums, and improving social security systems. In Japan, Inoue et al.⁶ estimated significant cost reductions by promoting generics, and Sase et al.⁷ used Bayesian networks to identify patients with high-cost type 2 diabetes. Globally, diabetes poses a major burden, with 537 million cases in 2021 projected to rise to 783 million cases by 2045, primarily in Asia, the Middle East, and Africa. Related healthcare costs are expected to increase from $966 billion in 2021 to $1.5 trillion by 2045.⁸ David et al.⁹ estimated lifestyle disease-related losses to be $3.5 trillion in Japan between 2010 and 2030. Studies have also shown that socioeconomic factors such as income and education influence medical spending.¹⁰

Recent machine learning (ML)-based studies can be broadly categorized into three groups: (1) those using insurance claims data to identify high-cost patients,^11–13 (2) those comparing algorithmic performance across models,^12,14 and (3) those leveraging electronic medical records or biometric data to predict cost patterns.^15,16 Alexandre et al.¹⁷ compared model performance using national health data in France, while Kurosaki et al.¹⁸ suggested the utility of Japan’s National Database data to address regional disparities.

Long-term cost projections have also been addressed. Getzen¹⁹ predicted a 5% annual increase by 2034. Other studies examined the impact of dental care on medical and LTC costs,^20–22 exploring how factors such as multimorbidity, past expenditures, and self-rated health status affect future costs.^23–25 Mori et al.²⁶ linked chronic disease complexity in those older than 75 to higher combined medical and LTC expenses. However, no studies have jointly analyzed future changes in medical, dental, and LTC costs at the individual level.

Identifying the key factors associated with rising healthcare costs is essential for developing effective preventive strategies and improving resource allocation under the LTCI system. Although many studies have aimed to predict future healthcare costs, few have explored the specific past-year health indicators and expenditure patterns that are most strongly related to subsequent changes in medical, dental, and LTC costs. Despite the existing literature on healthcare cost prediction, few studies have investigated the combined effects of prior-year medical, dental, and LTC costs on future expenditures in Japan’s LTCI population, highlighting a substantial research gap. Therefore, this study examined the relationship between prior-year health indicators and subsequent changes in medical, dental, and LTC costs among older adults with type 2 diabetes under Japan’s LTCI system. Instead of focusing on cost prediction, our objective was to identify the key factors associated with cost increases using ML models as an exploratory tool for variable importance analysis.

Methods

We used the datasets for Iwamizawa City in Hokkaido comprising LTC claims, including dental care expenditures (April 2017–March 2020), National Health Insurance medical claims (April 2017–March 2020), and late-stage older adult medical claims (April 2017–March 2020), as well as a matching table of the number of insured individuals and patient codes shown in the LTC claims (April 2017–March 2020). For the name of the disease or injury, we extracted data encompassing type 2 diabetes (including complications such as diabetic nephropathy) from the National Health Insurance claims and late-stage older adult medical claims. Furthermore, we linked the “presence or absence of type 2 diabetes” with LTC claims that include dental care costs. Table 1 presents the data items of the medical and LTC claims that were used, and Table 2 displays details of the datasets for patients with type 2 diabetes, including the dental and LTC costs. Explanatory variables were selected based on the literature and the specific objectives of this study. Many studies have used age, sex, past medical expenditures, and chronic disease indicators such as diabetes status as key predictors of future healthcare costs.^12–18 Accordingly, we adopted these core variables. As this study aimed to predict costs across three domains—medical, LTC, and dental services—additional variables related to LTC and dental service usage were included to better capture domain-specific characteristics and improve predictive performance. The complete list of explanatory variables is presented in Table 1. Figure 1 presents the ID linkages between the databases in graphic form. Owing to data limitations, certain variables, such as medication use, detailed comorbidities, and socioeconomic factors, were not available in our dataset.

Table 1.

Data items of medical and LTC claims.

