Corresponding Author: Gerald C. Hsu, eclaireMD Foundation, USA.
Diabetes is a metabolic disorder in which plasma glucose blood levels are abnormally high and the body cannot produce enough insulin or become insulin-resistant. Unfortunately, people with diabetes have a higher risk of getting cardiovascular disease and stroke. In order to fully understand a particular patient’s diabetes condition, the patients and their medical professionals need to have a broad sense of the disease, in-depth knowledge, and quick view of the patient’s metabolism status, i.e. overall health state. By having a simple indicator that dynamically demonstrates a general health status on a daily or momentary basis can benefit them.
The basic requirement for patients with type 2 diabetes (T2D) to control their disease is to know what their glucose values are. However, glucose testing is invasive, inconvenient, and costly. Most T2D patients are not performing the measurement on a regular basis for these reasons; therefore, it is important to find an alternative way to achieve a quick, easy, painless, low-cost, effective and accurate testing method available for these patients.
The author was diagnosed with T2D since 1995 and, by 2010, his health was in a “near-collapse” state. He suffered from five episodes of cardiac issues, high possibility of requiring kidney dialysis in the future, bladder problems, foot ulcer, and other medical complications. As a result, he decided to launch his research on diabetes in order to save his own life. He spent nearly 8-years (~20,000 hours) researching an effective way to help himself to control his diabetes conditions via a scientific, yet simplified, and effective lifestyle management method. His health condition comparison between 2010 and 2017 can be seen in (Figure 1).
Figure 1: Health Data Comparison Between 2010 and 2017
Since he is a mathematician, engineer, and an industrialist, he launched his efforts by performing the following standard steps:
- phenomena observation
- data collection
- equation development
- statistical analysis
- practical answers for problem-solving
- user-friendly tool to use
This paper describes how he achieved his health goals by monitoring his daily metabolism status and also predicting his daily glucose values automatically via his invented “math-physical medicine” approach and AI tool.
His glucose values include FPG (fasting plasma glucose value measured in the morning before breakfast) and PPG (postprandial plasma glucose values measured after first bite of food 2 hours later and 3 times per day).
All data was collected in its entirety from one patient only, himself, via a customized software over 7.5 years since 2012. His long and consistent education and work has provided practical experience on how important it is to collect and categorize “clean data” from the beginning. Otherwise, for many data analysis projects, research scientists spend 70% to 80% of their time and resources to clean up “dirty or contaminated data” before launching their real research work, which includes data process, analysis, and interpretation. As a result, he started his project by developing a software program since 2010, and by using the program the author was able to collect and process more than 95% of his data as “clean data” and needed very little data cleaning and organizing later on. This project does not need to be concerned with “data interference” and “data contamination” problems due to different sources of genetic conditions, various lifestyles, and contradicting data source interpretations. These data come from a consistent sample source, making it much easier for the author to dive into one variable and extract the buried information.
The author learned an important work ethic from Professor Norman Jones of MIT in the early 1970s about data integrity. In this study, he used his measured data as the base for future data comparison and research. He has safeguarded the integrity of his collected data and has never altered its original content or influenced its integrity. In this way, all results from using his developed prediction tools are compared against the measured glucose and A1C values.
2. Metabolism Monitoring
In the beginning, the author used advanced mathematics, nonlinear engineering modeling, big data analytics to develop a set of mathematical equations to describe the patient’s metabolism. He defined “metabolism” as a form of “energy body needs” which is an organic process containing easily measured metabolic conditions (weight, glucose, lipids, blood pressure) and daily lifestyle data (food, water, exercise, stress, sleep, healthy daily routine). This metabolism equation contained a total of 10 categories, 6 inputs and 4 outputs. In addition to the major categories, such as food and exercise, he also investigated the impact on his metabolism by traveling, water intake, bowel movement, urination, stress / tension / anxiety, disturbance on daily life routine pattern and healthy habits, psychological effect on the physiology of his health, etc. Overall, these 10 categories comprised of ~500 elements and ~1.5 million data collected and stored over 7.5 years. With such a big volume of data, a customized computer software program was developed for handling the data collection, processing, and analysis.
He defined two new terms known as the Metabolism Index (MI) and General Health Status Unit (GHSU). MI is the total score reflecting the body’s health condition (i.e. state of metabolism), which combined all of the 10 categories. GHSU is a moving average value of the past 90-days daily MI scores. The graph of this data (Figure 2) shows a person’s “health state”. The “break-even” line between a “healthy state” and an “unhealthy state” is 73.5%. This break-even percentage is calculated by the mathematical metabolism equation using standard healthy-state data, such as BMI 25, glucose 120 mg/dL, SBP/DBP 120mmHG/80mmHG, etc. A value below this percentage is regarded as “healthy” and a value above the line is “unhealthy”.
