Improving Prediction of High-Cost Health Care Users with Medical Check-Up Data

Use of big data from health insurance for assessment of cardiovascular outcomes

Outcome research that supports guideline recommendations for primary and secondary preventions largely depends on the data obtained from clinical trials or selected hospital populations. The exponentially growing amount of real-world medical data could enable fundamental improvements in cardiovascular disease (CVD) prediction, prevention, and care. In this review we summarize how data from health insurance claims (HIC) may improve our understanding of current health provision and identify challenges of patient care by implementing the perspective of patients (providing data and contributing to society), physicians (identifying at-risk patients, optimizing diagnosis and therapy), health insurers (preventive education and economic aspects), and policy makers (data-driven legislation). HIC data has the potential to inform relevant aspects of the healthcare systems. Although HIC data inherit limitations, large sample sizes and long-term follow-up provides enormous predictive power. Herein, we highlight the benefits and limitations of HIC data and provide examples from the cardiovascular field, i.e. how HIC data is supporting healthcare, focusing on the demographical and epidemiological differences, pharmacotherapy, healthcare utilization, cost-effectiveness and outcomes of different treatments. As an outlook we discuss the potential of using HIC-based big data and modern artificial intelligence (AI) algorithms to guide patient education and care, which could lead to the development of a learning healthcare system and support a medically relevant legislation in the future.

Development and Evaluation of Machine Learning-Based High-Cost Prediction Model Using Health Check-Up Data by the National Health Insurance Service of Korea

In this study, socioeconomic, medical treatment, and health check-up data from 2010 to 2017 of the National Health Insurance Service (NHIS) of Korea were analyzed. This year's socioeconomic, treatment, and health check-up data are used to develop a predictive model for high medical expenses in the next year. The characteristic of this study is to derive important variables related to the high cost of domestic medical expenses users by using data on health check-up items conducted by the country. In this study, we tried to classify data and evaluate its performance using classification supervised learning algorithms for high-cost medical expense prediction. Supervised learning for predicting high-cost medical expenses was performed using the logistic regression model, random forest, and XGBoost, which have been known to result the best performance and explanatory power among the machine learning algorithms used in previous studies. Our experimental results show that the XGBoost model had the best performance with 77.1% accuracy. The contribution of this study is to identify the variables that affect the prediction of high-cost medical expenses by analyzing the medical bills using the health check-up variables and the Korea Classification Disease (KCD) large group as input variables. Through this study, it was confirmed that musculoskeletal disorders (M) and respiratory diseases (J), which are the most frequently treated diseases, as important KCD disease groups for high-cost prediction in Korea, affect the future high cost prediction. In addition, it was confirmed that malignant neoplasia diseases (C) with high medical cost per treatment are a group of diseases related to high future medical cost prediction. Unlike previous studies, it is the result of analyzing all disease data, so it is expected that the study will be more meaningful when compared with the results of other national health check-up data.

Identifying patterns of clinical conditions among high-cost older adult health care users using claims data: a latent class approach

Objectives

To identify patterns of clinical conditions among high-cost older adults health care users and explore the associations between characteristics of high-cost older adults and patterns of clinical conditions.

Methods

We analyzed data from the Shanghai Basic Social Medical Insurance Database, China. A total of 2927 older adults aged 60 years and over were included as the analysis sample. We used latent class analysis to identify patterns of clinical conditions among high-cost older adults health care users. Multinomial logistic regression models were also used to determine the associations between demographic characteristics, insurance types, and patterns of clinical conditions.

Results

Five clinically distinctive subgroups of high-cost older adults emerged. Classes included “cerebrovascular diseases” (10.6% of high-cost older adults), “malignant tumor” (9.1%), “arthrosis” (8.8%), “ischemic heart disease” (7.4%), and “other sporadic diseases” (64.1%). Age, sex, and type of medical insurance were predictors of high-cost older adult subgroups.

