Detecting Multimorbidity Patterns with Association Rule Mining in Patients with Alzheimer’s Disease and Related Dementias

Abstract

Researchers estimate the number of dementia patients to triple by 2050¹. Dementia seldom occurs in isolation; it’s frequently accompanied by other health conditions². The coexistence of conditions further complicates the management of dementia. In this study, we embarked on an innovative approach, applying association rule mining to analyze National Alzheimer’s Coordinating Center (NACC) data. First, we completed a literature review on the utilization of association rules, heatmaps, and network analysis to detect and visualize comorbidities. Then, we conducted a secondary data analysis on the NACC data using association rule mining. This algorithm uncovers associations of comorbidities that are diagnosed together in patients who have Alzheimer’s disease and related dementias (ADRD). Also, for these patients, the algorithm provides the probability of a patient developing another comorbidity given the diagnosis of an associated comorbidity. These findings can enhance treatment planning, advance research on high-association diseases, and ultimately enhance healthcare for dementia patients.

Introduction

Alzheimer’s disease and related dementias (ADRD) are a collection of degenerative neurological disorders that progressively hinder cognitive function and memory³. ADRD is prevalent amongst older adults, as approximately 6.2 million individuals aged 65 and above in the United States (U.S.) are currently experiencing Alzheimer’s dementia⁴. It is a recurring observation that individuals with ADRD often present with coexisting health conditions, or multimorbidities². Multiple Chronic Conditions (MCCs) or multimorbidities are defined as the concurrent presence of two or more chronic health conditions⁵.

Current statistics reveal a complex prospect of comorbidities amongst dementia patients. Cardiovascular conditions, including hypertension, heart disease, and stroke frequently coincide with dementia⁶. Metabolic disorders like diabetes and obesity are also prevalent, potentially contributing to neurodegenerative processes⁷. Depression and anxiety are common psychiatric comorbidities, highlighting the intricate relationship between mental well-being and the gradual changes in cognitive function⁸. Respiratory conditions such as chronic obstructive pulmonary disease (COPD) are also outstanding, suggesting a broader impact of comorbidities across various physiological systems⁹.

A considerable amount of research has focused on examining how comorbidities might affect the advancement of ADRD. For example, a study by Leszek et al. (2021) suggested that cardiovascular disease can lead to Alzheimer’s Dementia¹⁰. Similarly, research by Cholerton et al. (2016) highlighted that numerous studies conducted within various populations have consistently demonstrated a link between type 2 diabetes and cognitive decline. Moreover, this study suggested that a patient with type 2 diabetes is a good candidate for a precision health approach to treating dementias¹¹. Furthermore, another research study uncovered a relationship between the number of comorbidities and the patient’s cognitive condition, with “comorbidity increas[ing] the rate of cognitive decline”¹². Also, the number of comorbid conditions increases as the severity of dementia progresses¹³. Thus, MCCs play a significant role in the management of ADRD as they provide valuable insights that can lead to improved clinical management, optimized treatments, early detection of health issues, better resource allocation, enhanced quality of life, and advancements in our understanding of the associated comorbidities¹⁴. This research is crucial for addressing the holistic healthcare needs of ADRD patients and improving their overall well-being.

The Apriori algorithm stands out as a valuable tool. It has the ability to reveal important associations among the comorbidities. Using the Apriori algorithm to analyze information from ADRD patients, we can detect hidden relationships between various comorbidities. For instance, the algorithm will show if certain combinations of chronic conditions tend to co-occur at a rate higher than what random chance would predict¹⁵. The algorithm’s ability to find complex connections means it can uncover insights that traditional statistical methods might overlook. Central to this research is the algorithm’s capability to sift through vast electronic health records and identify hidden connections. This approach goes beyond traditional statistical methodologies, exposing patterns that might have previously remained unnoticed. The Apriori algorithm’s proven success in healthcare research positions it as a potent tool to explore the complex comorbidities associated with ADRD.

Comorbidity studies have increasingly utilized heatmaps as an informative tool to visualize correlations between MCCs. Heatmaps provide a visual representation of the complex data by assigning colors to depict the strength of the correlation between the comorbidities. In these studies, each comorbidity is represented along both the x and y axes of the heatmap, and the intersecting cells are shaded according to the magnitude of the correlation. Lee et al (2017) conducted a study that utilized heatmaps to assess the association of comorbidities in older cancer patients undergoing chemotherapy¹⁶. Another study by Venkatakrishnan et al (2021), utilized heat mapping to visualize associations between comorbidities in COVID-19 patients¹⁷. The findings of both studies confirm that using heat mapping enhances the contextual understanding of the prevalence of comorbidities and enables visualization of associated comorbidities. Heat Mapping is an excellent visualization mechanism and will be beneficial to this research.