Figure 2: Metabolism Index (MI) and General Health Status Unit (GHSU) from 2012 to 2018.
3. Glucose Predictions
The author started with a simple task of predicting tomorrow’s weight output from the previous 3-day input of weight, food quantity, and bowel movement. The weight prediction is the pre-processor for predicting FPG in the morning which constitutes a minor part (about 20% – 25%) of A1C formation. Although there are five influential factors for FPG creation, he discovered and proved that weight is the predominant one.
The prediction of PPG, however, is a more complicated task since it involves about 15 influential factors that produce the PPG value.
He applied signal processing technology from geophysics and electronic engineering to decompose the human body’s highly nonlinear biomedical signal curves, such as the glucose wave, into multiple and different sub-waves created by each influential factor. He carefully checked each sub-signal waveform for its completeness, accuracy, and correlation with other curves, using time-series analysis, spatial analysis, and frequency-domain analysis, etc. Finally, he reintegrated these multiple sub-waveforms back to a predicted glucose curve to simulate the actual measured one. By developing many mathematical equations and analyzing their data using various statistical models, he was able to identify primary, secondary, and tertiary factors according to their respective contribution margins and importance levels on glucose creation. Those factors for FPG and PPG will be discussed in the “Results” section.
Over the past three years, he continuously explored and added some missing influential factors into the formation of the PPG signal. His purpose was to improve the predicted PPG waveform’s contents and accuracy while maintaining a high correlation with the measured PPG waveform.
He further improved his model via a “curve-fitting” trial and error engineering method which he learned from his defense work experience. He has continuously compared these two sets of data and improved the accuracy until it reached a high linear accuracy, while still maintaining a high correlation. High correlation means the trend of the predicted curve moving along with the measured curve like its “twin”. This predicted PPG also serves as the major part (about 75% – 80%) of his estimated daily A1C.
He also developed a machine-learning statistical algorithm to automatically adjust the conversion value of daily averaged glucose value from FPG and PPG to a combined daily A1C value. Finally, he specifically added in a “safety margin” which he learned from his nuclear power work experience. The reason for having the safety margin on top of the estimated A1C is to cover the possible variance generated by different chemical process and various environmental factors associated with the A1C testing done in the laboratory. This extra caution can provide a numerical safety buffer to avoid creating unnecessary panic on T2D patients while serving as a daily “early warning” to them before they have a chance to get their A1C tested.
The author spent 10 months during 2015-2016 investigating FPG. Initially, he exhausted all avenues to find possible connecting factors, despite finding a low correlation of ~9% between FPG and PPG. On 3/17/2016, he discovered the high correlation of 84% between FPG and Weight. He used 26,000 FPG-related data based on 1,505 days (1/1/2014 – 2/14/2018), to conduct statistical analyses. In the time-series diagram, there are three high periods and three low periods of Weight, along with the FPG curve following the Weight curve like its “twin”. In spatial analysis diagram of BMI vs. FPG (without time factor), there is a “quasi-linear” equation existing between two coordinates of BMI and FPG from point A (24.5, 102) to point B (27.0,142). The stochastic (random) distribution of data has two clear “concentration bands” stretched from lower left corner toward upper right corner. The +/- 10% band covers 67% of the total data and the +/- 20% band covers 94% of the total data. Only the remaining 6% of the total data is influenced by other secondary factors. The following graph (Figure 3) shows the close relationship between weight and FPG.
Figure 3: FPG and Weight Relationship (spatial analysis and time-series analysis)
During the 989 days (6/1/2015 – 2/14/2018), he had 2,967 meals and collected about 60,000 PPG-related data.
The findings of PPG and its corresponding influential factors are as follows:
Primary factor No.1, Carbs/sugar intake:
- average 14.7 grams per meal
- +60% correlation
- 38% contribution rate
Primary factor No.2, Walking exercise:
- average post-meal 4,300 steps
- -64% correlation
- 43% contribution rate
- Secondary factor, Temperature: 10%
- Combined 12 tertiary factors: 9%
The following chart (Figure 4) shows the decomposition of PPG waveform into 4 major Sub-Waveforms; whereas, the other graph (Figure 5) displays the positive impact of carbs/sugar intake on PPG and negative impact of walking exercises on PPG as well as a what-if analysis.
