This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Diabetes, is properly cited. The complete bibliographic information, a link to the original publication on http://diabetes.jmir.org/, as well as this copyright and license information must be included.
Complications from type 2 diabetes mellitus can be prevented when patients perform health behaviors such as vigorous exercise and glucose-regulated diet. The use of smartphones for tracking such behaviors has demonstrated success in type 2 diabetes management while generating repositories of analyzable digital data, which, when better understood, may help improve care. Data mining methods were used in this study to better understand self-monitoring patterns using smartphone tracking software.
Associations were evaluated between the smartphone monitoring of health behaviors and HbA1c reductions in a patient subsample with type 2 diabetes who demonstrated clinically significant benefits after participation in a randomized controlled trial.
A priori association-rule algorithms, implemented in the C language, were applied to app-discretized use data involving three primary health behavior trackers (exercise, diet, and glucose monitoring) from 29 participants who achieved clinically significant HbA1c reductions. Use was evaluated in relation to improved HbA1c outcomes.
Analyses indicated that nearly a third (9/29, 31%) of participants used a single tracker, half (14/29, 48%) used two primary trackers, and the remainder (6/29, 21%) of the participants used three primary trackers. Decreases in HbA1c were observed across all groups (0.97-1.95%), but clinically significant reductions were more likely with use of one or two trackers rather than use of three trackers (OR 0.18,
Data mining techniques can reveal relevant coherent behavior patterns useful in guiding future intervention structure. It appears that focusing on using one or two trackers, in a symbolic function, was more effective (in this sample) than regular use of all three trackers.
Diabetes is a cluster of metabolic conditions characterized by dysglycemia from defects in insulin secretion and/or unhealthy behaviors that cause debilitating complications and death [
There is consensus among researchers and clinical professionals that glucose self-monitoring, exercise, and diet-related health behaviors are important in effective T2DM management and well-regulated serum glucose levels [
Data mining (DM) refers to analytic approaches useful in detecting coherent patterns in large and complex datasets [
As utilization of electronic health records increases in health care, DM becomes more relevant [
In another diabetes study, DM algorithms were used to construct a model that predicted short-term changes in blood glucose [
Health coaching is a promising clinical role that stimulates and supports health behavior change in patients with varying SES, health problems and diagnosed chronic diseases. When connected with 24 hour/day/7 day/week mobile phone-based counseling, health coaching is associated with benefits for individuals affected by uncontrolled T2DM [
Our primary objective was to evaluate associations between the mobile phone monitoring of health behaviors, within a randomized controlled trial (RCT), and clinically significant reductions in glycated hemoglobin (HbA1c).
The RCT protocol was reviewed and approved by the York University research ethics board (Certificate #2012-033), and all patients provided written, informed consent to participate.
The RCT assessed T2DM patients (N=97) assisted by personal health coaches trained in behavior-change theories, practices, and counseling methods (see
CWP use data were extracted from NexJ Systems servers upon trial completion and compiled into .csv files stored on password-protected portable drives. Study participant IDs were matched with software user IDs, as data were anonymized and prepared for analyses.
Flowchart of enrollment.
To discover useful relationships between self-tracker use and HbA1c outcomes, we employed association-rule algorithms software to find coherent relationships in transactions represented by sets of items, termed “frequent item sets.” For example, when customer A buys bread and cheese, and customer B buys bread, cheese and burgers, bread and cheese appear frequently on both shopping lists. Therefore, bread and cheese are associated and they qualify as a frequent item set [bread, cheese]. Support is a term reflecting the measurement of association frequency, as defined by the percentage of observations to which the item sets belong. In our study, support was defined as the number of times an attribute value (such as
We used association rule algorithms to identify all common attributes in participants. The support threshold was fixed at a minimum of 5%, such that item sets were generated that occurred in at least 5% of the sample. We used the a priori association rule algorithm implemented by C language for the Unix/Linux environment.
