What Causes Patient Non-Adherence to Self-Monitoring of Blood Glucose?
- 1 Introduction and Motivation
- 2 Team Members
- 3 Dataset
- 4 Proposal Approach
- 5 Findings
- 6 Tool Critiques
- 7 Conclusion
- 8 Visualization Ideas
- 9 Reference
Introduction and Motivation
Patient non-adherence can result in significant risks to both the patient and the treatment provider. In the context of health care, non-adherence is defined as patients failing to follow the rehabilitation program or the advices (e.g. self-monitoring, medication, exercise, etc.) that are given by the provider. In treating chronic diseases like diabetes, self-management plays an important role, which requires patients to adjust their daily activities according to the indicators measured by medical instruments. Based on the thesis by Dr. Vaughn-Cooke  and the corresponding data, we use Tableau  and HCE  to visualize the blood glucose, daily self-monitoring of blood glucose (SMBG) testing frequency and daily adherence values of 99 diabetes patients over a 60-day period to find out the critical factors for the non-adherence to self-monitoring of blood glucose. Also, we analyze self-reported surveys and medical records to explore their relationship to the non-adherence.
In 2010 from June to September, 99 type 1 and 2 diabetes patients completed the study by Vaughn-Cooke . In the study, data were collected in three methods: (1) surveys that patients were required to fill out before and after the study, (2) medical records including health vitals and diabetes complications data obtained by the Penn State Hershey Diabetes and Obesity Institute Registry and (3) actual SMBG frequency with the blood glucose value downloaded from the glucose meter. The dataset and its description file can be accessed at:
- Visualize the relationship between the daily adherence and daily blood glucose values to emphasize the significance of treatment adherence for diabetes patients.
- Use visualization methods to explore the impacts of each variable of SMBG testing, e.g., days with or without testing, the outlier of blood glucose value, intended daily frequency, etc.
- Clustering the patients according to their self-reported information in the survey (Performance Shaping Factors - PSFs), e.g., age, race, education, weight, duration years of diabetes, etc., and use hierarchical data visualization techniques including treemaps  and HCE  to explore the relationship between those attributes and non-adherence.
- Try to design different visualization for provider-accessible Electronic health Records and patient-accessible Patient health records.
- Critique the tools: Tableau and HCE.
To provide a patients overview for treatment providers, we designed a visualization utilizing box plot (Figure 1). First, in the overview, by examining the plot of patients on the left side, we find that patients with high median daily glucose are likely to have low adherence. Second, we find that the adherence level of each patient tends to be stable, which means the patients are likely to keep their pace of daily self-monitoring of blood glucose. Third, from the color and size of each point, we find that the green points are always larger than the red ones, which reveals the insight that patients with higher SMBG testing frequency are likely to have higher treatment adherence. To verify our general findings, we drill down to visualize the temporal aspects.
Insight 1: Low Adherence, High Glucose
To drill down, we randomly filter out seven patients (#12, #16, #31, #37, #44, #52 and #60) and use Glucose Flow to visualize the temporal aspects of their glucose value, adherence and SMBG frequency. In this visualization, we find that patients’ glucose values are likely to increase when their adherence goes down, and vice versa. For instance, in the flow of patient #37, #31, #12 and #16, we can find a remarkable increase of glucose when the flow becomes red, which means the adherence goes down. In the flow of patient #31, the glucose increase first and then decrease in the following period, while the flow becomes red and changes back to green, which means the adherence goes up and then goes down. Moreover, by examining the rest patients (#52, #44 and #60), we can make the same conclusion. Based on this insight, the patients can use our visualization to monitor the “red” danger points, when their adherence goes down and potentially leads their glucose rise up.
Insight 2: High Frequency, High Adherence
To find out the relationship among glucose value, SMBG frequency and adherence, we compact the Glucose Flow visualization and show all the 99 patients in the same view. We tried to sort the flows either by average glucose value or by SMBG frequency, and surprisingly found that the SMBG frequency exhibited a remarkably high correlation with the adherence. As the flows are decreasingly sorted by average SMBG frequency, the flows at the top are with higher frequency than those on the bottom. And from the color, which encodes the adherence, we can find that the top flows are mostly in green color while the bottom flows are prevalently in red. From this visualization, we can make the conclusion that patients with high SMBG frequency are likely to have high treatment adherence.
Insight 3: Better Educated, Higher Adherence
In order to figure out if education level of the patients affects the treatment adherence, we cluster the patients into different groups based on the following Performance Shaping Factors (PSFs).
- Age: Considering people's habits and biological changes vary a lot during the life span, we divided the patients based on the following criteria according to Ageing on Wikipedia (http://en.wikipedia.org/wiki/Ageing):
- Early adulthood (20-39): 24 counts.
- Middle adulthood (40-59): 43 counts.
- Late adulthood (60-81): 32 counts.
- Education: Prior education is a positive factor on patient's understanding of adherence in the SMBG process. We also divided the patients to three groups:
- Less educated (level 1-2, high school): 43 counts.
- Medium educated (level 3, bachelor): 18 counts.
- Highly educated (level 4-5, graduate school): 38 counts.
- Delivery of Medical Information: Lots of prior studies illustrator that delivery of medical information is an important factor for treatment adherence. This score is obtained from the PSF survey according to Dr. Vaughn-Cooke’s thesis . The patient is asked to fill in 1-4 score for 10 questions. Nevertheless, the score varies from 1 to 31 with an average of 28, which indicates that this information is very subjective compared with the former two.