NHI claims and medical claims data	LTC data
Patients ID	Health insurance ID
Health insurance ID	LTC insurance ID
Age	Medical costs
Sex	LTC costs
Type 2 diabetes	Dental care costs
Medical costs	Months of facility service usage
	Number of months using in-home service
	Level of care

Table 2.

Information on the datasets of type 2 diabetes patients (including dental and LTC costs).

Data set of type 2 diabetes patients including dental and LTC costs	Under 75 years old		Over 75 years old		Total (people)
	With diabetes	Without diabetes	With diabetes	Without diabetes	Total (people)
FY2017	54	476	848	2,949	4,327
FY2018	68	509	990	3,545	5,112
FY2019	71	563	1,155	4,227	6,016

Figure 1.

ID linkage between the databases

We used data from individuals who had been diagnosed with type 2 diabetes at medical institutions. The diagnoses were made by physicians based on standard clinical criteria, including fasting blood glucose levels; oral glucose tolerance test results; and the presence of associated conditions such as obesity, hypertension, and dyslipidemia. However, we did not include detailed health status information such as diabetes severity or specific comorbidities in our analysis as this would require linkage to additional clinical datasets that were not available in the current data structure. Regarding data preprocessing, we addressed missing values and extreme outliers (e.g., unusually high or low medical expenditures) by excluding these records from the analysis.

Subsequently, from the datasets of patients with type 2 diabetes (including dental and LTC costs), we divided the target patients into those aged 74 or younger and those aged 75 or older. We used multiple ML models to create a one-year model for predicting medical costs, a one-year model to forecast dental care costs, and a one-year model to determine LTC costs. We then calculated the degree of importance of the variables for which R² and the root mean square error (RMSE) provided the most outstanding models. We created two types of models: one to estimate medical, dental, and LTC costs from 2017 to 2018 and one to establish medical, dental, and LTC costs from 2018 to 2019. Tables 3 and 4 show the responses and explanatory variables for the target data.

Table 3.

A model that predicts various costs from 2017 to 2018.

Response variable	FY2018 medical costs FY2018 LTC costs FY2018 dental care costs
Explanatory variable	FY2017 medical costsFY2017 LTC costsFY2017 dental care costs
Explanatory variable	Use of home care servicesUse of LTC facilitiesSex, Age, Care Level, Type 2 Diabetes Status.

Table 4.

A model that predicts individual costs from 2018 to 2019.

Response variable	FY2019 medical costs FY2019 LTC costs FY2019 dental care costs
Explanatory variable	FY2018 medical costsFY2018 LTC costsFY2018 dental care costs
Explanatory variable	Use of home care servicesUse of LTC facilitiesSex, Age, Care Level, Type 2 Diabetes Status.

Three types of ML models were used: random forest (RF), boosting trees (BT), and neural network (NN). We selected RF, BT, and NN as representative ML models because of their proven ability to capture nonlinear relationships and variable interactions, which are common in healthcare cost data. Similar model types have been successfully applied in previous studies of medical expenditure prediction. We selected RF for its robustness and interpretability through variable importance measures, BT for its high predictive accuracy and computational efficiency, and NN for its ability to model complex nonlinear relationships. These models have demonstrated strong performance in medical cost prediction research.^12–14 In addition, we conducted a preliminary analysis using linear regression to provide a baseline comparison. However, its performance was significantly lower than that of the nonlinear models; therefore, it was excluded from the final model comparison. For model evaluation, R² and RMSE were chosen following standard practices in regression-based healthcare cost modeling.^13,16 R² measures the proportion of variance explained by the model, whereas RMSE provides a direct assessment of the average prediction error. Both methods are widely used in healthcare expenditure prediction studies,^11,17 allowing for intuitive comparisons across models.

We compared the R² and RMSE and calculated the primary effect of the degree of importance of the variables for which R² and the RMSE provided the most outstanding models. Tables 5–7 present the parameters for each ML model. The statistical software used for these analyses was JMP Pro 17.0.0. In this study, hyperparameter tuning was performed manually using JMP Pro by testing multiple combinations of model settings (e.g., number of trees, learning rate, and maximum depth). The best-performing set of parameters was selected based on the model performance within the training dataset.

