IEEE Consumer Electronics Magazine - January/February 2020 - 33

Hence, we used the independent variables and
the BGL values of datasets from each PPG signal
of each participant to fit a random forest regression model and adaboost model.

Table 2. Results of prediction accuracy for adaboost and
random forest regressions.
Model type

Participant

Adaboost

Random forest

79.5%

79.9%

86.0%

87.5%

85.7%

78.7%

79.7%

85.2%

87.1%

85.4%

84.1%

85.9%

85.5%

89.7%

90.0%

81.8%

82.2%

87.8%

86.5%

Global

Experimental Approach
Each participant's data were experimentally
validated using the tenfold cross-validation and
leave-one-out cross-validation methods.13 If one
participated in nine measurement events, we used
all nonrepeated sampling sets, eight for training
and one for testing. The mean absolute error
(MAE) of validation sets was then estimated, and
the accuracy of models was determined. It is
important to bear in mind that each participant
had different levels of BGL. Thus, the accuracy of
each personalized model was determined by the
remainder of 1 minus the percentage of MAE in
the maximum value of BGL for each participant.

RESULTS AND DISCUSSION
Model Performance under Data Constraints
Table 2 summarizes the global and personalized model accuracies using the adaboost and
random forest regression models, respectively.
In general, the models performed satisfactorily
under the data constraints. The global model
accuracy for both methods was about 80%. The
average of the personalized models for both
methods increased to 85%. However, there were
some variations across different models within
an individual, and the greatest difference in the
accuracy of the random forest model was above
10% between participants G and C. The model
performance depended on different participant
identity instead of on the modeling.
Global versus Personalized Models
All the personalized models were more accurate than the global model for both adaboost and
random forest methods except for participant C.
Both method's accuracy improved a minimum of
2.3% and a maximum of 10.2% compared to the
global model.

Personalized

better than the adaboost models by about 1%,
although the adaboost model's accuracy for participants E, F, and I was slightly higher than that
of the random forest model.

CONCLUSION
Given the high accuracy of noninvasive BGL
estimates attained in this study (maximum of 90%
accuracy and over 80% in most cases), it is possible for some diabetic patients to estimate BGL
using the personalized models with limited PPG
data. However, we found that the data contained a
significant level of random noise, which, to some
extent, affected the model performance. The proposed noninvasive glucose monitoring method
and instrument requires greater clinical evaluation
with additional analyses to assess this technique's
applicability to a broader population. Nevertheless, the current study demonstrates the great
potential to produce a wearable and easy-to-use
glucometer to monitor diabetes health in motion.

& REFERENCES
1. P. Zimmet, K. G. M. M. Alberti, and J. Shaw, "Global
and societal implications of the diabetes epidemic,"
Nature, vol. 414, pp. 782-787, 2001.

Random Forest versus Adaboost Regressions
The model accuracy results for the random
forest and adaboost models were very similar.
The random forest models generally performed

January/February 2020

2. L. Chen, D. J. Magliano, and P. Z. Zimmet, "The
worldwide epidemiology of type 2 diabetes mellitus-
Present and future perspectives," Nature Rev.
Endocrinol., vol. 8, pp. 228-236, 2011.

IEEE Consumer Electronics Magazine - January/February 2020

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - January/February 2020