Chapter 8: Effects of Comorbidities (Chronic Illness) on Hospitalization and Mortality Risks: Available to Clinicians Through an App
-
Published:2024
8.1 Introduction
Chronic diseases are a major cause of mortality and morbidity, accounting for more than 75% of the healthcare burden in Europe and costing several hundred billion euros each year [1]. Treatment and care for these patients require coordinated input from a wide range of healthcare professionals over a period of many years. Assessing potential risks and benefits are a relevant issue for planning tailored decisions in an ‘acute on chronic’ setting. Comorbidity, the co-occurrence of two or more chronic diseases, usually produces an incremental increase in the need for hospitalization and in all-cause, 5-year mortality. These risks depend not only on the potentially new reason for admission, i.e., a new health concern, but also on which co-occurrence takes place and impacts the current reason for admission [2].
Efforts have been directed at identifying clinical risk groups in terms of costs, and several approaches have objectified scaling measures to this end. Examples are the CIRS (Cumulative Illness Rating Scale), [3] ICED (Index of Coexisting Disease), [4] and Clinical Risk Groups [5]. Other aggregated scales include gender and age in the disease clusters, e.g., ACGs (Adjusted Clinical Groups), [6] DRGs (Diagnosis-Related Groups), [7] and HRGs (Healthcare Resource Groups) [8]. Each creation procedure is different depending on the objectives pursued and the type of algorithm used. However, the usefulness of these approaches has been questioned because they are purely descriptive and thus not appropriate for making clinical decisions.
The scales described above use agglomerative hierarchical clustering methods dependent on the similarity of patient characteristics with different selection procedures, regardless of whether the procedures are supervised or unsupervised, and use mathematical inferences. Comorbidity clustering, however, allows the reallocation of a patient into comorbidity groups. This so-called divisive hierarchical risk clustering can provide relevant information on risks and can drive information for tuning the relevance and desirability of clinical interventions.
Likewise, the potential impact of a newly developed disease becomes visible as the clinician searches for the impact of a specific comorbidity.
In the present study, a divisive hierarchical risk clustering of circa four million adults observed during a period of five years was performed. Big data technologies were applied to prepare the data for modeling. A mobile app and a dashboard were developed to facilitate clinical use.
8.2 Challenges and Focus
The challenges described in this manuscript are the development of an application and a dashboard to allow the exploitation of the models for clinical risk stratification in such a way as to be useful for the clinician and to allow tailoring for an individual patient. The obtained information for a given patient is useful for intensifying the treatment in order to reallocate the patient into a better risk subset in the corresponding comorbidity group.
This led to the following two tangible goals:
A smart phone app
This app was to be the interface between modeling and clinical needs. It allows a rapid, real-time calculation of risk for an individual subject, visible via the smartphone app. A physician should be able to enter the gender, the age range, the diseases as the basis for the comorbidity group, and finally the values available, such as a lab or functional status, for the relevant clinical features. The application should then generate a report with both general and specific risks for the cohort of patients with a similar profile, including the risk for hospitalization itself or for 5year mortality for the group and that individual patient. Furthermore, the physician should be able to obtain the impact of each comorbid disease on hospitalization and mortality risk. Designed with input from a multi-disciplinary group, the process was to meet or exceed robust data and privacy protection legislation.
Dashboard
The aim of the dashboard was to provide an easy and intuitive way of exploring the data, such as how hospitalization and all-cause 5-year mortality change across different comorbidity groups or clusters. A further aim is to explore how the need for hospitalization and the projected mortality risk change across different comorbidity groups and differences within the comorbidity group or clusters. The dashboard should have a panel for each outcome measure. Each panel should offer data in multiple forms, including charts displaying information per selected comorbidity group.
8.2.1 Design, Privacy, Study Population, and Data Collection
Data from the universal healthcare system of the Valencian Community (Spain) were used. There is one unique electronic centralized clinical record per patient in this community. This database included 3,799,885 individuals older than 18 years in 2012. Total population data were extracted for the period from 1 January 2012 to 31 December 2016.