Table 5.

RF’s parameters.

Parameter	Value
Validation data ratio	0.2
Number of rows (Total)	505
Number of features	9
Number of Trees in the forest	100
Number of features per split	7
Bootstrap sampling rate	1
Minimum number of splits per Tree	10
Maximum number of splits per Tree	2,000
Minimum leaf size	5
Training rows	403
Validation rows	102

Table 6.

BT’s parameters.

Parameter	Value
Validation data ratio	0.2
Number of layers	30
Number of splits per Tree	5
Learning rate	0.084
Minimum leaf size	5
Training rows	397
Validation rows	108

Table 7.

NN’s parameters.

Parameter	Value
Validation data ratio	0.2
Number of nodes in the first layer	3
Transformation of explanatory variables	Yes

Results

Table 8 compares the accuracy of the 2018 and 2019 prediction models. For the 2018 prediction models, NN excelled in both the 74-or-younger and 75-or-older models for predicting medical costs. In the predictions of LTC costs, RF excelled in the 74-or-younger model, and NN excelled in the 75-or-older model. In the predictions for dental care costs, BT excelled in the 74-or-younger model, whereas RF excelled in the 75-or-older model.

Table 8.

Comparison of accuracy of the 2018 and 2019 prediction model.

		FY2018 medical costs forecast				FY2019 medical costs forecast
		Under75		Over75		Under74		Over75
		Training	Validation	Training	Validation	Training	Validation	Training	Validation
RF	R²	0.67	0.37	0.63	0.18	0.70	0.38	0.61	0.25
RF	RMSE	98,622.56	198,298.89	82,817.53	128,560.24	127,090.67	162,398.82	95,016.36	132,991.35
BT	R²	0.57	0.03	0.47	0.11	0.65	0.46	0.45	0.25
BT	RMSE	130,007.15	148,058.40	98,957.52	133,559.51	137,550.01	151,804.35	113,729.30	133,069.21
NN	R²	0.31	0.37	0.19	0.21	0.59	0.14	0.18	0.14
NN	RMSE	153,456.09	165,403.05	123,058.56	123,878.06	139,970.90	238,603.48	138,305.61	144,329.96
		FY2018 LTC costs forecast				FY2019 LTC costs forecast
RF	R²	0.88	0.76	0.81	0.57	0.91	0.76	0.91	0.81
RF	RMSE	371,440.05	515,925.72	9,966.22	17,059.95	332,722.35	627,791.53	383,940.88	588,365.09
BT	R²	0.88	0.71	0.91	0.84	0.90	0.76	0.88	0.81
BT	RMSE	362,057.81	592,860.00	382,754.75	503,488.31	348,473.22	634,476.86	453,632.77	591,160.06
NN	R²	0.80	0.70	0.84	0.86	0.84	0.82	0.82	0.79
NN	RMSE	479,919.71	569,714.82	513,190.97	482,185.29	467,597.59	465,612.46	557,989.71	607,238.16
		FY2018 dental care costs forecast				FY2019 dental care costs forecast
RF	R²	0.57	0.19	0.93	0.81	0.64	0.37	0.74	0.50
RF	RMSE	20,771.65	17,686.07	345,739.77	541,611.70	13,190.36	12,823.56	12,747.81	16,681.27
BT	R²	0.62	0.36	0.71	0.60	0.60	0.25	0.66	0.50
BT	RMSE	14,110.74	37,624.58	12,808.71	13,983.91	13,875.81	13,992.39	14,655.08	16,751.27
NN	R²	0.23	0.33	0.59	0.68	0.43	0.39	0.50	0.44
NN	RMSE	26,141.12	24,249.86	15,033.57	13,272.91	15,040.67	19,442.02	17,062.55	20,740.20

For the 2019 prediction models, BT outperformed the 74-or-younger model, whereas RF demonstrated superior predictive performance in the 75-or-older model for medical costs. For estimating LTC costs, the NN excelled in the 74-or-younger model, whereas the RF excelled in the 75-or-older model. For predictions of dental care costs, NN excelled in the 74-or-younger model, and RF excelled in the 75-or-older model.

