Using Quantitative Medicine to Control Type 2 Diabetes 

Glucose – PPG

Section 4: Glucose – PPG

Date: 10/24/2016  13:00

Approximately 9 to 10% of the populations in Eastern Asia, Europe, and USA have diabetes, Latin America has about 8.5%, while Africa has about 5% of its people with this disease.  Particularly in developing nations, they do not have sufficient resources to do frequent glucose testing, let alone A1C testing.  An easily accessible and free tool for early prediction will be very helpful for diabetic patients to control their disease, especially in the developing nations.  In addition, Apple devices are too expensive for many to purchase.  Another fact is that most senior citizens are more likely to be diabetic candidates and they may not be as in tune with modern technology and APP’s.  Therefore, my charitable medical research organization, eclaireMD Foundation, has developed both iPhone-based and PC web-based tools for different type of users.  The PC version of my tools are much easier to access by patients in developing countries and elderly populations.  My weight and glucose prediction tool can be downloaded for free from Apple’s APP Store by searching for the keyword of eclairemd, and then click the “Wellness” product on the screen.

On August 3, 2010, my lab test results showed my A1C as 10.0%, ACR as 116 mg/mmol, and triglyceride level as 1,161 mg/dL.  After receiving a serious warning from my primary physician, I decided to change my overall lifestyle.  I began collecting my health and lifestyle data on January 1, 2012.  To date, approximately 5-years’ worth of complete data has been collected.  In my Amazon’s cloud database, I have stored about one million data points regarding my health and lifestyle, which includes original inputted data, system calculated data, and AI generated data.  In Figure 4-1: Daily glucose and 90-days average glucose, within 5.5 years, my glucose levels varied between 86 mg/dL and 227 mg/dL with an average value of 127 mg/dL.  In Figure 4-2: My lab-test A1C values were between 6.3% and 7.1%, with an average value of 6.6%, and my mathematical simulated A1C values (between 6.3% and 8.3% with an average value of 7.3%) had a prediction accuracy of 89%.  This reduced accuracy resulted from my build-in “conservative” design in order to provide patient early warning.

Figure 4-1: Daily Glucose and 90-days Average Glucose (1/3/2012 – 7/24/2017)
Figure 4-2: Mathematical Simulated A1C and Lab-tested A1C Comparison

The laboratory A1C test result reflects the average glucose values for the past 90 days. The eclaireMD tool calculates a mathematically simulated A1C value based on the user’s daily glucose input data. However, this mathematically simulated A1C value has a built-in AI to automatically customize the calculation according to the changes of the user’s biomedical conditions.  Frequent calibration by inputting laboratory test results can improve the accuracy of the mathematically simulated A1C results. My mathematical A1C curve spectrum has completely “enveloped” the actual lab-tested A1C data from 2012 to 2017. This means that my simulated A1C can provide me with an upper-bound early warning before I get tested for the laboratory A1C.

Figure 4-3 lists all of my lab-test A1C values. Although I have taken into account both the lifespan of red blood cells and developed different linear and nonlinear mathematical models, glucose still has a somewhat unpredictable output value due to a highly nonlinear biomedical body system that it changes with time. There are many elements that affect glucose values and I will discuss more of my research results in following sections regarding these elements.

Figure 4-3: List of My Past Lab-tested A1C Values (2000 – 2017) and Corresponding Mathematically Simulated A1C Values

From 2012 through 2017, the overall A1C curve remains at a somewhat steady state, i.e. around 6.6%, the most important factor of controlling my diabetic condition is that I have been decreasing the dosage of my medication for over two years’ time frame and completely removed all of my diabetes medications in December of 2015. Between 2012 and 2013, I was taking three different diabetes medications, such as Januvia 100 mg, Actoplus Met 15 mg/850 mg, and Metformin 2000 mg.  I started to decrease the number of drugs along with reducing the dosage amounts from the beginning of 2014. By December 8, 2015, I completely stopped taking my diabetes medication.  I must point it out that during this period, I witnessed different degrees of “withdrawal symptoms” which were similar to a drug addict’s detoxification process. Within one month of removing or reducing medication, my glucose chart fluctuated greatly, with many ups and downs, without any clear and reasonable explanation.

