Using Quantitative Medicine to Control Type 2 Diabetes 

Glucose – FPG

Section 5: Glucose – FPG

Date: 7/7/2017 – 8/13/2017

My Initial Fasting Glucose Study
During the course of data collection and analysis, I have noticed some deviation that exists between tested data and simulated data. My first attempt was to identify the most important elements affecting glucose and A1C. Some of them will be addressed in the following discussions. Around April 2015, I became intrigued by the difference between higher fasting glucose in the morning (Dawn Phenomenon) and lower fasting glucose in the morning. Without possessing a complete and clear knowledge of how the liver and pancreas work in terms of creating and controlling glucose level, I decided to study the fasting glucose via a “physics” research approach: i.e. observe phenomena, collect data, create hypothesis, define inputs, derive governing equation, identify outputs, investigate relationship between inputs and outputs, plug all inputs into the equation and calculate outputs, and validate the accuracy of outputs. If the outputs are wrong, then repeat the process again until hypothesis is proven correct. In modern terminology, this is another way to describe the “macro-view” approach known as ”big data” collection and analysis. By October of 2016, I have collected nearly 1,000 pre-breakfast fasting plasma glucose (FPG) data.

The prediction of FPG (fasting plasma glucose) is very different from the prediction of PPG (postprandial plasma glucose). At first, I decided to use my 90-days average daily glucose as the initial condition for predicting FPG. Please see Figure 5-1: Predicted FPG value based on my preliminary finding of 360-days data. It indicates that, although the daily FPG goes up and down and it is difficult to predict, but after a long period of time, the averaged FPG settles around the 90-days average glucose value. The deviation between the predicted and actual is 1.6% and my predictions have reached to 98.4% accuracy. It should be noted that the above analysis and tentative conclusion were based on data available prior to 10/20/2016. During this period, most of my fasting glucose data are very close to my averaged daily glucose value.

Figure 5-1: Predicted FPG value based on my preliminary finding of 360-days data

My understanding of FPG was suddenly changed on 11/23/2016
I arrived in Honolulu on 11/22/2016 and stayed there for two months. The next morning, my FPG jumped to 158 mg/dL. At first, I thought it was due to traveling and changing living environment; however, it has persistently stayed at a high level, above 140 mg/dL. I was puzzled by the questions related to quantitative characteristics of FPG, such as: What is the exact causes of the surge? How to predict its pattern? How high will it jump? How to control the high FPG from happening again? My past understanding and practice of PPG control via diet and exercise have had very negligible impact on my FPG control, since it is produced in early morning hours when I am sleeping. For the next 3 months, I searched for related information by reading more than 100 papers and articles, asking questions to a few internal medicine doctors, etc. I also conducted investigations of all possible relationships between FPG and all of my collected input elements on the cloud server, and performed numerous statistical correlation analyses. For example, I calculated the current day’s FPG against the same day’s post-breakfast glucose and previous day’s post-dinner glucose. The purpose of these two studies was to see whether FPG contributed to the increase of the same day’s post-breakfast glucose or received a “left-over” impact from the previous day’s post-dinner glucose. Please see those two low correlation coefficients from Figure 5-2: Two correlation studies between FPG and two different PPG values.

Figure 5-2: Two Correlation studies between FPG and same day’s post-breakfast PPG and between FPG and previous day’s post-dinner PPG

I even tried some “tricks” mentioned in some papers and articles, e.g. eating snacks before sleep, chewing a piece of candy during midnight, measuring my glucose data every hour between 2am and 7am (you can image that the quality of my sleep was badly affected). None of them worked. However, I still did not want to retake my diabetes medications. In addition, I refused to take the suggestion from a few papers to visit a physician to get insulin shots. These elevated FPG values finally pushed up my lab-tested A1C from 6.4% to 6.7% (5% increase). By the spring of 2017, I still could not find any clues to develop a useful and accurate FPG prediction model in order to control my higher fasting glucose in the mornings.

My Recent Fasting Glucose Research after 3/17/2017
After collecting 5 more months of higher FPG values (11/23/2016 – 4/30/2017), followed by another 3 months of efforts on reducing my weight (5/1/2017 – 7/30/2017), I finally have sufficient data to investigate my FPG situation. I have chosen a data set in a period of 16 months (about 500 days), from 4/1/2016 to 7/28/2017. During this period, I further subdivided them into two “equal length” sub-periods of 8 months (about 250 days) each. In the first sub-period from 4/1/2016 through 11/23/2016, my average weight was 172 lbs. and the average FPG was 110 mg/dL. However, in the second sub-period from 11/23/2016 through 7/28/2017, my average weight was 176 lbs. and the averaged FPG was about 130 mg/dL. As a result, the averaged FPG has a 20 mg/dL increased amount due to my average weight increase of 4 lbs. Actually, during this entire period (4/1/2016 – 7/28/2017), my minimum weight was 171 lbs. and the maximum weight was 179 lbs. It should be noted that during the second sub-period after 11/23/2016, there were no significant changes in my lifestyle, including food, exercise, stress, sleep, water intake, life regularity, temperature, living environment, etc.