Regarding the most outstanding ML model, Tables 9–11 depict the degree of importance of the variables in the cost predictions for 2018 and 2019 and their primary effects. Figures 2–7 display a graph that compares the primary effects of different 2018 and 2019 models.

Table 9.

Degree of importance of variables related to predictions of medical costs of the ML model that was adopted, and their primary effects.

FY2018 medical costs forecast				FY2019 medical costs forecast
Under75_NN		Over75_NN		Under74_BT		Over75_RF
Variable importance	Main efect	Variable importance	Main efect	Variable importance	Main efect	Variable importance	Main efect
Use of LTC facilities	0.583	FY2017 dental care costs	0.244	FY2018 medical costs	0.505	FY2018 medical costs	0.639
FY2017 medical costs	0.168	Use of home care services	0.223	Use of LTC facilities	0.361	Use of LTC facilities	0.129
Care level	0.034	Use of LTC facilities	0.2	FY2018 LTC costs	0.024	Care level	0.019
FY2017 LTC costs	0.028	FY2017 medical costs	0.178	Care level	0.022	Use of home care services	0.008
Type 2 diabetes status	0.017	Care level	0.036	Type 2 diabetes status	0.006	FY2018 dental care costs	0.013
FY2017 dental care costs	0.014	FY2017 LTC costs	0.012	Sex	0.003	Type 2 diabetes status	0.009
Age	0.004	Sex	0.002	Use of home care services	0.001	FY2018 LTC costs	0.007
Sex	0.003	Age	0.0003	Age	0.001	Age	0.006
Use of home care services	0.001	Type 2 diabetes status	0.0003	FY2018 dental care costs	1.00E-03	Sex	0.003

Table 10.

Degree of importance of variables related to predictions of LTC costs of the ML model that was adopted, and their primary effects.

FY2018 LTC costs forecast				FY2019年 LTC costs forecast
Under75_RF		Over75_NN		Under74_NN		Over75_RF
Variable importance	Main efect	Variable importance	Main efect	Variable importance	Main efect	Variable importance	Main efect
FY2017 LTC costs	0.506	FY2017 LTC costs	0.583	Use of LTC facilities	0.458	FY2018LTC costs	0.577
Use of LTC facilities	0.38	Use of LTC facilities	0.181	FY2018LTC costs	0.294	Use of LTC facilities	0.316
Care level	0.028	Care level	0.102	Use of home care services	0.005	Care level	0.02
Use of home care services	0.014	Use of home care services	0.033	Care level	0.005	Use of home care services	0.009
Age	0.004	FY2017 dental care costs	0.014	Age	0.003	FY2018 medical costs	0.005
Sex	0.001	FY2017 medical costs	0.001	FY2018 medical costs	0.001	Age	0.0030
FY2017 dental care costs	0.001	Sex	0.001	FY2018 dental care costs	0.001	FY2018 dental care costs	0.0010
FY2017 medical costs	0.0003	Age	0.001	Sex	0.001	Sex	0.0010
Type 2 diabetes status	0.0001	Type 2 diabetes status	0.0004	Type 2 diabetes status	0.001	Type 2 diabetes status	0.0003

Table 11.

Degree of importance of variables related to predictions of dental care costs based on the ML model that was adopted, and their primary effects.