Comorbidity studies harnessed the power of network analysis graphs as an insightful method for visualizing comorbidity correlations and associations. These studies construct intricate networks in which diseases are represented as nodes, and the relationships between them as edges. The resulting graph offers a comprehensive and intuitive view of correlations between MCCs, highlighting not only the presence of comorbidities but also the comorbidities they coexist with. Ieva and Bitonti (2018) conducted a network analysis study of comorbidities diagnosed in patients with heart failure, they concluded that: “network analysis can be considered a useful tool in epidemiologic framework when relational data are the objective of the investigation, since it allows to visualize and make inference on patterns of association among nodes (here HF comorbidities) by means of both qualitative indexes and clustering techniques”¹⁸. By utilizing network analysis, researchers gain a deeper understanding of how comorbidities cluster and interact, enabling them to identify key players in disease co-occurrence and develop more targeted strategies for patient care and management. This approach has proven invaluable in uncovering complex comorbidity patterns and advancing our knowledge of the intricate web of health conditions that individuals may face.

To investigate the connections between MCCs and ADRD, we initially conducted a comprehensive literature review on PubMed, focusing on the utilization of associated rule mining, heat mapping, and network analysis graphs within the context of National Alzheimer’s Coordinating Center (NACC) data. Our investigation revealed a research gap in this area. While some studies have applied these techniques in comorbidity research, none of them, to our knowledge, have employed these methods to investigate ADRD comorbidities using NACC data. Our research approach will be innovative in utilizing these techniques and will contribute to the advancement of knowledge that’s necessary in the field of MCCs within the realm of ADRD. In this research, we will uncover associations among comorbidities diagnosed alongside ADRD. Moreover, we will determine the likelihood of a patient acquiring another comorbidity after being diagnosed with a related comorbidity in the presence of ADRD.

To accomplish this project, we conducted a retrospective observational study using the NACC survey data collected from 42 Alzheimer’s Disease Research Centers throughout the U.S. . The NACC patient database captures patients’ clinical information as part of the Uniform Data Set (UDS). The UDS comprises longitudinal data collected annually through standardized clinical evaluations conducted at the Alzheimer’s Disease Centers funded by the National Institute on Aging (NIA)¹⁹. The NACC data is an ideal source for secondary data analysis on ADRD comorbidities. It’s a large and diverse dataset with a standardized collection procedure emphasizing ADRD.

Methods

The Apriori algorithm - an associated rule data mining tool -was initially developed to perform market basket analysis, identifying items frequently purchased together in single transactions. However, it was found to have valuable applications in healthcare, enabling the exploration of frequent symptoms, patterns, and comorbidities within patient data: “Apriori algorithm-based association rule analysis provides interpretable and intuitive results to inform general trends in the database. It has already been introduced in the field of exploring the comorbidity of attention-deficit hyperactivity disorder”²⁰.

The Apriori algorithm has the ability to reveal important associations among comorbidities. Using the Apriori algorithm to analyze information from ADRD patients, we can detect hidden relationships between various comorbidities. The algorithm allows us to determine what combinations of chronic conditions co-occur more frequently at a rate greater than what random chance would predict¹⁵. This capacity is particularly relevant in studying comorbidities in ADRD patients. The algorithm operates on the principle that if a set of items is frequent, then all of its subsets must also be frequent: “It processes by identifying noticeable rules among frequent patterns that are mined first through establishing threshold values of support and confidence”. This notion of “prior knowledge” supports the algorithm’s effectiveness in identifying strong associations among items. The algorithm’s functionality relies on two key metrics: support and confidence. Support quantifies the frequency of a specific itemset within the dataset, while confidence gauges the likelihood that a rule connecting two itemsets holds true.

Additionally, the Apriori algorithm employs pruning, a mechanism that discards itemsets falling below a defined support threshold, optimizing computational efficacy. The Apriori algorithm identifies the associations between the comorbidities as “rules” or frequent patterns. The left-hand side (LHS) would include a comorbidity and the right-hand side( RHS) would include the comorbidity that is highly associated with the comorbidity defined in the LHS. All the rules generated will fall within the predetermined confidence and support criteria^20-21.