- Poor delivered (1~25): 37 counts.
- Partly delivered (25~30): 35 counts.
- Mostly delivered (30+): 27 counts.
These three factors answered when and how well a patient is educated to some extent: Did the patient get well educated in recent years? Did the patient have a good understanding of the medical information of diabetes treatment? Finally, we plot TreeMaps to illustrate their influence on adherence.
From the TreeMap, We can easily figure out that 3 groups with a higher average adherence have remarkable features: Both of the first and second group in adherence have the highest education level (4~5); while the highest group's age is among 18~39 and the second group is 40~59. Considering the experiment is conducted in 2010, these people spent their university years in 1970~2010 with a better education in medical care and health. So that could be a vital feature corresponding to treatment adherence. To verify our guess, we move the age classification 10 years younger, resulting in another TreeMap:
Via this visualization, the highly-educated patients in an early-adulthood have the most adherence to the diabetes treatment. However, the patients in middle adulthood with a medium education have the least adherence. A possible reason could be: compared with the least educated group, who would obey more to the doctors' diagnostic instructions, they on the contrary, care less on their diabetes treatment adherence.
Finally, we turn to the subjective score of medical information delivery. Again, we draw a tree map to see if this matches the objective education level:
This time, we found that the highest group has the highest education and highest delivery of medical information while the red group is a little confusing, with medium education level and medium delivery response. The other two red ones are separately the least educated with the highest "delivery" and the medium educated with the least information delivery. For the former group, I think the subjective response made them misunderstand the importance of medical delivery. For the other group, personally I agree that they do not get the significance of treatment adherence.
In conclusion, thanks to the TreeMap, we figure out that when and how well the patient is educated is greatly related to treatment adherence. Generally speaking, the better Educated, the higher adherence.
Insight 4: Psychological Factors and Relationships Do Affect the Adherence
This time, we plot scatter points and employ filters to look into the psychological factors including anxiety, depression, adherence motivation, support from families and interpersonal needs. The color again indicates the average adherence during the past 60 days. The size indicates the number of patients.
From the red squares, we can clearly see that some patients with extremely low motivation; extremely high anxiety and depression are likely to have lower adherence than others. Moreover, non-adherence of some patient may due to extremely low self-motivation.
Accoring to G. Zimet, et.al. 1988 , patients who gain high support from families are likely to have high adherence. This is indeed true for our data. Almost all patients with over 80 score of support have above 75% average adherence to diabetes treatment.
In conclusion, psychological factors and relationships do affect the adherence. For higher adherence, the patient had better has less anxiety and depression; higher motivation and more support from families also helps a lot.
- Hall of Fame
- Strive for consistency: whenever a dimension or a measure is modified in a sheet; the corresponding visualization in the dashboard is automatically updated
- The outputs in PC, web and iPad maintain splendid consistency in Tableau.
- Connecting different databases in Tableau is easy, which make it out-stand from other visualization software.
- Hall of Shame
- It is impossible to create a novel color scheme with more than 2 colors using Tableau. The provided color editor only provide two-color gradients; modifying the center of the palette would change the starting or ending color.
- One variable is either classified as a dimension or a measure. But sometimes, a variable (such as average percentage of adherence) could be regarded both a dimension and a measure. So if we could press "Ctrl" and drag one variable to measure or dimension windows to copy them, it would be much more convenient.
- Hall of Fame
- It can provide great visualization if the data has internal hierarchy and numbers.
- Hall of Shame
- Bug: The visualization disappears when the bar height goes down to a certain value.
I find converting data format and combining two sheets together is a tedious task for TimeSearcher.
In this application project, we investigate into the datasets of 99 patients including treatment adherence; gloclose; education; demographic data; psychology survey data and etc. Finally, we have the following insights:
- Patients' glucose values are likely to increase when their adherence goes down, and vice versa.
- Patients with high SMBG frequency are likely to have high treatment adherence.
- The better educated of the patients, the higher average adherence they have.
- For higher adherence, the patients had better decrease anxiety and depression; try to gain higher motivation and more support from families.
Is that possible to integrate our visualization into the glucometer?
How shall we integrate visualization to assist patients for the treatment decision making process?
-  M. Vaughn-Cooke, “A Multidimensional Information System for Human Reliability Assessment: Applied to Patient Adherence,” Diss. The Pennsylvania State University, 2012.
-  P. Hanrahan, “Tableau Software White Paper-Visual Thinking for Business Intelligence,” Tableau Software, Seattle, WA, 2003.
-  J. Seo and B. Shneiderman, “Interactively Exploring Hierarchical Clustering Results,” Computer (Long. Beach. Calif)., vol. 35, no. 7, pp. 80–86, 2002.
-  T. Asahi, D. Turo, and B. Shneiderman, “Using TreeMaps to Visualize the Analytic Hierarchy Process,” Inf. Syst. Res., vol. 6, no. 4, pp. 357–375, 1995.
-  Zimet, Gregory D., et al. "Psychometric Characteristics of the Multidimensional Scale of Perceived Social Support." Journal of personality assessment 55.3-4 (1990): 610-617.
-  Ben Shneiderman, "Treemap Art", http://treemapart.wordpress.com, 2013.