The registry includes patient demographics, medications, vital status, medical history, and laboratory data. Patient data underwent a two-step process for privacy purposes. The 1st step was the pseudo-anonymization of all data in the record, and in the 2nd step, automated and manual reviews were performed by masking or deleting variables that could lead to the re-identification of patients. Spanish Law 3/2018 for Data Protection and Guaranty of Digital Rights and corresponding European norms were followed [9]. The Committee for Ethics and Clinical Trials of the Hospital Clinico of Valencia approved the study prior to data extraction with registration number F-CE-Geva-15.
Diseases are listed using the ICD-9 coding system. Aggregate data in the system had been collected by the database system on multiple occasions and from different locations, i.e., wherever the patients had received care. This could mean that not all the baseline variables that were required to adjust for potential confounding were available at the time of inclusion. A 6-month window was therefore defined around the time of data extraction in order to gather the most complete information on lab values and control variables (e.g., body mass index (BMI), blood pressure, serum cholesterol fractions, creatinine, and diabetes marker HbA1c).
For each patient, the set of clinical features was brought into the study database as ordinal values with several levels for each clinical variable. Health-related features measured for each individual were paired with the discrete values assigned to the different intervals. Table 8.1 shows the list of clinical measures included in the modeling.
8.3 Data Management Methods
The architectural layers and the specific technologies deployed in the study are summarized in Figure 8.1. The technological stack was mainly inspired by the Cloudera distribution Hadoop Search for reference. A distributed Cloudera cluster was created using two servers: Operating System, Ubuntu Server 18.04; CPU, Intel(R) Xeon(R) Gold 5118 CPU @ 2.30 GHz with 12 cores; RAM, 1536 GB RAM DDR2 ECC (24 modules, 64 GB, 2666 MHz); and HDD, 5.24TB SSD RAID5 & 32.74TB HDD SATA RAID5. This stack enforced distributed data processing to maximize the use of computing resources.
The data flow in this architecture could be summarized as follows: First, the original data were extracted and processed into a set of files suitable for analysis. This set of files was stored in the Hadoop Distributed File System (HDFS). Next, queries were executed with Hive or SparkSQL and generated a data set with the relevant features. Afterward, this second data set was used to train the different models using Spark and Skitlearn Search as references. Finally, models were stored on an external server and accessed with the Representational State Transfer Application Programming Interface (REST API).
8.3.1 Mathematical Modeling
The mathematical modeling consisted of two different stages: First, the hierarchical clustering identified the distinct comorbidity groups that produce an increment in hospitalization and/or all-cause mortality, taken from a total of 19 diseases and with up to four co-morbid diseases per group (Table 8.1). In the second stage, the patient model aimed to offer a detailed report on each cluster of comorbidities. Previously determined, the risks studied were genderand age-dependent, and a prior manual stratification of these factors was established. For this, each gender population was divided into five different age groups.
8.4 Results
Data learning and computations were robust, and the translation from descriptive to clinically relevant with respect to privacy legislation was achieved. Output aspects, such that the clinician could assess and use the data, are described below. A digital system displaying the stratification of risks for hospitalization and mortality in chronic disease clusters has been developed, which includes 19 chronic diseases, 5 age groups, and gender.
Information about the different data available to work with in the distinct stages of the project. The table shows five main groups of information: gender and ageranges that help in handily dividing the population into ten different sub-populations, creating one group for each Cartesian combination of gender and age-range; the 19 diseases used to establish the different comorbidity groups with a hierarchical and statistical outlook; the clinical features, another group of ordinal variables used as explanatory variables for the patient models in each comorbidity group; and the two objective variables (risks) of the modularizations.