While the Apriori algorithm will serve as the means to decipher MCC associations, it will be important to visualize the associations amongst the comorbidities. Two methods that allow us to visualize correlations are heatmaps and network analysis graphs. Heatmaps are graphical representations used to visualize data by assigning specific colors to data values, creating a visual pattern that highlights variations and trends within the dataset. Heatmaps are commonly employed in data analysis to reveal patterns, clusters, or correlations within large sets of data. By using a color spectrum, they provide an intuitive way to identify high or low values, making it easier to interpret complex information^16-17.

On the other hand, network analysis graphs are a powerful tool for studying and visualizing unique relationships and interactions among entities in a system or dataset. These entities can be nodes representing individual elements or objects, and the connections or edges between them represent relationships or interactions. Network analysis helps visualize dynamics within complex systems, such as social networks, transportation networks, or biological interactions¹⁸. By creating a visual representation of these connections, network graphs make it easier to analyze and understand how different components are linked and influence one another.

Both heatmaps and network analysis graphs are invaluable tools in the realm of data analysis. They each offer unique insights into complex data patterns and associations. While heatmaps excel at revealing patterns within numerical data matrices^16-17, network analysis graphs provide a visual means to explore unique connections between entities, shedding light on the relationships that might not be immediately apparent from the raw data¹⁸. Furthermore, these methods can be complementary when examining comorbidities among ADRD patients. Heatmaps will be used to visualize associations and patterns from ADRD patients highlighting potential relationships between different conditions. Network analysis graphs, on the other hand, will show how various diseases or conditions are interconnected within the patient population. Together, these techniques enable us to visualize the association rules generated by the apriori algorithms.

Results

Study Participants

To prepare the data for this research, a meticulous process was undertaken to streamline the information. Initially, a challenge emerged due to the presence of multiple entries per patient. This resulted from patients having records for both their initial visit and subsequent follow-up visits. However, for the purposes of this research, the sequence of visits held no significance for the point being. The primary objective is whether comorbidities are diagnosed at any point for patients with a confirmed ADRD diagnosis. Consequently, a decision was made to consolidate rows pertaining to multiple entries per patient. This transformation ensured that each patient was represented by a single row in the dataset.

In preparing the data for analysis, additional adjustments were made. Diagnostic information that wasn’t numerical in nature was removed, retaining only their numerical counterparts. Furthermore, the distinction between various types of a single disease became irrelevant within the scope of this investigation. The focal point during this stage was centered on determining whether a patient had been diagnosed with a specific condition or not. Consequently, codes indicating disease types were omitted from the dataset. All these adjustments resulted in 7231 entries for patients with ADRD. The characteristics pertaining to the study participants are identified in Table 1 and are sorted by gender. Table 1 was generated using the “table1” R package.

Table 1:

Participant characteristics and gender differences.

We generated a binary data frame from the filtered data, containing only values of ones and zeros, in order to facilitate data analysis. Binary data enables us to recognize patterns and to use the patient information in the apriori algorithm. This revamped dataset contained details about patients with confirmed ADRD diagnoses, while also encompassing information about concurrently diagnosed MCCs. The dimensions of the binary data frame contained 7231 rows, each corresponding to a distinct patient, and 18 variables that represent different comorbidities. Among this cohort of 7231 patients, the most recurrently diagnosed comorbidities, ranked in descending order of prevalence, were hypercholesterolemia, hypertension, arthritis, urinary incontinence, and bowel incontinence (Figure 1). Hypercholesterolemia led the list with 4198 patients receiving this diagnosis, followed by hypertension with 3957 patients, arthritis with 3503 patients, urinary incontinence with 3083 patients, and bowel incontinence with 1778 patients.

Figure 1.

Prevalence of diagnosed comorbidities in ADRD patients. Diseases are abbreviated as follows: HYPCHOL = Hypercholesterolemia present, HYPERT = Hypertension present, ARTH = Arthritis present, URINEINC = urinary incontinence present, BOWLINC = bowel incontinence present.

Apriori Algorithm

To apply the association rule to this dataset, we utilized the “arules” R package. Before applying the Apriori algorithm, we transformed the existing data frame into a transaction matrix. This conversion was crucial to align the data with the Apriori algorithm’s requirements. In using the Apriori algorithm, we opted for patterns with a confidence level surpassing 0.60, a support exceeding 0.10, and a lift greater than 1.0, as illustrated in Figure 2. It’s worth noting that this selection criterion aligns with criteria used by other research studies that have harnessed the Apriori algorithm for similar data analysis tasks^21-25. A confidence level exceeding 0.60 indicates a strong association between the comorbidities, suggesting a high probability that the presence of one disease is dependent on the presence of another. A support greater than 0.10 indicates that these patterns occur frequently enough to be of particular interest. Moreover, a lift greater than 1.0 highlights that the observed associations are not coincidental occurrences but rather demonstrate a true correlation between the comorbidities. This standardized approach allowed us to extract meaningful insights and associations from our data.