I have beening a long term diabetic patient for almost 20 years. I realized that have knowledge, a reliable tool, and strong willpower are the three main keys to control this disease. During the past 4 years, I have applied my acquired medical knowledge and developed my own tools to help control both my weight and glucose. The main tools are my weight predictor and my post-meal glucose predictor, which were developed on April 11, 2015 and June 1, 2015 respectively. I also developed my fasting glucose predictor on July 31, 2017. It should be noted that the metabolism model (MI and GHSU) I developed in 2014 is the foundation for improvements in my overall health. When I could predict my weight and glucose beforehand, it became much easier for me to adjust the amount and quality of my food, along with the frequency and intensity of my exercise. In other words, if I can wisely adjust my input parameters, my output values will most likely be automatically adjusted for the better. It should also be noted that the body is organic (nonlinear) and dynamic (changes with time), so we have to constantly monitor for signs of change.

Two fundamental rules must be followed in terms of using these prediction tools. First, I had to follow the prediction model’s precise suggestions regarding how to input those value. Second, after measuring my weight or glucose, I was not allowed to go back to randomly readjust my original input data in order to increase the accuracy of prediction (unless there were some new findings or facts that I had just learned or realized from applying this prediction experience). Some degree of AI has been built into the system as well, but the entire biomedical system must be continuously observed and monitored along the way. That is why I created a non-profit medical research foundation to persistently work in depth on this subject, even after my passing.

I have conducted an accuracy study of the PPG prediction for two adjacent periods of 115 days each: period A from 3/1/2016 to 6/23/2016 and period B from 7/1/2016 to 10/23/2016. All of my post meal glucose values during period A were obtained via the glucometer, while during period B, only 33% of my post meal glucose values were measured. The remaining 67% was based on the predicted PPG as the actual PPG values. Whenever I ate my meals at restaurants where no reliable nutrition information was readily available, or I cooked a new dish at home using new food materials, I would measure my post meal glucose by the finger piercing method. Period A of 100% measured PPG via the glucometer has a slightly lower accuracy rate (~97%) than period B of 30% measured PPG via the glucometer (~98%). In theory, if 100% of the actual PPG values are based on the predicted PPG data (i.e., 0% of PPG values via the glucometer), then the accuracy rate will be 100%. In addition, the correlation coefficients between predicted PPG and actual PPG of both periods are quite high (>76%), see Figure 4-4: Accuracy analysis of predicted PPG values.

Figure 4-4: Accuracy analysis of predicted PPG values

The conclusion from this experiment is that I could eliminate the finger-piercing method and still get a very high accuracy rate (>97%) on my post meal glucose results. This is an important finding that my PPG prediction method is proven to be extremely reliable for most Type 2 diabetes (T2D) patients, whose average daily glucose level fall between 100 mg/dL and 400 mg/dL. They will be able to use my PPG prediction tool to control their diabetes. This was my original goal to conduct the reliability study to remove the burden and cost associated with finger piercing and test strips. Hopefully, this glucose prediction model and tool can reduce both cost and pain for worldwide diabetes patients, especially for patients in underdeveloped nations.

No significant Correlation between FPG and PPG
I have conducted a statistical analysis of the correlation coefficient (r) and coefficient of determination (r2: r square) of 90-days moving average values of FPG vs. PPG during the complete period of 6/1/2015 through 7/26/3017. The results of both r (-3.3%) and r2 (0.11%) from these two sets of data are very low, which means that the FPG values almost have no correlation with PPG values at all, see Figure 4-5. The outcome makes a lot of sense from the biomedical point of view. Fasting glucose is the combination effect of glucose produced by the liver and insulin produced by the pancreas during sleeping hours. On the other hand, during the awakening hours, both the diet control and exercise have direct impact on PPG values.

Figure 4-5: No correlated relationship between FPG and PPG during a period (6/1/2015 – 7/26/2017)

Weighted contribution factor analysis of FPG and PPG on A1C
Important dates and events regarding my glucose are as follows:
(1) I started to record and collect my PPG data on 1/1/2012;
(2) I started to record my FPG values on 6/1/2015;
(3) I stopped to take my diabetes medication on 12/8/2015
(4) My FPG values suddenly jumped from ~110 mg/dL to above 140 mg/dL on 11/23/2016 and remained high for 5 months. My FPG values dropped again on 6/5/2017 after my weight was reduced.