| Gender (number of individuals) | |
|---|---|
| Man | (1,689,036) |
| Woman | (1,828,192) |
| Age-ranges (number of individuals) | |
|---|---|
| 25+ | (433,394) [25–39] |
| 40+ | (1,125,338) [40–54] |
| 55+ | (896,373) [55–69] |
| 70+ | (640,417) [70–84] |
| 85+ | (421,706) [85–8] |
| List of diseases selected (number of individuals) | |
|---|---|
| 1. Anemia (632,229) | 11. Dislipidemia (1,330,191) |
| 2. Anxiety (1,119,416) | 12. Hypertension (1,134,860) |
| 3. Asthma (155,172) | 13. Heart Failure (135,323) |
| 4. Atrial Fibrillation (180,232) | 14. Hypothyroidism (256,199) |
| 5. Cardiomiopathy (29,673) | 15. NAFLD (65,189) |
| 6. Cirrhosis (112,602) | 16. Osteoarthritis (270,860) |
| 7. COPD (332,735) | 17. Stroke (182,743) |
| 8. Chronic Kidney Disease (144,444) | 18. Vascular Dementia (31,404) |
| 9. Dementia (187,546) | 19. Venous Thromboembolism (49,752) |
| 10. Diabetes (489,547) |
| Clinical features levels (number of records) | |
|---|---|
| 1. Smoker: | 0: No Smoker, 1: Smoker |
| 2. TAS: (mmHg) (2,140,114) | 0: < 100, 1: 100 – 139, 2: > 139 |
| 3. TAD: (mmHg) (2,143,090) | 0: < 80, 1: 80 – 89, 2: > 89 |
| 4. Glucose: (mg/dL) (2,404,397) | 0: 0 – 125, 1: > 125 |
| 5. HbA1c: (%) (1,031,080) | 0: 0 – 7, 1: > 7 |
| 6. Creatinine: (mg/dL) (2,401,173) | 0: 0 – 1.5, 1: > 1.5 |
| 7. LDL Cholesterol: (mg/dL) (2,221,718) | 0: 0 – 100, 1: > 100 |
| 8. HDL Cholesterol: (mg/dL) (2,235,509) | 0: 0 – 44, 1: > 44 |
| Risk (Hospitalization and mortality) of interest (number of records) | |
|---|---|
| 1. Exitus: | 0: non-dead, 1: dead. (355,021 subjects) |
| 2. Hospitalization: | 0: the cause of hospitalization is none of the diseases of the comorbidity group, 1: the cause of hospitalization is one or more of the diseases that forms the comorbidity group. (3,966,827 admissions) (1,646,059 subjects) |
Software architecture. Cluster software employed to create the distributed cluster for computation. All brands shown can be trademarks or registered trademarks with their respective owners.
8.4.1 Smartphone App
The pictures shown in Figure 8.2 represent screens of the smart phone application. The first step consists of gender selection, then the physician selects the age range, and then the set of comorbid diseases. This organizes the patients into groups. The physician can then enter values for relevant clinical features to enable the two risk assessments (hospitalization and 5-year mortality risks). The application shows a report with general and specific endpoints for the group and the patient. Finally, the physician can select more information about each of the risks and receive more details for each one via a new screen.
Smartphone app. Screens to introduce the different data: category, age range, diseases, clinical features, and finally obtain the distinct risk endpoints or factors and the detailed report about the distinct risk information and the relevance of the clinical variables for the patient risk prediction.
8.4.2 Dashboard
In order to provide an easy and intuitive way of exploring the huge volume of data analyzed, a reporting web dashboard has been designed and developed. The dashboard is organized with one panel for each of these two risks being analyzed. Each panel shows charts that display information related to the selected comorbidity group. Figures 8.3, 8.4, and 8.5 show the different data charts displaying aspects of mortality risk as a function of comorbidities.
In Figure 8.3, two bar charts display information on the category and age distributions of the specific comorbidity cluster that the user has selected, showing the mortality risk and the number of affected patients.
For a more general overview of the data of interest, a scatter plot can be generated (Figure 8.4). This chart displays each cluster as a different data point, with the number of patients on the Y-axis and the relevant factor on the X-axis. Once the user selects a set of up to four diseases, the chart updates to show the comorbidity cluster of the selected diseases, as well as all the clusters that contain them. The user can also filter by age group and/or category in order to see more specific data; he/she can zoom in/out, hover over the elements to see the information of a specific cluster, and download the images.
Univariate descriptive analysis. Category and age distribution for a chosen comorbidity cluster. On the left, the mortality over time is shown for patients with heart failure and hypertension, distributed in age and gender categories. On the right, the distribution of patients and gender categories with heart failure and hypertension is shown.
Number of patients versus the mortality risk over time (presented as a percentage). The picture shows aggregated information for each comorbidity group for a given category and age, focusing on the clusters’ mortality risk.
Diseases conditional graph. It shows conditioned relations between different comorbidity clusters. The darker the hue between the related paths, the greater the probability of the conditioned transition. The larger the node, the greater the corresponding risk for that comorbidity group.