Among the 97 T2DM study completers, the present analysis considers patients in the experimental group (n=48). As our objective was to evaluate associations between the use frequency of different trackers in the CWP in relation to HbA1c outcomes, we selected the change in HbA1c over 24 weeks (
We proceeded to discretize the attributes to implement the association rule algorithm. For the four tracker attributes (
Discretization of tracker use frequency a-d.
Attribute | Group | Range | Meaning |
Glucose count | 1 | 0-0 | Zero |
2 | 5-24 | Up to once per week | |
3 | 24-120 | 1-4.9 times/week | |
4 | 120-192 | 5-7.9 times/week | |
5 | 192-336 | 8-13.9 times/week | |
6 | 336-520 | 14-21.9 times/week | |
Exercise count | 1 | 0-0 | Zero |
2 | 0-5 | Minimal (>0-4) | |
3 | 5-12 | Once every other week | |
4 | 12-24 | 0.5-0.9 time/week | |
5 | 24-48 | 1-1.9 times/week | |
6 | 48-72 | 2-2.9 times/week | |
7 | 72-96 | 3-3.9 times/week | |
8 | 96-150 | 4-6 times/week | |
Food count | 1 | 0-0 | Zero |
2 | 1-5 | >0-5 over 6 months | |
3 | 5-24 | Up to 1 time/week | |
4 | 24-96 | 1-3.9 times/week | |
5 | 96-168 | 4-6.9 times/week (1/day) | |
6 | 168-336 | 7-13.9 times/week (1-1.9/day) | |
7 | 336-504 | 14-20.9 times/week (2-2.9/day) | |
8 | 504-720 | 21+ times/week (3+/day) | |
Generic count | 1 | 0-0 | Zero |
2 | 0-24 | 1 time/week | |
3 | 24-96 | 1-3.9 times/week | |
4 | 96-168 | 4-6.9 times/week (0.5-1/day) | |
5 | 168-336 | 7-13.9 times/week (1-2/day) | |
6 | 336-672 | 14-17.9 times/ week (2-4/day) | |
7 | 672-840 | 18-34.9 times/ week (4-5/day) | |
8 | 840-1200 | 35-50 times/week (5-7/day) |
aAttribute=type of tracker used.
bGroup=discretization group.
cRange=frequencies of use for allocation to group.
dMeaning=frequency of use in terms of use per week/day.
For the
The above pre-processing approaches produced redundancies in the input data for the association rule algorithm. Therefore, we applied three approaches to post-processing after the associations were generated. First, if a frequent item set contained only
Second, for a frequent item set, if
In this frequent item set, the
Third, after considering the above two situations, redundant item sets existed in the result. For example,
Pre- and postprocessing were implemented by Python using the Linux environment. Attribute selection and categorization were used to determine how system use was associated with change in HbA1c per participant. A minimum clinically significant change approach was used to determine what proportion of participants demonstrated 0.5% (5.5 mmol/mol) or greater reductions of HbA1c and used trackers at variable levels of intensity [
Last, a Fisher’s exact test for count data was applied to determine the statistical relationship between use of all three trackers, use of one or two trackers, and HbA1c reductions. For this test, the 10 intervention participants who used the software but did not have a clinically significant reduction in HbA1c were used as a comparison group. Results were considered significant at the
In total, 48 intervention patients completed the 24-week trial. Only 39 patients used the CWP software to track health behaviors. Of this software software-user sample, 29 reduced their HbA1c measure by clinically significant levels at trial conclusion, defined as 0.5% (5.5 mmol/mol) or greater (see
Subject demographics.