FY2018 dental care costs forecast				FY2019Dental care costs forecast
Under75_BT		Over75_RF		Under75_NN		Over75_RF
Variable importance	Main efect	Variable importance	Main efect	Variable importance	Main efect	Variable importance	Main efect
FY2017 dental care costs	0.96	FY2017 dental care costs	0.965	FY2018 dental care costs	0.767	FY2018 dental care costs	0.975
Use of home care services	0.004	Age	0.006	Age	0.05	Sex	0.001
Age	0.004	Care level	0.002	Use of LTC facilities	0.026	FY2018LTC costs	0.001
Care level	0.003	FY2017LTC costs	0.003	FY2018LTC costs	0.018	Care level	0.001
FY2017 medical costs	0.001	FY2017 medical costs	0.002	FY2018 medical costs	0.012	Age	0.001
Sex	0.0002	Sex	0.0010	Use of home care services	0.0130	FY2018 medical costs	0.0010
FY2017LTC costs	0.0003	Use of home care services	0.0004	Sex	0.0060	Type 2 diabetes status	0.0010
Use of LTC facilities	0.0002	Type 2 diabetes status	0.0003	Care level	0.0060	Use of LTC facilities	0.0004
Type 2 diabetes status	0.0001	Use of LTC facilities	0.0002	Type 2 diabetes status	0.0010	Use of home care services	0.0002

Figure 2.

Comparison of the degree of importance of variables for the 74-or-younger model (medical costs)

Figure 3.

Comparison of the degree of importance of variables for the 75-or-older model (medical costs)

Figure 4.

Comparison of the degree of importance of variables for the 74-or-younger model (LTC costs)

Figure 5.

Comparison of the degree of importance of variables for the 75-or-older model (LTC costs)

Figure 6.

Comparison of the degree of importance of variables for the 74-or-younger model (dental care costs)

Figure 7.

Comparison of the degree of importance of variablesfor the 75-or-older model (dental care costs)

Table 12 outlines the results obtained from analyzing the models that predicted medical, LTC, and dental care costs and their characteristics. In the model for medical costs, the previous year’s medical expenses and usage of facility services had a significant influence. Moreover, the presence or absence of type 2 diabetes, LTC costs, and status of dental visits tended to have little influence. In the LTC cost model, the previous year’s LTC costs and the usage or non-usage of facility services had a significant influence. The presence or absence of type 2 diabetes, medical costs, and status of dental visits tended to have little influence. In the model for dental care costs, the previous year’s costs had a significant influence, with the influence of other explanatory variables being extremely minimal. Healthcare status and LTC tended to have only a minor influence on dental care costs in the following year.

Table 12.

The results of various prediction models 1 year later and their characteristics.

	Medical costs model for the next year	LTC costs model for the next year	Dental care costs for the next year
Under 75	Use of LTC facilities:36∼58%Medical costs:16∼50%Prediction accuracy:Low	Use of LTC facilities:30∼50% LTC costs:40%Prediction accuracy:High	Dental care costs:76∼96%Prediction accuracy:Low
Over 75	Medical costs:24∼63%Use of LTC facilities:12∼20%Use of home care services:0∼20%Prediction accuracy:Low	LTC costs:58%Use of LTC facilities:18∼31%Prediction accuracy:High	Dental care costs:97%Prediction accuracy: Varying
Characteristics of results	・The impact of previous year’s medical costs and use of LTC facilities is significant.・Prediction accuracy is low.・The status of type 2 diabetes and the cost of LTC and dental care costs have little impact	・The impact of previous year’s LCT costs and use of LTC facilities is significant.・The status of type 2 diabetes and the cost of medical and dental care costs have little impact	・The impact of previous year’s dental care costs is significant.・Very small impact of explanatory variables other than dental care costs・Medical and LTC conditions have a fairly small impact on dental care costs in the next year

Figure 8 presents a relational diagram illustrating the degree of influence of year X’s medical, LTC, dental, and other costs on the expenses for year X + 1.

Figure 8.

Degree of influence on the medical, LTC, and dental care costs one year later

Discussion

An analysis of medical claims targeting LTCI users revealed that medical costs for year X as well as the usage status of facilities and home care services influenced year X + 1 medical costs. Year X’s LTC costs and the usage status of facility services may influence the LTC costs of year X + 1. Only year X’s dental care costs appeared to influence the dental care costs of year X + 1. Moreover, the presence or absence of type 2 diabetes, dental care expenses (status of dental visits), and level of care had an extremely small influence on medical, LTC, and dental care costs 1 year later.