Conclusions

Profiling patterns of clinical conditions among high-cost older adults is potentially useful as a first step to inform the development of tailored management and intervention strategies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12939-022-01688-3.

Machine learning versus regression modelling in predicting individual healthcare costs from a representative sample of the nationwide claims database in France

BACKGROUND

Innovative provider payment methods that avoid adverse selection and reward performance require accurate prediction of healthcare costs based on individual risk adjustment. Our objective was to compare the performances of a simple neural network (NN) and random forest (RF) to a generalized linear model (GLM) for the prediction of medical cost at the individual level.

METHODS

A 1/97 representative sample of the French National Health Data Information System was used. Predictors selected were: demographic information; pre-existing conditions, Charlson comorbidity index; healthcare service use and costs. Predictive performances of each model were compared through individual-level (adjusted R-squared (adj-R²), mean absolute error (MAE) and hit ratio (HiR)), and distribution-level metrics on different sets of covariates in the general population and by pre-existing morbid condition, using a quasi-Monte Carlo design.

RESULTS

We included 510,182 subjects alive on 31st December, 2015. Mean annual costs were 1894€ (standard deviation 9326€) (median 393€, IQ range 95€; 1480€), including zero-claim subjects. All models performed similarly after adjustment on demographics. RF model had better performances on other sets of covariates (pre-existing conditions, resource counts and past year costs). On full model, RF reached an adj-R² of 47.5%, a MAE of 1338€ and a HiR of 67%, while GLM and NN had an adj-R² of 34.7% and 31.6%, a MAE of 1635€ and 1660€, and a HiR of 58% and 55 M, respectively. RF model outperformed GLM and NN for most conditions and for high-cost subjects.

CONCLUSIONS

RF should be preferred when the objective is to best predict medical costs. When the objective is to understand the contribution of predictors, GLM was well suited with demographics, conditions and base year cost.

An integrated model for medical expense system optimization during diagnosis process based on artificial intelligence algorithm

In the era of artificial intelligence, the healthcare industry is undergoing tremendous innovation and development based on sophisticated AI algorithms. Focusing on diagnosis process and target disease, this study theoretically proposed an integrated model to optimize traditional medical expense system, and ultimately helps medical staff and patients make more reliable decisions. From the new perspective of total expense estimation and detailed expense analysis, the proposed model innovatively consists of two intelligent modules, with theoretical contribution. The two modules are SVM-based module and SOM-based module. According to the rigorous comparative analysis with two classic AI techniques, back propagation neural networks and random forests, it is demonstrated that the SVM-based module achieved better capability of total expense estimation. Meanwhile, by designing a two-stage clustering process, SOM-based module effectively generated decision clusters and corresponding cluster centers were obtained, that clarified the complex relationship between detailed expense and patient information. To achieve practical contribution, the proposed model was applied to the diagnosis process of coronary heart disease. The real data from a hospital in Shanghai was collected, and the validity and accuracy of the proposed model were verified with rigorous experiments. The proposed model innovatively optimized traditional medical expense system, and intelligently generated reliable decision-making information for both total expense and detailed expense. The successful application on the target disease further indicates that this model is a user-friendly tool for medical expense control and therapeutic regimen strategy.

Characteristics and resource utilization of high-cost users in the intensive care unit: a population-based cohort study

Background

Healthcare expenditure within the intensive care unit (ICU) is costly. A cost reduction strategy may be to target patients accounting for a disproportionate amount of healthcare spending, or high-cost users. This study aims to describe high-cost users in the ICU, including health outcomes and cost patterns.

Methods

We conducted a population-based retrospective cohort study of patients with ICU admissions in Ontario from 2011 to 2018. Patients with total healthcare costs in the year following ICU admission (including the admission itself) in the upper 10th percentile were defined as high-cost users. We compared characteristics and outcomes including length of stay, mortality, disposition, and costs between groups.