|
SOI |
Full intervention sample (n=48) | |
Age in years, mean (range) SD | 53.4 (26-68) 10.7 | 53.1 (26-74) 10.9 | |
|
|||
|
Baseline | 8.88% (73.6) 1.30 | 8.69% (71.5) 1.32 |
|
6 months | 7.52% (58.7) 0.95 | 7.88% (62.6) 1.17 |
|
Reduction (baseline to 6 months) | 1.36% (14.9) | 0.81% (8.9) |
|
|||
|
Male | 11 (33) | 17 (35) |
|
Female | 18 (66) | 31 (65) |
|
|||
|
Less than high school | 5 (17) | 10 (21) |
|
High school | 14 (48) | 17 (35) |
|
College diploma or vocational training | 6 (21) | 11 (23) |
|
University degree | 4 (14) | 8 (17) |
|
Did not disclose | 0 (0) | 2 (4) |
|
|||
|
Owns a car | 10 (34) | 19 (40) |
|
Has access to car | 8 (28) | 9 (19) |
|
No access to car | 11 (38) | 19 (40) |
|
Not disclosed | 0 (0) | 1 (2) |
|
|||
|
Unemployed | 10 (35) | 16 (33) |
|
Student | 1 (3) | 3 (6) |
|
Part-time | 0 (0) | 1 (2) |
|
Full-time | 8 (28) | 13 (27) |
|
Retired | 3 (10) | 6 (13) |
|
Self-employed | 5 (17) | 6 (13) |
|
Work in the home (take care of children) | 2 (7) | 2 (4) |
|
Not disclosed | 0 (0) | 1 (2) |
|
|||
|
0-9999 | 4 (14) | 9 (19) |
|
10,000-25,000 | 8 (28) | 10 (21) |
|
25,000-50,000 | 7 (24) | 12 (25) |
|
50,000-75,000 | 2 (7) | 3 (6) |
|
75,000-up | 3 (10) | 4 (8) |
|
Did not disclose | 5 (17) | 10 (21) |
As seen in
The Fisher’s exact test for count data indicated subjects who used the software (n=39) were more likely to achieve a clinically significant reduction in HbA1c if they used one or two trackers than if they used all three trackers (OR 0.18,
Tracker usage pattern.
Mobile phone tracker | Users, n | Mean reduction of HbA1c, % | SD |
Food only | 2 | 1.95 | 1.20 |
Glucose only | 7 | 1.74 | 1.00 |
Exercise only | 0 | 0 | 0 |
Glucose + exercise | 11 | 0.97 | 0.38 |
Glucose + food | 3 | 1.07 | 0.45 |
Food + exercise | 0 | 0 | 0 |
Glucose + food + exercise | 6 | 1.55 | 0.49 |
Fisher’s exact test for count data on three trackers versus lessa.
|
Achieved clinical HbA1c reduction (≥0.5%) | Did not achieve clinical HbA1c reduction (<0.5%) |
All 3 trackers | 6 | 6 |
<3 trackers | 23 | 4 |
aOR 0.18 and
Breakdown of smartphone usage group.
Food tracker.
Blood glucose tracker.
Exercise tracker.
The glucose tracker was the most frequently used CWP function. Altogether, 22.9% (11/48) of subjects tracked their blood glucose with the software 1-4.9 times per week
The food tracking function followed a similar pattern, although with slightly less frequent use. We found 9/48 (18.8%) of participants who achieved at least a 0.5% reduction in HbA1c used the food tracker a minimal amount (one time or less per week), while 7/48 (14.6%) used the system 1-3.9 times per week, and 4/48 (8.3%) used the system 4-6.9 times per week (see
The least used tracker (and the only tracker that was never singularly used) was the exercise tracker, which was used by 7/48 (14.6%) of the intervention participants between 0.5-0.9 times per week (
Glucose tracker use.
