smartCasaplus – risk-based geriatric health coaching for multimorbid insured persons

Chronic diseases are increasingly determining the disease pattern in an aging society and leading to changing healthcare needs. A more preventative-oriented healthcare approach is particularly important in this regard, as the burden of disease is influenced by various risk factors resulting from the interaction between risky behavior, psychosocial stress, and socioeconomic factors. Multimorbidity, particularly in old age, is closely linked to the increased prevalence of chronic diseases. Depending on the definition, this means the simultaneous occurrence of at least two or more diseases (Puth et al. 2017). A recent analysis of diagnostic data from 67 million statutory health insurance patients in 2023 shows a multimorbidity prevalence of 31.4% for the overall cohort and 78.2% for those aged 65 and over (Schmitz et al. 2023). Multimorbidity was defined as the presence of at least three diseases over a period of at least three quarters.

> Multimorbidity differs from the related concept of comorbidity, which refers to the combined effects of additional diseases in relation to the index disease in an individual and often occurs in the context of disease clusters. Some diseases and combinations of diseases are particularly common in multimorbidity, with three main patterns emerging from the multitude of combinations found in studies: first, combinations of cardiovascular and metabolic diseases; second, mental disorders (often in combination with physical diseases); and third, musculoskeletal diseases associated with different physical diagnoses (Güthlin et al. 2020). As the number of diseases increases, the classification of symptoms and the selection of appropriate treatment options become more complex, the resulting interactions less manageable, and the risk of adverse interactions increases. This increases the likelihood of adverse side effects, medication errors, and drug interactions that can even lead to hospitalization. Multi-medication often causes additional nonspecific symptoms such as fatigue, loss of appetite, dizziness, confusion, tremors, or falls. This can lead to additional medication prescriptions and hospital admissions.
Multimorbid geriatric patients, in particular, have a higher risk of premature death, hospitalization, and longer length of stay than those with a single chronic illness. Multimorbidity is also associated with poorer functional capacity and health-related quality of life, depression, and the use of multiple medications (polypharmacy), as well as higher socioeconomic costs. In light of this, the question arises of a service that goes beyond the possibilities of indication-based approaches and, above all, caters to the situation of people whose health situation is determined less by illnesses than by mutual symptoms, complaints, and functional limitations resulting from the complex interplay of various chronic and acute illnesses, the respective contribution to the overall situation of which is difficult to determine and, moreover, is often secondary to coping with everyday life. In this case, the (geriatric) patient will only be able to be an "expert in their illness" in the traditional sense to a limited extent.
Complex disease processes pose new challenges for medical care, as they increase the urgency of consistent patient and user orientation: on the one hand, tailoring care to individual patient needs, and on the other, strengthening patient orientation by supporting empowerment, self-management, and personal responsibility of those affected (Burger 2013: 16). A person's self-management skills are particularly relevant in multimorbidity, as individuals must simultaneously manage multiple problems and providers who may operate in different sectors (Leijten 2018). Healthcare should therefore consider the individual's needs and preferences and prioritize common treatment goals (DEGAM 2018).
This gives rise to the need for an individualized, patient-centered approach. Central to the integrated care of patients with multiple chronic diseases is a holistic understanding of the individual in their individual life situation. The main focus here is on the individual's health, well-being, and self-care capabilities, as well as available resources, such as support from family members (Burger/Habel 2014).