Results

Among 370,061 patients included, 37,006 were high-cost users. High-cost users were 64.2 years old, 58.3% male, and had more comorbidities (41.2% had ≥3) when likened to non-high cost users (66.1 years old, 57.2% male, 27.9% had ≥3 comorbidities). ICU length of stay was four times greater for high-cost users compared to non-high cost users (22.4 days, 95% confidence interval [CI] 22.0–22.7 days vs. 5.56 days, 95% CI 5.54–5.57 days). High-cost users had lower in-hospital mortality (10.0% vs.14.2%), but increased dispositioning outside of home (77.4% vs. 42.2%) compared to non-high-cost users. Total healthcare costs were five-fold higher for high-cost users ($238,231, 95% CI $237,020–$239,442) compared to non-high-cost users ($45,155, 95% CI $45,046–$45,264). High-cost users accounted for 37.0% of total healthcare costs.

Conclusion

High-cost users have increased length of stay, lower in-hospital mortality, and higher total healthcare costs when compared to non-high-cost users. Further studies into cost patterns and predictors of high-cost users are necessary to identify methods of decreasing healthcare expenditure.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12913-021-07318-y.

Assessing the Value of Unsupervised Clustering in Predicting Persistent High Health Care Utilizers: Retrospective Analysis of Insurance Claims Data

BACKGROUND

A high proportion of health care services are persistently utilized by a small subpopulation of patients. To improve clinical outcomes while reducing costs and utilization, population health management programs often provide targeted interventions to patients who may become persistent high users/utilizers (PHUs). Enhanced prediction and management of PHUs can improve health care system efficiencies and improve the overall quality of patient care.

OBJECTIVE

The aim of this study was to detect key classes of diseases and medications among the study population and to assess the predictive value of these classes in identifying PHUs.

METHODS

This study was a retrospective analysis of insurance claims data of patients from the Johns Hopkins Health Care system. We defined a PHU as a patient incurring health care costs in the top 20% of all patients' costs for 4 consecutive 6-month periods. We used 2013 claims data to predict PHU status in 2014-2015. We applied latent class analysis (LCA), an unsupervised clustering approach, to identify patient subgroups with similar diagnostic and medication patterns to differentiate variations in health care utilization across PHUs. Logistic regression models were then built to predict PHUs in the full population and in select subpopulations. Predictors included LCA membership probabilities, demographic covariates, and health utilization covariates. Predictive powers of the regression models were assessed and compared using standard metrics.

RESULTS

We identified 164,221 patients with continuous enrollment between 2013 and 2015. The mean study population age was 19.7 years, 55.9% were women, 3.3% had ≥1 hospitalization, and 19.1% had 10+ outpatient visits in 2013. A total of 8359 (5.09%) patients were identified as PHUs in both 2014 and 2015. The LCA performed optimally when assigning patients to four probability disease/medication classes. Given the feedback provided by clinical experts, we further divided the population into four diagnostic groups for sensitivity analysis: acute upper respiratory infection (URI) (n=53,232; 4.6% PHUs), mental health (n=34,456; 12.8% PHUs), otitis media (n=24,992; 4.5% PHUs), and musculoskeletal (n=24,799; 15.5% PHUs). For the regression models predicting PHUs in the full population, the F1-score classification metric was lower using a parsimonious model that included LCA categories (F1=38.62%) compared to that of a complex risk stratification model with a full set of predictors (F1=48.20%). However, the LCA-enabled simple models were comparable to the complex model when predicting PHUs in the mental health and musculoskeletal subpopulations (F1-scores of 48.69% and 48.15%, respectively). F1-scores were lower than that of the complex model when the LCA-enabled models were limited to the otitis media and acute URI subpopulations (45.77% and 43.05%, respectively).

CONCLUSIONS

Our study illustrates the value of LCA in identifying subgroups of patients with similar patterns of diagnoses and medications. Our results show that LCA-derived classes can simplify predictive models of PHUs without compromising predictive accuracy. Future studies should investigate the value of LCA-derived classes for predicting PHUs in other health care settings.