Comparison and verification with prior research

We found that the medical costs of the previous year influenced the same costs in the following year and that the degree of this influence was high, especially among older people. Masuhara et al.²³ stated that, although medical costs show wide and random variations during youth and middle age, they may be closely correlated with past medical costs among older adults. We obtained similar outcomes. However, as the results indicated that predictive accuracy was extremely low, with an R² of 0.37 in the 74-or-younger model and 0.21 in the 75-or-older model, there is a need to reconsider the explanatory variables and target period.

Among participants aged 74 or younger, the usage or non-usage of facility services in the previous year had a significant influence on LTC costs in the following year, and among participants aged 75 or older, the LTC costs in the previous year had a significant influence on such expenditures in the following year. Under Japan’s LTCI program, the usage of LTC facility services is always costly. This indicates a pattern of people aged 74 or younger seeing their LTC costs rise dramatically the year after having used such facilities for the first time in the previous year, as very few people aged 74 or younger use these services. In other words, although this cannot be explained using the models based on our results, even among people aged 75 or older, LTC costs increased the following year if they had used LTC facility services.

The LTCI system varies by country. The LTCI system originated in the Netherlands. However, Germany and Japan were the first countries to legislate a public LTCI program and have a history of over 20 years of implementing numerous reforms and improvements. Germany’s LTCI Act prioritizes home-based care as well as prevention and rehabilitation as its two basic principles. Hence, even if an analysis similar to that employed in our study were conducted using German data, we assume that the outcomes would differ from the findings obtained in Japan, where numerous LTC facility services are offered.²⁷

The dental care costs accrued in the previous year significantly influenced the same costs in the following year. Takeuchi et al. and Kiuchi et al. noted the possibility of dental care interventions helping optimize medical and LTC costs.^21,22 However, our findings imply that the degree of their contribution was not significant.

Our findings showed that high-accuracy ML models differed according to the participants’ age and the type of cost model. Alexandre et al., who used ML, indicated that if medical cost prediction was the objective, RF should be prioritized, whereas Ugochukwu et al., who analyzed models for determining medical insurance costs, revealed that the RF model’s prediction errors were small, a finding that differed from that of our study.^12,18 Therefore, based on these results, when predicting medical, dental, and LTC costs, instead of assuming that there is an ML model that is optimal for all situations, it is important, especially when forecasting medical costs, to identify appropriate methods and techniques that match the data each time before embarking on analyses and comparisons.

Limitations

Medical insurance and LTCI systems are not internationally standardized, and the type, quality, and price of services differ by country. It is highly unlikely that applying the current analytical techniques to other countries and simply comparing their results will generate useful data. Because patients’ self-payment ratios also differ, the development of coefficients and other items that correct for such differences becomes necessary.

Moreover, because of the circumstances related to the periods of the databases that were available for use (2017–2019), we could produce only 2 years’ worth of prediction models for the following 1 year, resulting in a model that estimated the period only from years X to X + 1. However, substantial individual differences are present across the years when major medical, dental, and LTC costs accrue, and this frequency increases with age. Using long-term databases and building models that include patients’ personal health-related events in the medical, dental, and LTC sectors (especially for surgeries and other expensive events), we can establish more precise models for estimating medical, dental, and LTC costs.

Another limitation of this study was the exclusion of potentially important variables such as detailed clinical indicators, medication use, and socioeconomic status. These factors are known to influence healthcare expenditure and may improve model performance if included in future analyses. Additionally, although ML models were employed, full hyperparameter optimization and regularization procedures were limited, which might have affected predictive accuracy. The dataset might also contain unmeasured sources of bias or residual confounding factors, and the exclusion of records with missing values might have introduced selection bias, potentially influencing the generalizability of the results.

Conclusions

We investigated the influence of personal medical, dental, and LTC costs on such expenses in the short-term and quantitatively described the degree of their influence. To the best of our knowledge, this is the first study to combine a full range of personal medical, dental, and LTC costs to make predictions.