A graph chart shows the relations between different comorbidity clusters (Figure 8.5). The corresponding risk for each comorbidity group or disease has been previously computed by counting the individuals that show the risk versus the number of individuals in the corresponding group. To this end, each dot in the graph represents a different comorbidity cluster. The initial dot is the comorbidity cluster selected by the user, comprised of N diseases. For each dot, further dots represent the impacts of N + 1 comorbidities on the risks for hospitalization and mortality.
When a user selects a specific age group and category, information regarding the risks and number of patients affected by each comorbidity can be displayed in two ways. On the one hand, an additional information pop-up appears when the user hovers over a specific dot. On the other hand, the size of the dot represents the incidence of that risk in the specific case. The total number of patients involved in the analysis is used to gradually change the color between the overall outcome dot and the dot the user is hovering over.
8.5 Learnings and Outcomes
This chapter describes the process and outcomes of this study, which used big data strategies to translate comorbidity and risks into a clinically useful set of platforms. It developed a divisive hierarchical tree strategy to stratify the risk of hospitalization and mortality for an assortment of comorbidities and comorbidity bundles into clusters of chronic diseases and their corresponding relevant clinical features. This methodology allows better insight into a specific patient group or even an individual patient by conditioning the estimations to the pertinent comorbidity group and the selected relevant explicative variable values. The information for a given patient is useful for making a choice about intensifying treatment to reallocate the patient into a better risk subset in the corresponding comorbidity group or seeing that this effect is unlikely. The potential impact of a newly acquired disease is observable. Furthermore, the software created can be launched over a big data cluster in a transparent way, thus increasing the computational power transparently and dynamically.
8.6 Discussion
In the study described in this chapter, a digital system displaying the stratification of risks for hospitalization and mortality in chronic disease clusters has been developed, which includes 19 chronic diseases, 5 age groups, and gender. The user interface is implemented in two ways: the first identifies the different significant comorbidity clusters for each gender, the age ranges, up to four comorbidities, and their respective hospitalization and mortality risks. The second identifies the most relevant general clinical features for each cluster from a set of eight features and introduces prediction. The system provides information on the risk of each cluster, the contribution of each disease, and the clinical features associated with each, thereby allowing the physicians to make individually tailored decisions for treatment.
The study was conducted using data from the unique Electronic Health Record (EHR) of the general population in a Mediterranean population (Valencian Community, Spain). An interesting observation about this EHR is that the level of quality of life and the diagnosis reported seem to underreport chronic disease and focus on incidental health issues without comorbid context. However, in two recent reviews, the prevalence of chronic health conditions was higher in medical records than in other data sources such as administrative data or even health surveys, [10,11] while the more symptomatic chronic diseases are more poorly recorded [12,13].
Previously, the different machine learning approaches used to obtain clinical risk groups focused mainly on stratifying the severity of comorbid diseases or getting information on potential costs, regardless of gender or age. Examples of the former were CIRS, ICED, and the Kaplan Index, while of the latter, ACGs, DRGs, and HRGs. The procedure for group creation differed depending on the pursued objectives and the kind of algorithm used. One approach has used an agglomerative hierarchical clustering method, [14] similar to the approach used in the present study, while other ACGs (Adjusted Clinical Groups), [6] DRGs (Diagnosis-Related Groups), [7] and HRGs (Healthcare Resource Groups), [8] automatically establish the groups of comorbidities by searching for similarities in the clinical measures of patients [15]. An unsupervised disease clustering technique based on a multidimensional nonlineal projection (UMAP) has also been described [16,17].
The divisive hierarchical risk clustering used in the present study uses a two-step approach to achieve the desired outputs. In comparison, the more general modularizations consider all the information in one step of modeling. The two-step method permits creating specific models for each of the comorbidity groups, which also include only the clinical features required by the group. This method can save time because it evaluates only the clinical features needed for the corresponding group. Our results show that the accuracy of the different models depends on gender, age range, the impact of the group, and the relevant features specific to the group. Note that in our approach, even within the same cluster, clinical features will result in different outcomes for the two risks of hospitalization and all-cause 5-year mortality.