Application of Bayesian network and regression method in treatment cost prediction

Charging according to disease is an important way to effectively promote the reform of medical insurance mechanism, reasonably allocate medical resources and reduce the burden of patients, and it is also an important direction of medical development at home and abroad. The cost forecast of single disease can not only find the potential influence and driving factors, but also estimate the active cost, and tell the management and reasonable allocation of medical resources. In this paper, a method of Bayesian network combined with regression analysis is proposed to predict the cost of treatment based on the patient's electronic medical record when the amount of data is small. Firstly, a set of text-based medical record data conversion method is established, and in the clustering method, the missing value interpolation is carried out by weighted method according to the distance, which completes the data preparation and processing for the realization of data prediction. Then, aiming at the problem of low prediction accuracy of traditional regression model, this paper establishes a prediction model combined with local weight regression method after Bayesian network interpretation and classification of patients' treatment process. Finally, the model is verified with the medical record data provided by the hospital, and the results show that the model has higher prediction accuracy.

Predicting Future Service Use in Dutch Mental Healthcare: A Machine Learning Approach

A mental healthcare system in which the scarce resources are equitably and efficiently allocated, benefits from a predictive model about expected service use. The skewness in service use is a challenge for such models. In this study, we applied a machine learning approach to forecast expected service use, as a starting point for agreements between financiers and suppliers of mental healthcare. This study used administrative data from a large mental healthcare organization in the Netherlands. A training set was selected using records from 2017 (N = 10,911), and a test set was selected using records from 2018 (N = 10,201). A baseline model and three random forest models were created from different types of input data to predict (the remainder of) numeric individual treatment hours. A visual analysis was performed on the individual predictions. Patients consumed 62 h of mental healthcare on average in 2018. The model that best predicted service use had a mean error of 21 min at the insurance group level and an average absolute error of 28 h at the patient level. There was a systematic under prediction of service use for high service use patients. The application of machine learning techniques on mental healthcare data is useful for predicting expected service on group level. The results indicate that these models could support financiers and suppliers of healthcare in the planning and allocation of resources. Nevertheless, uncertainty in the prediction of high-cost patients remains a challenge.

Association between serum uric acid and obesity in Chinese adults: a 9-year longitudinal data analysis

OBJECTIVES

Hyperuricaemia has been reported to be significantly associated with risk of obesity. However, previous studies on the association between serum uric acid (SUA) and body mass index (BMI) yielded conflicting results. The present study examined the relationship between SUA and obesity among Chinese adults.

METHODS

Data were collected at Guangdong Second Provincial General Hospital in Guangzhou City, China, between January 2010 and December 2018. Participants with ≥2 medical check-up times were included in our analyses. Physical examinations and laboratory measurement variables were obtained from the medical check-up system. The high SUA level group was classified as participants with hyperuricaemia, and obesity was defined as BMI ≥28 kg/m². Logistic regression model was performed for data at baseline. For all participants, generalised estimation equation (GEE) model was used to assess the association between SUA and obesity, where the data were repeatedly measured over the 9-year study period. Subgroup analyses were performed by gender and age group. We calculated the cut-off values for SUA of obesity using the receiver operating characteristic curves (ROC) technique.

RESULTS

A total of 15 959 participants (10 023 men and 5936 women) were included in this study, with an average age of 37.38 years (SD: 13.27) and average SUA of 367.05 μmol/L (SD: 97.97) at baseline, respectively. Finally, 1078 participants developed obesity over the 9-year period. The prevalence of obesity was approximately 14.2% for high SUA level. In logistic regression analysis at baseline, we observed a positive association between SUA and risk of obesity: OR=1.84 (95% CI: 1.77 to 1.90) for per-SD increase in SUA. Considering repeated measures over 9 year for all participants in the GEE model, the per-SD OR was 1.85 (95% CI: 1.77 to 1.91) for SUA and the increased risk of obesity were greater for men (OR=1.45) and elderly participants (OR=1.01). In subgroup analyses by gender and age, we observed significant associations between SUA and obesity with higher risk in women (OR=2.35) and young participants (OR=1.87) when compared with men (OR=1.70) and elderly participants (OR=1.48). The SUA cut-off points for risk of obesity using ROC curves were approximately consistent with the international standard.