As a result of our analyses, which target patients with type 2 diabetes who also use LTCI, we found that medical costs in year X + 1 were notably influenced by medical costs in year X as well as the usage or non-usage of facilities and home-based care services. The LTC costs in year X + 1 were significantly influenced by medical costs in year X and the usage or non-usage of facility services. Dental care costs in year X + 1 were significantly influenced by dental care costs in year X. Moreover, the presence or absence of type 2 diabetes, dental care costs (status of dental visits), and level of care had an extremely small influence on medical, LTC, and dental care costs 1 year later. Because the predictive accuracy of LTC costs was high in our study, the results are reliable, especially as a model for predicting LTC costs in year X + 1.

Our results indicated that high-accuracy ML models differed according to age and cost models. Thus, when building prediction models that link medical, LTC, and dental care costs, it is important to set up individually tailored techniques each time that match the data before performing analyses and making comparisons.

Footnotes

ORCID iD

Teppei Suzuki

Ethical considerations

This study was approved by Graduate School of Health Sciences, Hokkaido University Research Ethics Committee (approval no. 22-50-2) on December 13, 2024.

Author contribution

TS designed the study, performed the data analysis, and drafted the manuscript. HS, HE, TA, and WN contributed to data collection and field investigation. KO provided critical advice on the study design and supervised the overall research process. All authors reviewed and approved the final version of the manuscript.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was funded by KAKENHI (Grant Number 22K17319), provided by the Japan Society for the Promotion of Science.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.*

References

Ministry of Health, Labour and Welfare . FY2023 trends in medical expenses -MEDIAS- [Internet]. Ministry of Health, Labour and Welfare, 2024. [cited 2025 Jun 16]. Available from: https://www.mhlw.go.jp/topics/medias/year/23/index.html

Ministry of Health, Labour and Welfare . Overview of actual statistics on nursing care benefits, etc. fiscal year 2023. Ministry of Health, Labour and Welfare, 2024. Available from: https://www.mhlw.go.jp/toukei/saikin/hw/kaigo/kyufu/23/index.html

Ministry of Health, Labour and Welfare . Trends in health and nutrition policy in Japan [Internet]. Ministry of Health, Labour and Welfare, 2024. Available from: https://www.mhlw.go.jp/content/10904750/001128375.pdf

Ministry of Health, Labour and Welfare . Data on trends in medical expenses [Internet]. Ministry of Health, Labour and Welfare, 2024. Available from: https://www.mhlw.go.jp/bunya/iryouhoken/database/zenpan/iryou_doukou.html

Japan Society for the Prevention of Lifestyle Related Diseases . Statistics on lifestyle related diseases [Internet]. Japan Society for the Prevention of Lifestyle Related Diseases. Available from: https://seikatsusyukanbyo.com/

Inoue

Maswana

, et al. Forecasting of future medical care expenditure in Japan using a system dynamics model. Inquiry 2022; 59: 00469580221091397, Available from: https://pubmed.ncbi.nlm.nih.gov/35420903/

Sase

Kumagai

Suzuki

, et al. Characteristics of type-2 diabetics who are prone to high-cost medical care expenses by Bayesian network. Int J Environ Res Public Health 2020; 17(15): 5271, Available from: https://www.mdpi.com/1660-4601/17/15/5271

International Diabetes Federation . IDF diabetes Atlas. 10th ed. International Diabetes Federation, 2021. Available from: https://diabetesatlas.org/resources/previous-editions/

Bloom

Chen

Kuhn

, et al. The economic burden of chronic diseases: estimates and projections for China, Japan, and South Korea. J Econ Ageing 2020; 17: 100163, Available from: https://www.sciencedirect.com/science/article/abs/pii/S2212828X18300744

10.

Yetim

İlgün

Çilhoroz

, et al. The socioeconomic determinants of health expenditure in OECD: an examination on panel data. Int J Healthc Manag. 2021; 14(4): 1265–1269.

11.