Figure 2.

Rules generated by the apriori algorithm with a confidence value > 0.60 and a support value >0.10. Diseases are abbreviated as follows: BOWLINC = bowel incontinence present, URINEINC = urinary incontinence present, HYPERT = Hypertension present, ARTH = Arthritis present, HYPCHOL = Hypercholesterolemia present, DIABET = Diabetes present at visit, SLEEPAP =Sleep apnea present.

Consider the first rule in Figure 2, with a confidence of 0.9381, this means that out of all the transactions where patients have bowel incontinence (LHS item), about 94% of those transactions (patients) have urinary incontinence (RHS item). Additionally, a support value of 0.2307 indicates that 23% of patients in the dataset have both bowel and urinary incontinence. In other words, these two conditions co-occur relatively frequently amongst our ADRD patients. A coverage value of 0.2459 means that approximately 25% of our patients are diagnosed with either urinary or bowel incontinence. Lastly, A lift greater than 1.0 (in this case, 2.20) suggests that there is a positive correlation between both conditions, meaning that the presence of one condition (e.g., “bowel incontinence”) increases the likelihood of the other condition (e.g., “urinary incontinence”) occurring by a factor of 2.20. Therefore, the apriori algorithm proves that there is a strong connection between these two comorbidities.

All 22 rules generated by the Apriori algorithm exhibit a lift value greater than one, indicating a positive correlation among the listed diseases in Figure 2. For instance, the second rule suggests that the presence of both hypertension and bowel incontinence significantly raises the likelihood of urinary incontinence, with nearly 94% of ADRD patients in the dataset who have been diagnosed with hypertension and bowel incontinence also confirmed to have urinary incontinence. This proves a strong association among urinary incontinence, bowel incontinence, and hypertension in ADRD patients. Furthermore, the remaining rules in Figure 2 illustrate that urinary incontinence is strongly linked with bowel incontinence and arthritis, as well as hypercholesterolemia and bowel incontinence. Notably, hypercholesterolemia exhibits a high association with diabetes, and a similar trend is observed for hypertension as well. Figure 2 offers further opportunities to uncover additional correlations through a similar analytical approach. These findings provide valuable insights into the dynamic of comorbidities within this patient population. The association rules generated by the apriori algorithm can help healthcare providers enhance their diagnoses and treatment plans for ADRD patients; taking into account the potential development of MCCs associated with their existing comorbidities.

Heatmap

To visualize the associations identified by the Apriori algorithm, we used a heatmap to display the non-binary comorbidity data. Heatmaps are a widely used graphical representation that uses colors to display the values of a matrix. In a heatmap, each cell’s color reflects its value, making it easy to spot correlations or trends based on the shade of color. We updated the comorbidity data frame to a matrix so a heatmap can be generated using the “ggplot2” R package as shown in Figure 3.

Figure 3.

Heatmap representation of the correlations “rules” generated by the Apriori Algorithm.

At a glance, we can easily visualize the rules generated by the apriori algorithm. In Figure 3, we see dark purple or almost indigo cells that signify a strong correlation value of 1, while lighter shades of purple represent decreasing levels of correlation. For example, the first rule generated by the apriori algorithm is very apparent in this heatmap. The strong association between urinary and bowel incontinence is represented by the darkest shade of purple in the heatmap. Additionally, in the hypercholesterolemia column, there is a visible but less intense shade of purple corresponding to the diabetes and hypertension cells, illustrating the existence of a strong association. Moreover, the hypertension column exhibits a visible purple hue at the intersection of diabetes and hypercholesterolemia, exhibiting an association between those comorbidities as well. This heatmap provides valuable insight when it comes to visualizing association rules, enhancing our understanding of MCC associations in the ADRD population.

Network Analysis

Following the heatmap, we conducted a network analysis using the R package “bootnet” on our dataset. The network analysis showcases some fascinating connections between the comorbidities, some of which were generated through the apriori algorithm in addition to some new connections. For example, the first rule generated by the apriori algorithm is portrayed in Figure 4. The strong association between urinary and bowel incontinence is represented by a significantly thick line, referred to as an “edge”, connecting these nodes that symbolize the comorbidities. Similarly, hypercholesterolemia and hypertension share a strong connection, expressed by a visible line in the network analysis. Both of these conditions also exhibit links and associations with diabetes, indicating a complex dynamic among these comorbidities.