In the past, my correlation analyses were focused on relationship between input categories (e.g. diet, exercise, living environment, etc.) and output categories (e.g. glucose, weight, blood pressure, etc.). I am a professional engineer with strong mathematics, physics, and computer science background. Most of my past 50-years training as an engineer is following this guiding principle: to identify the relationship between inputs and outputs of a system. I have worked diligently trying to solve this particular FPG problem for almost 4 months. As a result, around 3am of March 17, 2017, I dreamt about the possibility of one output category could also be served as the influential input factor of another output category. This influential input category was ”Weight”! This was my “out-of-box thinking” resulting from my persistent digging into the same problem for 4 months (from 11/23/2015 through 3/17/2017). As previously indicated, existing input categories, such as food quantity and exercise amount, were inputs of our weight. By now, I thought about my weight to function as the primary and direct input of my FPG. Both weight and glucose belong to the category of outputs. After breaking out from my previously trained engineering thinking pattern and constraints, I was then able to look into the FPG problem with a totally different new angle.

The summary results of the high correlation of 84% between weight and FPG are shown in Figure 5-3: Weight and FPG during the period of 4/1/2016 through 7/28/2017. In addition, a high correlation can be seen in the charts observed in Figure 5-4, when I plot them out under the selection criteria of weight data of >176 lbs. and FPG data of >130 mg/dL. I realized that, during the past 5 years, I have already controlled my PPG very tightly via diet, exercise, and others. However, during this 4-months winter sub-period, I have eaten too much “between-meal” snacks, e.g. nuts. This eating pattern did not contribute much on my post-meal glucose values; however, it did result in a 4 to 8 lbs. of weight increase which in turn caused the increase of 20 mg/dL on the average FPG. Lesson to be learned here is to watch out for snacks.

Figure 5-3: Weight and FPG during the period of 4/1/2016 through 7/28/2017
Figure 5-4:Strong correlation for weight > 176 lbs. and FPG > 130 mg/dL (4/1/2016–7/28/2017)

Of course, the increased FPG values will most likely push up my A1C values. As I mentioned earlier, my FPG contributes approximately 25% and my PPG contributes about 75% to the predicted A1C value. Therefore, I verified this FPG impact on my A1C from my two following A1C lab tests (4/9/2017 and 6/1/2017), both with 6.7% value, 0.3 increased value from previous A1C of 6.4% (5% increase).

Correlation comparison between Weight and FPG vs. Weight and PPG
My interpretation of the low correlation coefficient (~1%) between post-meal glucose and weight is due to the fact that weight is only a second-level influence on PPG. The other two lifestyle factors, diet via carbs and sugar control (r be equal to 60%) and regular exercise (r be equl to -27%), are true primary factors for PPG. Our body weight (one of body’s output) is controlled by our overall lifestyle, food quantity and exercise (two of body’s inputs) are included as well. During the daytime, we can manipulate both diet and exercise to control our post-meal glucose level to an optimal level. However, during sleeping hours, it is a completely different story. For most cases, during sleep time, other lifestyle factors cannot be regulated by us and therefore, our brain takes over the total control of the operation for our internal organs. When the brain senses that body weight has been increased, it will give orders to the liver to produce glucose around 3am for storing future needed energy. If liver produces excessive glucose, then the brain gives order to the pancreas to produce insulin to balance the glucose level. However, for a diabetic patient, the function of the pancreas has been compromised; therefore, the glucose control mechanism will cause our bodies to have a higher fasting glucose value in the morning (FPG). Since diet and exercise cannot alter FPG directly during sleep, the secondary factor, weight, becomes the input factor and comes into play. Of course, this is my interpretation without a thorough understanding of biomedicine.

By the end of May 2017, I have figured out the inter-relationship between weight and fasting glucose, and therefore developed a predicted fasting glucose equation. In Figure 5-5, I have summarized my predicted FPG values under 6 different body weights in the morning based on the least square mean calculation of past 90-days data.

Figure 5-5: My predicted FPG values under 6 different body weights in the morning

The final comparison results of FPG and PPG of 3 meals between predicted and actual glucose during the period of 6/1/2016 through 7/30/2017 are displayed in Figure 5-6. It shows that the accuracy percentages of all glucose are extremely high and all of the correlation coefficients are quite high as well. For the FPG case, the accuracy is around 99% and the correlation is between 48% – 76% (for both daily glucose and 90-days moving average glucose cases). This Figure illustrates my newly developed fasting glucose prediction model, on 7/30/2017, is highly reliable as with my post-meal glucose prediction model developed two years ago, on 6/1/2015.

Figure 5-6: Accuracy and Correlation between predicted glucose and actual glucose for FPG and 3 PPGs

The next step is to expand my developed FPG prediction equation to be suitable for the general public, i.e. other diabetic patients. This part of research, using trial-and-error, took two months (June and July of 2017) to complete. Figure 5-7: Using my data of weight & FPG to generalize it into BMI & FPG for other patients, illustrates how I developed this predicted fasting glucose step by step, including considerations of its possible future extensions.