Vellela

Pushpalatha

Sarathkumar

, et al. Advanced intelligence health insurance cost prediction using random forest. ZKG Int 2023; VIII(1): Available from: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4473700

12.

Orji

Ukwandu

. Machine learning for an explainable cost prediction of medical insurance. Mach Learn Appl 2024; 15: 100516. Available from: https://www.sciencedirect.com/science/article/pii/S2666827023000695

13.

Maisog

, et al. Using massive health insurance claims data to predict very high-cost claimants: a machine learning approach. arXiv Preprint [Internet]. 2019 [cited 2025 Jun 6];arXiv:1912.13032. Available from: https://arxiv.org/abs/1912.13032

14.

Duncan

Loginov

Ludkovski

. Testing alternative regression frameworks for predictive modeling of health care costs. N Am Actuar J 2016; 20(1): 65–87.

15.

Taloba

Abd El-Aziz

Alshanbari

, et al. Estimation and prediction of hospitalization and medical care costs using regression in machine learning. J Healthc Eng 2022; 2022: 7969220, Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC8906954/

16.

Panay

Baloian

Pino

, et al. Predicting health care costs using evidence regression. Proc 13th Int Conf Ubiquitous Comput Ambient Intell 2019; 31(1): 74. Available from: https://www.mdpi.com/2504-3900/31/1/74

17.

Vimont

Leleu

Durand-Zaleski

. Machine learning versus regression modelling in predicting individual healthcare costs from a representative sample of the nationwide claims database in France. Eur J Health Econ 2022; 23: 211–223.

18.

Kurosaki

Yoshimura

. Medical expenses for diabetes care in Japan: analysis of inter-prefecture differences. Jpn J Public Health 2020; 67(8): 501–508. Available from: https://www.jstage.jst.go.jp/article/jph/67/8/67_19-060/_article/-char/ja/

19.

Getzen

. Getzen model of long-run medical cost trends: update for 2024–2034+ [Internet]. Schaumburg (IL): Society of Actuaries, 2023. [cited 2025 Jun 6]. Available from: https://www.soa.org/resources/research-reports/2023/2024-getzen-model-update/

20.

Jevdjevic

Listl

Beeson

, et al. Forecasting future dental health expenditures: development of a framework using data from 32 OECD countries. Community Dent Oral Epidemiol 2021; 49(3): 256–266. Available from: https://pubmed.ncbi.nlm.nih.gov/33252147/

21.

Takeuchi

Sato

Suma

, et al. Associations of oral health status and dental health service utilization with dental and medical expenditures. J Dent Health 2017; 67: 160–171, Available from: https://www.jstage.jst.go.jp/article/jdh/67/3/67_160/_article/-char/ja/

22.

Kiuchi

Takeuchi

Saito

, et al. Differences in cumulative long-term care costs by dental visit pattern among Japanese older adults: the JAGES cohort study. J Gerontol A Biol Sci Med Sci 2024; 79(9): glae194, Available from: https://pubmed.ncbi.nlm.nih.gov/39101529/

23.

Masuhara

Konishi

Choui

, et al. Statistical theory on the simulation of lifetime medical expenditures using receipt data. Jpn Assoc Health Care Adm [Internet] 2015; 9(1): 47–56. Available from: https://www.jstage.jst.go.jp/article/jaha/9/1/9_47/_article/-char/ja/

24.

Gao

Yuan

. Forecasting the health transition and medical expenditure of the future elderly in China: a longitudinal study based on Markov chain and two part model. Front Public Health 2022; 9: 774140. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC8792851/

25.

Miyauchi

Hashimoto

Nishimura

. Health state transition model by binary expression and cubic lattice representation corresponding to the specific health checkup. Int J Affect Eng [Internet] 2021; 20(2): 49–55. Available from: https://www.jstage.jst.go.jp/article/ijae/20/2/20_IJAE-D-20-00019/_article/-char/ja

26.

Mori

Hamada

Yoshie

, et al. The associations of multimorbidity with the sum of annual medical and long-term care expenditures in Japan. BMC Geriatr 2019; 19: 69.

27.