The geriatric training program Casaplus
Frailty and a lack of knowledge in dealing with multiple, concurrent chronic illnesses often lead to a lack of independence in the home and social environment among older people. To prevent unwanted escalations of crisis situations, maintain or strengthen everyday competence, and delay the need for care, preventive measures can effectively support age-specific health promotion (Hildebrandt: 2011). One such offering is the geriatric training program Casaplus: Since 2007, it has offered telephone health coaching for multimorbid geriatric patients with insurance, in which approximately 30,000 people have already participated nationwide. It is intended as a complement to standard medical care measures and implements recommendations for managing multimorbid older people by selecting suitable program participants through risk stratification, joint goal definition taking into account needs and preferences, prioritizing the issues and issues that cause the greatest distress from the perspective of those affected, and supporting self-management. Casaplus pursues a holistic and person-centered approach, providing advice and training across indications rather than focusing on specific illnesses. Casaplus targets insured persons with an increased risk of hospitalization and enables older people to receive the appropriate support for each individual.
The Casaplus coaching program comprises a total of seven counseling modules that address the risks typical for geriatrics described above (see Figure 1). First, a detailed medical history is taken to identify predisposition to age-specific risks such as malnutrition, risk of falls, adverse drug reactions, or social isolation. Program participants are then assigned to one or more counseling modules by their personal health coach, depending on the counseling goal and individual preference or level of distress.
Building on this, insured persons are regularly sensitized to potential health risks by their personal advisor, trained in symptom management and early crisis detection, and, if necessary, guided to appropriate care structures. In 2023, program participants received support in an average of three to four counseling modules. If the program participant is assigned to one of the three modules—falls, malnutrition or dehydration, or care and social situation—they are offered outreach counseling by a nursing service. A nationwide network of nursing services is available for this purpose. The specially trained nursing professionals gain a detailed on-site overview of the participant and their living situation, provide recommendations for action, and, if necessary, intervene in a low-threshold manner, for example, by pointing out tripping hazards. Module changes during the program, e.g., due to interim events (e.g., further indication, unexpected inpatient stay), are possible and occur regularly.
Casaplus thus complements the care of geriatric patients in their familiar environment and strengthens their everyday skills. MedicalContact AG's telephone health coaching is needs-based. The goal is to capture as many factors as possible that influence health outcomes and tailor support accordingly. Upon request, the consultation can also be extended to family members of the insured person.

smartCasaplus – with risk-based candidate identification
After ten years of successful use, Casaplus underwent a comprehensive relaunch in 2017, initially for VIACTIV policyholders. This involved the intervention itself, for example, by adding the oncology counseling module and updating the counseling content of the other modules. The Casaplus coaching program was also expanded to smartCasaplus, featuring identification based on continuously learning risk models that enable ongoing adaptation of the identification process based on the latest evaluation results and data sources.

The smartCasaplus KEP hospital admission prediction model
The success of Casaplus depends on identifying insured individuals with a sufficiently high hospitalization risk. To this end, MedicalContact AG has been using a hospital admission prediction model (KEP model) from DCC Risikoanalytik GmbH since 2017. This model is continuously adapted to the Casaplus program and features the following key changes from the previous model:
• In contrast to the LOH model originally used, it is not a static but a learning prediction model, which is subject to a continuous change process and is adapted for each identification.
• The hospital admission prediction model was developed from a general model to a specific model, each focusing on the hospital cases that can be influenced by Casaplus. For this purpose, the CEP target variable to be predicted (a list of ICD10 discharge diagnoses) is adjusted to the indications that were demonstrably successfully influenced in the previous evaluation period.
• In addition, a gap-based risk forecast is performed, i.e., a gap of two to three quarters between the identification time and the forecast period is considered for participant recruitment. This includes the time required for information, recruitment, and the admission interviews (see Figure 1). This prevents the predicted inpatient stay from occurring before the intervention can take effect.
• Furthermore, the smartCasaplus-KEP model uses the most up-to-date billing data available, taking into account the delays typical for statutory health insurance (GKV) in individual main service areas. For this purpose, a data status analysis and quality assurance are performed at each identification point in time, and the current LAG pattern (quarters with data delays for each main service area) usable for model application is extracted. For model adjustment, this specific LAG pattern is projected into the past, taking into account the GAP quarters, so that the target variable is fully available in the current database for at least twelve current months (see Figure 2).
• Finally, the CEP model was expanded to include regional attributes to take into account the very different regional supply structures.

The current KEP model is used to identify those insured persons in the population whose Casaplus-specific hospitalization risk is at least three times higher in the forecast period.

The smartCasaplus impact forecast model
In addition to the CEP model, the impact model is another key component of the program's ongoing focus on those insured persons with the highest probability of success. During the annual evaluation, those program participants are identified for whom a positive effect on the target variable can be measured in the second program year. As part of the ongoing adjustment of the impact model, these participants form the target group IMP = 1. The impact prediction model is adjusted analogously to the CEP model based on the current data (see Fig. 2) and evaluation results. By applying the impact model to all candidates identified through the CEP prediction, it is possible to focus recruitment on those insured persons with a sufficiently high hospitalization risk that can also be influenced by geriatric coaching.