By now, there are sufficient glucose data in my cloud server, so I can conduct a big data analytics on the composite impact of FPG and PPG on A1C value. First of all, I defined a new term, Adjusted A1C, as an interim processed variable to conduct this study. The equation is:
Adjusted A1C = a x FPG + b x PPG & b = 100% – a

where a and b are A1C’s weighted contribution factor (%) of FPG and PPG, respectively.

I have used the data from the period from 12/8/2015 through 7/27/2017 to eliminate the data impact from medication. During this period, I have collected 8 sets of lab-tested A1C data (test performed quarterly) over 2 years, which are used as the base for this comparison study. I then chose 4 different sets of weighted contribution factors for data analysis:
a = 0%, 15%, 25%, 50%;
and the corresponding
b = 100%, 85%, 75%, 50%.

This 4 sets of factors are to calculate 4 adjusted A1C curves in order to compare them against the eclaireMD simulated mathematical A1C curve and lab tested A1C data points, and the results are shown in Figure 4-6. By this comparison analysis, I observed the following phenomena:
(1) The contribution ratio of FPG 25% and PPG 75% provides the highest accuracy 100% and correlation 99.8% between adjusted A1C and simulated A1C. This make a perfect sense since simulated A1C uses 1 FPG and 3 PPG values each day as its base of calculation plus some degrees of AI modification.
(2) The more deviation of the contribution ratios away from the set of 25% & 75%, the lower accuracy and correlation become, however, they are still quite high in all cases.
(3) I deliberately put a case of both FPG and PPG at 50% level, i.e. they have equal contribution on making A1C value. The turning point date is 11/23/2016, PPG was high prior to that date and FPG was higher after that date. On adjusted A1C curve, you can see that simulated curve is higher than adjusted prior to 11/23/2016 and vice versa for after 11/23/2017.

Figure 4-6: 4 sets of adjusted A1C values in comparison with eclaireMD simulated A1C and 9 different lab-tested A1C data

In late 2016, I conducted an analysis of lower FPG contribution factors (below 12%) and its results are shown in Figure 4-7, the statistical Comparison Study of A1C Values based on 3 sets of A1C values: eclaireMD predicted, lab-tested, and Adjusted. The results show that the case of 0% of FPG and 100% of PPG is the closest combination to eclaireMD predicted A1C, which also possess the highest correlation coefficient (60%). However, the case of 11.4% of FPG and 88.6% of PPG mix gives a perfect match between adjusted A1C and lab-tested A1C. This means that, during the period of 8/2/2015 through 1/3/2017, my weighting contribution factors of FPG and PPG to A1C value are probably falling within the range of FPG < 10% and PPG > 90%. This lower FPG contribution percent is due to my average lower fasting glucose (~110 mg/dL) prior to 11/23/2016. Please note that I have adopted the common medical community’s convention of using the lab-tested A1C as the basis of all of my glucose statistical comparison study.

Figure 4-7: Statistical Comparison Study of A1C Values based on (1) eclaireMD predicted A1C, (2) lab-tested A1C, and (3) Adjusted A1C

During this 20-months period since 12/8/2015, my average lab-tested A1C is 6.6% (real lab-tested values fluctuating between 6.4% and 6.7%) without taking any diabetes medication. Furthermore, the most probable weighted contribution factors: <11.4% of FPG and >88.6% of PPG, also indicate that my type 2 diabetes is very well under controlled by applying a better lifestyle management without taking any diabetes medications. However, this observation is based on my highly accurate PPG prediction model and my near-constant low FPG (~110 mg/dL) prior to 11/22/2015. My FPG values suddenly increased to 158 mg/dL on 11/23/2016 and remained relatively high through 6/3/2017. I have discovered that, at present day of 7/27/2017, the best mix ratio of FPG & PPG has changed to 25% & 75%.

Figure 4-8: Two periods’ A1C data comparison on actual lab testing dates shows that both of my simulated and adjusted A1C curves are “enveloping” most of my lab-tested A1C data. This is what I called “conservative design” but it still remains a high accuracy of predictions. This proves that a patient can use my tool to predict his or her A1C results before lab tested day.

Figure 4-8: Two periods’ A1C data comparison on actual lab testing dates