Figure 4:

ADRD Comorbidity Network Analysis

Furthermore, our network analysis has exposed additional connections that go beyond the limitations of the apriori algorithm, which includes preselected support, confidence, and lift criteria. Possibly this lack of constraints allowed us to discover an additional connection not identified in the rules generated by the apriori algorithm. This expanded insight into comorbidity associations is priceless for ADRD healthcare providers. According to the analysis in Figure 4, Angina displays a strong connection to myocardial infarction, which in turn has a connection with congestive heart failure, represented by a notably thick line. This unique chain ultimately leads to atrial fibrillation, demonstrating that these cardiovascular conditions are associated with one another. The network analysis graph portrayed the unique comorbidity associations within ADRD patients. Utilizing network analysis offers a comprehensive and visually intuitive tool for healthcare professionals to better understand MCC associations in ADRD patients.

Discussion

The primary goal of this research was to explore prevalent comorbidities that are diagnosed together in patients who have ADRD and to provide the probability of a patient developing another comorbidity given the diagnosis of an associated comorbidity. Through the apriori algorithm, we found a strong association between urinary and bowel incontinence in ADRD patients, even though those diseases are not the most prominent in the NACC data (Figure 1). Additionally, this association seems unexpected as the comorbidities are involved with different physiological systems - urinary and digestive. However, in our data, the commonality shared by bowel and urinary incontinence patients is that they are both diagnosed with ADRD.

ADRD is characterized by degeneration of brain tissue, including loss of nerve cells, accumulation of beta-amyloid - an abnormal protein, and the development of neurofibrillary tangles²⁶. Urinary incontinence is the inability of a person to control their bladder or urination²⁷ and similarly, bowel incontinence is the inability to control bowel movements²⁸. The urinary and bowel incontinence association calls for further research, particularly in brain signaling, prompting a person’s body to urinate or complete a bowel movement. The Pontine Micturition Center (PMC) is a functional area within the pons which are located in the brainstem. The PMC is responsible for prompting urination²⁹. On the other hand, the pudendal nerve is the peripheral nerve that originates from the sacral spinal cord and it initiates bowel movements as it manages the reflexive control of the external anal sphincter and other pelvic floor muscles³⁰. Although the co-occurrence of both urinary and bowel incontinence remains less obvious, a physical connection between the spinal cord and pons is very obvious in Figure 5, illustrating the interconnected regions of the brain³¹.

Figure 5:

Drawing of the brain anatomy showing the proximity of the spinal cord and pons(Adapted from Terese Winslow Medical And Scientific Illustration, https://www.teresewinslow.com/head-and-neck)³¹.

According to a study by Malykhina et al (2012), “There are ample clinical and experimental data that show interactions between lower urinary (LUT) and lower gastrointestinal (GI) activity. Clinically, LUT dysfunctions often coincide with gastrointestinal dysfunctions and vice-versa…The exact mechanisms and the timespan of these interaction changes are still largely unknown”³². While the urinary and digestive systems are separate physiological systems with distinct functions, the proximity of the signaling epicenters for both urination and bowel movements suggests a potential avenue for understanding the association, especially when considering the neuropathological impacts of ADRD on a person’s brain.

Conclusion

Our research has generated multiple associations of MCCs in patients with ADRD through associated rule mining. By employing heatmaps and network analysis graphs, we visualized the intricate patterns of co-occurrence among various comorbidities in ADRD patients. Our results provide a foundation for more informed medical decision-making and tailored patient care strategies.

While the Apriori algorithm yielded strong associations, it’s essential to consider the limitations of this approach. The algorithm identifies associations and co-occurrences, but it doesn’t inherently establish causal relationships. To address this, future research could explore incorporating the time of diagnosis. By integrating information about the time of diagnosis, we can enhance our understanding of causality, enabling the development of more robust causal inference models. This approach would lead to the creation of more robust models for drawing causal inferences. Such development would not only boost our capacity to forecast future comorbidities but also deepen our insight into the fundamental mechanisms influencing MCCs.

The significance of our discoveries extends beyond direct patient care, widely impacting both clinical practice and research. The associations detected can help medical professionals in predicting upcoming comorbidities for their patients. This knowledge can drive preventative care and tailored intervention strategies. Furthermore, researchers can leverage these insights to uncover novel pathophysiological relationships between associated comorbidities.

In conclusion, utilizing association rule data mining through the Apriori algorithm is a valuable method for detecting complex relationships within the realm of MCCs in patients with ADRD. Future research endeavors utilizing temporal considerations will provide a deeper level of insight. This research holds the potential to bring about even greater benefits for patients, research advancements, and informed medical decisions.