Figure 4: Decomposition of 4 Sub-Waveforms of PPG
Figure 5: Impact of Carbs/Sugar Intake and Exercise on PPG
and What-If Analysis
The ~3,000 meal photos were analyzed against six million food data collected from the United States Department of Agriculture (USDA) and stored in a cloud server. All food data were sorted based on country, franchise restaurants, individual cafes, home-cooked meals, and airline food.
Here are some comparison results:
- Airline food (both airport and inflight) – 136 mg/dL
- Restaurant food (both franchise and individual) – 127 mg/dL
- Home Cooking (includes all nations) – 111 mg/dL
The concluding results of high accuracy and correlations for two glucose predictions are as follows:
- The predicted FPG vs. measured FPG achieved a linear accuracy 99.8% (118.42 mg/dL vs. 118.62 mg/dL) and 98.6% correlation.
- The predicted PPG vs. measured PPG achieved a linear accuracy 99.3% (119.37 mg/dL vs. 120.16 mg/dL) and 71.3% correlation.
Please see the diagram (Figure 6) of the comparison of his estimated Daily A1C vs. lab-tested A1C values.
Figure 6: Estimated Daily A1C Curve (with 7% to 15% safety margin) and Lab-Tested A1C Data Since 2010
All of these concluding results have been programmed into a practical and AI software tool by the author. The requirements for the program were to measure his morning body weight and to take his three meal photos before starting his meal to store into the software. This AI tool can then immediately display his predicted fasting glucose as well as his 3 post-meal glucose values on the iPhone screen. If one particular predicted PPG value is too high, he can then decrease his carbs/sugar intake amount or rearrange his meal’s portion to get a new meal photo which reduces to a lower predicted PPG reading. The AI glucometer tool was developed entirely based on his research of glucose prediction technology, using optical physics, signal processing, and all of the mathematics and engineering techniques mentioned in this article.
The highly accurate predicted glucose values, both FPG and PPG, can provide an effective tool for T2D patients to control their diabetes condition. In addition, the sophisticated and dynamic metabolism index (MI) prediction can provide a snapshot of their daily overall health state.
Other than this full-length article, an abstract is also available for an overview. In addition, the author also displayed a table of summary of different points between existing knowledge domain and math-physical medicine, see (Table 1).
Table 1: Different Points Between Existing Knowledge Domain and Math-Physical Medicine Approach
Limitation of Research
This article is based on data collected from one T2D patient’s 8-years metabolic conditions and lifestyle details (i.e. his own data). It does not cover genetic conditions and life style details of other diabetes patients. His BMI was >31 (obese) in 2000. However, during the period of 2010 – 2018, he has reduced his BMI from 29.5 to 24.5 (overweight). Therefore, his conclusions and findings should be reverified for patients who are either underweight or obese. The author believes in his own work’s results, findings, and conclusions, which are based on a solid academic background and a careful and thorough process of identifying the system’s basic characters, developing various mathematical equations and statistical models, using modern computer science tools and sophisticated AI techniques. However, other T2D patients need to be cautious about applying his finding, results, and conclusions under different metabolic conditions.
The author has never hired any research assistant or associate to help with his work except for a part-time computer programmer (~3 working hours per day). He applied his own invention of a “Software Robot” created during 2001-2009 and AI knowledge he learned to produce his customized computer software for this research project and diabetes control.
This project was self-funded by using his own money that was earned from a successful high-tech venture in Silicon Valley. He did not receive any financial assistance or grants from any institution or organization.
He created this math-physical medicine approach by himself in order to save his own life. Although he has read many medical books, journals, articles, and papers, he did not specifically utilize any data or methodology from other medical work. All of his research is his original work based on data he collected from his own body and using computer software he developed during the past 8-years. Therefore, no major problems were associated with data interference or data contamination. In addition, his knowledge, information, technique, and methodology of mathematics, physics, engineering, and computer science came from his lifelong learning from schools or industries and should not be listed as medical references. This is the reason he does not have a reference section in his research article.
First and foremost, the author wishes to express his sincere appreciation to a very important person in his life, Professor Norman Jones at MIT. Not only did he give him the opportunity to study at MIT, but he also trained him extensively on how to solve problems and conduct scientific research with a big vision, integrity, and honesty.
The author would also like to thank Professor James Andrews at the University of Iowa. He helped and supported him tremendously when he first came to the United States. He believed in him and prepared him to build his engineering foundation during his undergraduate and master’s degree work.