HbA1c diff. | Glucose |
Glucose |
Glucose |
|||
Users, % | n | Users, % | n | Users, % | n | |
0.1 | 29.2 | 14 | 12.5 | 6 | 12.5 | 6 |
0.2 | 27.1 | 13 | 12.5 | 6 | 12.5 | 6 |
0.3 | 25 | 12 | 12.5 | 6 | 12.5 | 6 |
0.4 | 22.9 | 11 | 12.5 | 6 | 12.5 | 6 |
0.5 | 22.9 | 11 | 12.5 | 6 | 10.4 | 5 |
0.6 | 20.8 | 10 | 10.4 | 5 | 10.4 | 5 |
0.7 | 18.8 | 9 | 10.4 | 5 | 10.4 8 | 5 |
0.8 | 16.7 | 8 | 10.4 | 5 | 10.4 | 5 |
0.9 | 16.7 | 8 | 10.4 | 5 | 10.4 | 5 |
1.1 | 14.6 | 7 | 10.4 | 5 | 8.3 | 4 |
1.2 | 10.4 | 5 | 8.3 | 4 |
|
|
1.3 | 10.4 | 5 | 8.3 | 4 |
|
|
1.5 | 8.3 | 4 | 8.3 | 4 |
|
|
Food tracker use.
HbA1c diff. | Food up to 1x/wk | Food 1-3.9x/wk | Food 4-6.9x/wk | |||
Users, % | n | Users, % | n | Users, % | n | |
0.1 | 20.8 | 10 | 16.7 | 8 | 8.3 | 4 |
0.2 | 18.8 | 9 | 16.7 | 8 | 8.3 | 4 |
0.3 | 18.8 | 9 | 14.6 | 7 | 8.3 | 4 |
0.4 | 18.8 | 9 | 14.6 | 7 | 8.3 | 4 |
0.5 | 18.8 | 9 | 14.6 | 7 | 8.3 | 4 |
0.6 | 16.7 | 8 | 12.5 | 6 | 8.3 | 4 |
0.7 | 16.7 | 8 | 10.4 | 5 | 8.3 | 4 |
0.8 | 16.7 | 8 | 10.4 | 5 | 8.3 | 4 |
0.9 | 14.6 | 7 | 8.3 | 4 | 8.3 | 4 |
0.11 | 14.6 | 7 |
|
|
|
|
0.12 | 8.3 | 4 |
|
|
|
|
0.13 | 8.3 | 4 |
|
|
|
|
Exercise tracker use—Part 1.
HbA1c diff. | Exercise up to 1x/wk | Exercise <5x in 6 months | Exercise 1x every other wk | |||
Users, % | n | Users, % | n | Users, % | n | |
0.1 | 16.7 | 8 | 16.7 | 8 | 14.6 | 7 |
0.2 | 16.7 | 8 | 14.6 | 7 | 14.6 | 7 |
0.3 | 14.6 | 7 | 14.6 | 7 | 14.6 | 7 |
0.4 | 14.6 | 7 | 12.5 | 6 | 12.5 | 6 |
0.5 | 14.6 | 7 | 12.5 | 6 | 12.5 | 6 |
0.6 | 12.5 | 6 | 12.5 | 6 | 8.3 | 4 |
0.7 | 12.5 | 6 | 12.5 | 6 | 8.3 | 4 |
0.8 | 12.5 | 6 | 12.5 | 6 | 8.3 | 4 |
0.9 | 12.5 | 6 | 10.4 | 5 | 8.3 | 4 |
0.11 | 12.5 | 6 | 10.4 | 5 | 8.3 | 4 |
0.13 | 8.3 | 4 | 10.4 | 5 |
|
|
0.15 |
|
|
8.3 | 4 |
|
|
Exercise tracker use—Part 2.