The creation of divisive hierarchical groups permits an individual that belongs to a specific group to pertain as well to another of the more general relevant groups from which an individual comes. Moreover, the use of interpretable models, like decision trees, [18] also allows for their use as recommender systems. With these models, it is possible to obtain a specific location for a patient in the tree and know the clinical variables to enhance in order for a patient to improve their risk position within the group. Furthermore, it can help to forecast the other diseases that are likely to occur in a patient in order to improve the patient’s allocation into a more specific group. In addition, this methodology permits the discovery of conditioned knowledge in a natural way, something that could help physicians discard other relevant diseases that might alter the considered factor. Notwithstanding that, it allows different predictions for a patient that is in a sparsely populated group for which it has not been possible to obtain models.
8.6.1 Perspectives
This project focused on the data management and data presentation forms and included choices for models implemented within the data analysis. Its ultimate purpose is to support the clinician in forming policy for an individual patient based on the extensive database information. Validation and implementation were not parts of the study. A clinical and field discussion still needs to be done, potentially driven by specific medical specialties, as well as case/control studies to assess performance in terms of treatment selection and risk reduction.
In identifying the different clusters and the impact of the two risks within the cluster, the methodology used in the current study has limitations. The divisive hierarchical approach, used to obtain the different comorbidity groups, has a higher impact when the number of patients in that group is smaller. This can result in a smaller number of samples for model estimation if there are a larger number of different groups. Consequently, we excluded groups with less than 51 subjects. By increasing the number of individuals by including data from other EHRs, we increased the precision of the comorbidity groups. Collecting more data will increase the capacity to study more complex groups with more cohorts and, in the future, expand the number of diseases included in the research.
The number of clinical parameters is sparse in the EHR, although they are typically relevant for prognosis and risk calculations. However, a more diverse set of clinical variables would improve the results of the models, providing higher accuracy.
Along with the number of clinical features, another relevant point is the conceptual (temporal) frame. Assigning a value to a parameter does not take into consideration the variability of each parameter, potentially introducing bias into the assignment of variables with higher variability. A way to improve this could be to consist of the establishment of a dynamical time interval related to the natural variability of the clinical variable. The methodology used in the current study allows a further approach to a specific patient model by conditioning the risk estimations to the pertinent comorbidity group and selecting relevant and explicative variable values for that group. Furthermore, the created software can be launched over a big data cluster in a transparent way, thus increasing the computational power transparently and dynamically. In the near future, a case/control study will evaluate performance in terms of treatment selection and risk reduction.
8.7 Conclusion
The study described in this chapter goes to the heart of big data use for clinical medicine. By creating user interfaces that allow the clinician to use real data to support choices for a group or even an individual patient with an app or dashboard. The study lays the groundwork for procedures and technologies that others can follow and extend.
With 19 chronic conditions and millions of patient health records to work from, this concept will resonate as the need for objective strategies for consumption in healthcare and quality of life gains traction.
Source Indentification
Icons of the app made by author from www.flaticon.com
FreePik. Gender identity pack. Woman. Retrieved from https://www.flaticon.com/freeicon/woman_1864518.
FreePik. Gender identity pack. Man. Retrieved from https://www.flaticon.com/freeicon/man_1864509.
FreePik. World Pride pack. Female symbol. Retrieved from https://www.flaticon.com/freeicon/femenine_949792.
FreePik. Family-7 pack. Girl. Retrieved from https://www.flaticon.com/freeicon/girl_375304.
itim2101. Avatar pack. Business Woman. Retrieved from https://www.flaticon.com/freeicon/business-woman_1439675.
FreePik. Family-7 pack. Mother. Retrieved from https://www.flaticon.com/freeicon/mother_375257.
itim2101. Avatar pack. Worker. Retrieved from https://www.flaticon.com/freeicon/worker_1439706.
FreePik. Family-7 pack. Grandmother. Retrieved from https://www.flaticon.com/freeicon/grandmother_375303.
Eucalyp. Medical pack. Medical check. Retrieved from https://www.flaticon.com/freeicon/medical-check_946299.
icon8. Green sandglass icon. Retrieved from https://www.iconsdb.com/greenicons/sandglass-icon.html.
Vectors Market. Medical. Cardiogram. Retrieved from https://www.flaticon.com/freeicon/cardiogram_607580.
Dashboard graphics created by using the libraries https://www.chartjs.org and https://www.amcharts.com (https://www.amcharts.com/docs/v4/chart-types/forcedirected/).