The smartCasaplus SignIn forecast model
The SignIn prediction model is the third component of the smartCasaplus identification process. It determines the probability that identified candidates will actually enroll in the program. The goal of this model is to initially contact those insured persons most likely to participate in the program during the recruitment process, maximizing the acquisition rate and making optimal use of limited coaching resources. For the quarterly SignIn model adjustment, the target group (SignIn = 1) consists of all candidates who have been successfully recruited for the program in the past three years, i.e., who have participated in the program for at least 181 days. The SignIn prediction is also based on the current data basis according to the LAG-GAP principle in Figure 2.
The model quality is assessed based on the area under the receiver operator curve (ROC), the area under the curve (AUC). The ROC describes the ratio of the true positive detection rate (y-axis – sensitivity) to the false positive detection rate (x-axis – specificity) for all thresholds (points on the curve). The closer the AUC approaches the ideal value of 1, the more favorable the ratio between the highest possible sensitivity (high true positive rate) and the highest possible specificity (low false positive rate) at the optimal threshold (peak).
To optimize the AUC of the smartCasaplus forecast models, the originally used logistic regression methods are being gradually replaced by AI-based, high-dimensional deep learning (MLP multi-layer perceptron with 2,453 input nodes, two hidden layers, and ADAM learning or regression tree methods with 2,453 attributes, 12 levels, six nodes, and Bayesian parameter optimization). The currently used forecast models achieve an AUC between 0.78 and 0.88 (see Fig. 3).
As part of the smartCasaplus candidate list creation process, a Combiscore is calculated, which incorporates the three forecast models with varying weightings. The highest weighting is given to the CEP risk, and the lowest to the sign-in probability.

Evaluation of the smartCasaplus program

A total of 2,951 VIACTIV insured persons were enrolled in the smartCasaplus program as of June 30, 2023. Of these, 2,164 were valid participants on the cutoff date, i.e., had received at least 181 days of care. For the required retrospective long-term study, the 1,673 valid participants enrolled between September 15, 2017 (relaunch) and June 30, 2022, were assigned to the 10 half-year cohorts shown in Figure 4 based on their enrollment date.
The 100 smartCasaplus participants (sCP) who newly enrolled in a disease management program (DMP) during their sCP participation were excluded from the intervention group to be evaluated. The 1,228 participants who were previously enrolled in a DMP remained in the intervention group. However, to minimize bias due to program overlap between sCP and DMP, these sCP+DMP participants were compared with other DMP participants (without sCP participation) (see below).
The study population of the evaluation includes approximately 242,000 VIACTIV insured persons for each cohort,
• were fully insured during the evaluation period (4 quarters before the start of the intervention until the end of the data status) or died after the start of enrollment of the respective cohort,
• were between 60 and 100 years old on the day of enrollment,
• for which none of the 51 exclusion ICDs was documented as a confirmed outpatient or inpatient main or secondary diagnosis during the matching period (4 quarters before the start of the intervention) and
• for whom no care level > 2 and no inpatient care was documented during the matching period.

A key challenge in implementing retrospective evaluations based on claims data is the identification of suitable virtual control groups. This involves forming a comparison group of insured individuals who met the study population's inclusion and exclusion criteria, but whose outcome risk is comparable to that of the intervention participants.

Risk-adjusted propensity score matching (RAPSM)
The propensity score matching (PSM) method has been established as the most important method for identifying retrospective comparison groups in health care analysis (Austin 2014). This method involves testing multiple individual attributes in a statistical modeling procedure for their significance in explaining participation in an intervention. This model is subsequently applied to determine a propensity score for all insured persons in the study population. By identifying a "best-matching" partner for each insured person in the intervention group, "matched pairs" are formed. This ensures that, even when multiple attributes from the areas of demographics, regionality, morbidity, and therapy are taken into account, each participant in the intervention group is assigned the most similar comparison partner and that a comparison group of at least equal size is identified within the study population. The evaluation of the intervention effects on the outcome parameter is subsequently carried out using two-way comparative analyses of the primary outcome parameter between the intervention and comparison groups.
A major shortcoming of the classic PSM method is that the dependent variable in propensity modeling is the binary coded intervention participation and not the outcome parameter to be assessed (Kim et al. 2018; Schneeweiss 2009). As a result, while the identified comparison group has a similarly high probability of participating in the intervention, this similarity does not relate to the risk of the outcome to be assessed. However, the ability of the propensity model to explain the risk of the outcome to be measured is crucial for the validity of the retrospective comparison group identified in PSM.
In the Risk-Adjusted Propensity Score Matching (RAPSM) procedure, developed between 2020 and 2022 in the German-Canadian collaborative project AIR_PTE (AI-based Risk Prediction and Treatment Effect Estimation), the attributes of the propensity model are selected using an upstream AI-based risk model of the outcome parameter to be assessed. The RAPSM procedure no longer uses all available attributes as the basis for the subsequent PSM, but only those identified as risk factors using Shapley analysis (Yongchan 2021). The propensity model formed on this basis now explains not only the similarity regarding the intervention but also the similarity regarding the outcome risk (Zahn et al. 2023). As a result, the comparison group identified using RAPSM not only has a similar distribution of propensity scores, but also a similar distribution of outcome risk scores. The subsequent effect analysis is thus able to determine the effect size of the intervention on the primary outcome parameter in a cross-sectional manner (within a defined time interval after the start of the intervention), but also in a longitudinal manner (by continuously comparing the development of the outcome parameter in the intervention and comparison groups over several time intervals).