HbA1c diff. | Exercise 0.5-0.9x/wk | Exercise 1-1.9x/wk | Exercise 2-2.9x/wk | |||
Users, % | n | Users, % | n | Users, % | n | |
0.1 | 14.6 | 7 | 8.3 | 4 | 8.3 | 4 |
0.2 | 14.6 | 7 | 8.3 | 4 | 8.3 | 4 |
0.3 | 14.6 | 7 |
|
|
8.3 | 4 |
0.4 | 14.6 | 7 |
|
|
8.3 | 4 |
0.5 | 14.6 | 7 |
|
|
|
|
0.6 | 14.6 | 7 |
|
|
|
|
0.7 | 12.5 | 6 |
|
|
|
|
0.8 | 12.5 | 6 |
|
|
|
|
0.9 | 10.4 | 5 |
|
|
|
|
0.11 | 8.3 | 4 |
|
|
|
|
Tracker usage group by clinical HbA1c reduction (≥0.5%).
In the emergent field of mobile phone–assisted health coaching, a recurring question is whether the expense of integrating mHealth technology justifies the benefits. This highlights the need to further investigate for whom (which subpopulations) these technologies are useful and which technologies are most useful for them. In addressing these questions, this study pilots a method of investigating RCT participants, focusing on those who derived clinically significant benefit from participation
By employing attribute categorizations, tracker-use frequencies were determined per subject in an SOI of significant users who also derived clinically significant glucose reductions. This dataset enabled comparison of subjects who used single trackers versus multiple trackers (two or three trackers). Descriptive results indicated that in subjects achieving a 0.5% HbA1c reduction or better, singular use of the glucose and food monitoring was undertaken, while, in contrast, in the same group, no singular use of the exercise tracker was undertaken. In dual-tracker use, glucose and exercise trackers were employed by more subjects than glucose and food tracking. Interestingly, food and exercise (as dual) trackers were not used by any of the subjects who met criteria for 0.5% HbA1c reduction or greater. Last, all three trackers were used by 12.5% (6/48) of subjects and were associated with a substantive HbA1c mean reduction (1.55% or 16.9 mmol/mol).
The Fisher’s exact test for count data indicated that for subjects who used the software (n=39) and used all three trackers, there was a significantly lower likelihood of achieving a clinically significant reduction in HbA1c than for those who used a lesser number (ie, one or two trackers) (OR 0.18,
Data mining is often used to process large amounts of data. One limitation of the pilot application of data mining in this study was the relatively small user sample size. Nonetheless, the association rule algorithm technique offers a foundation with which to study larger datasets of mHealth tracking technologies as they become available. In terms of diabetes intervention, this was an RCT of typical size (48 intervention participants) and, altogether 10,695 uses of the mobile phone app were analyzed (about 62 uses per month per participant who used the software). Although future DM studies may address larger datasets, this pilot demonstrated application in an RCT dataset within which >10,000 app uses were analyzed (averaging ~1 use per day).
In summary, this study points to a future when the mobile monitoring of health behaviors will increase and provide digital signals representing engagement in discrete behaviors and daily-weekly-monthly outcomes. Whereas previous associations between counseling and outcomes were difficult to obtain and often based on retrospective self-report, mobile phone monitoring offers ongoing records that precisely reflect status improvements, their stability, and fluctuations (eg, relapsing patterns). Altogether, with the increasing collection of wearable data, we may derive a quantifiable perspective on health changes that instructs the patient and health coach in improving chronic disease management.
Connected Wellness Platform
data mining
glycated hemoglobin
randomized controlled trial
subpopulation of interest
type 2 diabetes mellitus
The authors would like to thank NexJ Systems Inc. for their partnership in this trial and for the use of the Connected Wellness Platform as a clinical research tool. Funding was provided by the Public Health Agency of Canada and the Federal Development Agency of Southern Ontario. We offer special thanks to the staff of the Black Creek Community Health Centre and trial participants from the Jane-Finch community of Toronto, Ontario.
Joel Katz is supported by a Canadian Institutes of Health Research Canada Research Chair in Health Psychology.
The authors would like to acknowledge with sadness the untimely passing of study co-author, colleague, and friend, Dr. Nicholas Cercone. Dr. Cercone’s expertise and mentorship on data mining theory and technique was invaluable. His inspiring and supportive presence will be deeply missed.
None declared.