CONCLUSIONS

Our study observed higher SUA level was associated with increased risk of obesity. More high-quality research is needed to further support these findings.

Trading-Off Machine Learning Algorithms towards Data-Driven Administrative-Socio-Economic Population Health Management

Together with population ageing, the number of people suffering from multimorbidity is increasing, up to more than half of the population by 2035. This part of the population is composed by the highest-risk patients, who are, at the same time, the major users of the healthcare systems. The early identification of this sub-population can really help to improve people’s quality of life and reduce healthcare costs. In this paper, we describe a population health management tool based on state-of-the-art intelligent algorithms, starting from administrative and socio-economic data, for the early identification of high-risk patients. The study refers to the population of the Local Health Unit of Central Tuscany in 2015, which amounts to 1,670,129 residents. After a trade-off on machine learning models and on input data, Random Forest applied to 1-year of historical data achieves the best results, outperforming state-of-the-art models. The most important variables for this model, in terms of mean minimal depth, accuracy decrease and Gini decrease, result to be age and some group of drugs, such as high-ceiling diuretics. Thanks to the low inference time and reduced memory usage, the resulting model allows for real-time risk prediction updates whenever new data become available, giving General Practitioners the possibility to early adopt personalised medicine.

Toward Business Process Innovation in the Big Data Era: A Mediating Roles of Big Data Knowledge Management

While recent debate recognizes the importance of big data (BD) and knowledge management (KM) in firm performance, there has been a paucity of literature regarding big data analytics technological (BDAT) and knowledge exploration-exploitation capabilities (KEEC) in the context of business process innovation (BPI). This study aims to identify whether BD and KM can be established in these emerging issues. We used a survey questionnaire to collect data from various firms and industries. We used structural equation modeling (SmartPLS and SPSS) to validate the research model with a sample of 155 companies in a developing country such as Indonesia. The result demonstrates a positive relationship between KEEC and BPI, followed by several significant findings such as BDAT with KEEC; KEEC on big data knowledge management (BDKM); BDKM and BPI; and BDAT on BDKM. In contrast, BDAT is nonsignificant for direct relationship on BPI, and interestingly, it becomes a significant result after mediated by BDKM. Similarly, BDKM has successfully mediated the relationship between KEEC and BPI. The management level ideally develops and increases such a knowledge creation/acquisition practices and BDAT in an organization to gain more meaningful benefits from these two capabilities. BDAT, KEEC, and BDKM simultaneously are a clear antecedent approach, which ultimately results in flexibility, effectiveness, and effectivity of BPI. The cases of this research are profit firms in a developing country such as Indonesia. A future study could be considered in different settings such as type of industries or more specific company's type, the economy level of countries (comparing between developed and developing countries), and environmental dynamical. A novel field of study is the inclusion of knowledge exploration-exploitation and BDAT that drives BPI.

Machine-learning-based prediction models for high-need high-cost patients using nationwide clinical and claims data

High-need, high-cost (HNHC) patients—usually defined as those who account for the top 5% of annual healthcare costs—use as much as half of the total healthcare costs. Accurately predicting future HNHC patients and designing targeted interventions for them has the potential to effectively control rapidly growing healthcare expenditures. To achieve this goal, we used a nationally representative random sample of the working-age population who underwent a screening program in Japan in 2013–2016, and developed five machine-learning-based prediction models for HNHC patients in the subsequent year. Predictors include demographics, blood pressure, laboratory tests (e.g., HbA1c, LDL-C, and AST), survey responses (e.g., smoking status, medications, and past medical history), and annual healthcare cost in the prior year. Our prediction models for HNHC patients combining clinical data from the national screening program with claims data showed a c-statistics of 0.84 (95%CI, 0.83–0.86), and overperformed traditional prediction models relying only on claims data.