Result of the current smartCasaplus evaluation
For each of the 10 cohorts, a risk-adjusted control group is formed using the RAPSM procedure. Participation in the six disease management programs (to reduce DMP overlay effects) and the sign-in probability (to reduce sign-on bias) are additional attributes included in the RAPSM model. The 1:1 allocation of the most similar comparison partner (CG) to each intervention participant (IG) is performed using nearest neighbor matching without replacement and under the condition that they have never participated in the intervention. The effects on the primary outcome parameter, hospital costs, and the secondary outcome parameters, total service costs (excluding dentist fees and sickness benefits) and morbidity allocation (HMG) from the health fund, are evaluated iteratively for all 10 cohorts in the 1st to 4th year of the intervention relative to the start of the intervention (see Figure 4).
A distribution-based pairwise extreme value adjustment is carried out, in which all couples in which the target value of one partner deviates by more than 2 times the standard deviation from the expected value of the target value distribution in the respective intervention year are excluded.
As can be seen in Figure 5 based on the development of mean hospital costs per insured person, the program effect on the intervention group (IG-red) can be determined on this basis as an absolute deviation from the comparison group (VG-blue) in the matching period and each of the 4 intervention years.
In addition to the number of adjusted pairs and the absolute mean difference, their significance as a p-value of a two-sided test and their relative deviation in percent are determined. As a criterion for successful matching, smartCasaplus requires a non-significant mean difference of < = 5% during the matching period.
The results of the smartCasaplus evaluation 2023 show a significant effect (p < = 0.05) of geriatric coaching on the primary outcome parameter hospital costs totaling EUR 3,824 per participant in all 4 intervention years.
By applying the same evaluation principle to the secondary outcome parameter total service costs (TSP), the 2023 program evaluation measured a TSP dampening effect of a total of EUR 3,892 according to the difference-to-difference (D2D) principle in the first four program years.
With regard to the secondary outcome parameter morbidity allocation (HMG-Zuw), these expenditure reductions are offset by a revenue reduction effect of EUR 938 in the first four program years, which only occurs from the second program year onwards due to the prospective nature of the morbidity allocation.
The return-on-investment (ROI) calculation for the smartCasaplus program shown in Figure 6 takes into account not only the cost and revenue containment but also the average program costs per participant for the service provider MedicalContact AG. In the first year of the intervention, the acquisition costs incurred before the start of the intervention are also included. As shown in Figure 6, the smartCasaplus evaluation 2023 results in an ROI of 1.1 in the first program year and a total ROI of 1.7 in the first four program years.

Conclusion
Data-based risk prediction makes it possible to use modern AI-based identification processes to focus health insurers' patient-centered care programs on high-risk patients whose risk is particularly susceptible to being influenced by the respective intervention. AI-based methods such as RAPSM can also be used in the ongoing evaluation of care programs to determine valid and statistically significant effects on expenditures and revenues and use these to continuously optimize the identification processes. The Health Data Use Act, which comes into force on February 1, 2024, creates further opportunities for health insurers and service providers to implement these methods in healthcare practice.
As can be seen from the example of the geriatric coaching program smartCasaplus, focusing on geriatric coaching using adaptive prognosis models enables both the economical use of the insurance contributions used for this purpose and patient-centered, individualized care for insured individuals in particular need. Individual coaching programs such as smartCasaplus can reduce the demographically driven growing gaps in holistic and patient-centered care, particularly in the psychosocial area, and achieve significant effects for the benefit of the statutory health insurance community.