Figure 5-7: Relationship between FPG and BMI for a 26-months period with a higher FPG sub-period and a lower FPG sub-period

Based on my understanding of both biomedicine and mathematics, it is my educated guess that there is a “skewed S-shape” (or a better description of an “escalator” shape) existing for the ranges of both BMI<24.5 and BMI>27.5. Let me try to explain it in simple terms. If you just apply the straight line as the prediction tool, you will reach to a glucose value of below 60 mg/dL very quickly (insulin shock) when your BMI is dropping below 24. So, I bent the straight prediction line upward. Similarly, if your BMI is rising beyond 28, by applying this straight line as the prediction tool, you will reach to a glucose over 180 mg/dL very quickly. Unless your physical health has been mistreated for a long period of time, your glucose value should be below 200 mg/dL. That is why I also bent the straight prediction line downward. These two bending effects make the prediction equation curve’s shape looks like an escalator. As of now, I do not have sufficient data to support these two “extreme” scenarios. I also have no intention to push myself either into insulin shock or become a severe diabetic patient again. However, as a scientist, it is my responsibility to make the best educated hypothesis and then verify it later.

Since 3/18/2017, I have started to lose weight again and, by 6/1/2017, I finally reduced my weight to 171 lbs. once more. From 11/23/2016 through 6/1/2017, I have experienced 6 months of higher FPG. Therefore, I needed at least another 6 months of data, until 12/1/2017, to maintain my low-weight and low-FPG level. Once I reached that day, I will be able to perform another analysis which would cover the following 3 equal length sub-periods:
(1) 6/1/2016-11/30/2016 with low FPG
(2) 12/1/2016-5/31/2017 with high FPG
(3) 6/1/2017-11/30/2017 with low FPG

Through this experiment, I can then confidently verify my conclusion with highly reliable correlation between weight and fasting glucose. (Note: My latest A1C lab test on 8/18/17 is 6.4%, which is reduced by 5% compared to the previous reading of 6.7% on 6/1/2017. The decrease is caused by the lower FPG values during this period based on my weight reduction of ~10 lbs.)

As shown in Figure 5-7: Relationship between FPG and BMI for a 16-months period with a higher FPG sub-period and a lower FPG sub-period, I am now highly confident to offer this analysis model to other diabetic patients. I do understand that the human body is a high degree of nonlinear system and it is difficult to oversimplify it by any simple mathematical equation. For example, through part of my previous correlation study, I have also discovered the existence of the close relationship between sleep quality and FPG, but my collected data regarding this area is still insufficient. I will continue my data collection and analysis in this area in order to understand the nonlinearity and other disturbances of fasting glucose prediction.

Based on the collected data from 1/1/2014 through 7/30/2017 (~1,300 days), using BMI and FPG as x- and y-coordinates, a straight-line relationship existed between them, [(24.5, 100) and (27.5, 140)], which is the basic relationship between weight and FPG. It should be noted that a high correlation coefficient of 82.5% existed between the 30-days moving average weight and 90-days moving average FPG.

On top of these findings, I also calculated the following 2 probability ranges of data deviation due to other potential disturbance on the “secondary” factors on FPG:
– Actual FPG value falls into the range of +10 and -10 mg/dL: 51%
– Actual FPG value falls into the range of +20 and -20 mg/dL: 86%

This means that there are approximately 14% of actual FPG data are falling outside of this wider band due to other secondary factors.

These findings are illustrated in Figure 5-8: Sensitivity study results between weight and FPG for a period of 16 months (4/1/2016 – 7/30/2017).

Figure 5-8: Sensitivity study results between weight and FPG for a period of 16 months (4/1/2016 – 7/30/2017)

After consolidating all of my data together, including weight, fasting glucose, post-meal glucose, and daily average glucose, I chose a complete period from 1/1/2014 through 7/30/2017 to conduct the correlation analyses between weight and glucose, including both FPG and PPG. The results are shown in both Figure 5-9: A complete period (1/12014 – 7/30/2017) of 1,300 days’ data exhibit a clear and strong correlation between weight and FPG and Figure 5-10: A complete period (1/1/2014 – 7/30/2017) of 1,300 days’ data with Comparison study results between weight and 3 glucose values (FPG, PPG, Daily Glucose). The conclusion is that glucose is closely tied with weight, which has been known for quite a long time in the medical community. However, in my research and tool development, I have proved its existence via physics and engineering research approach and big data analytics. I have also discovered and utilized two very different, yet highly accurate, prediction models for FPG and PPG, respectively. Figure 5-11: Comparison between predicted FPG and actual FPG vs. weight, provides some insights of my FPG model.

Figure 5-9: A complete period (1/1/2014 – 7/30/2017) of 1,300 days’ data exhibit a clear and strong correlation between weight and FPG
Figure 5-10: A complete period (1/1/2014 – 7/30/2017) of 1,300 days’ data with Comparison study results between weight and 3 glucose values (FPG, PPG, Daily Glucose)
Figure 5-11: Comparison between predicted FPG and actual FPG vs. weight