3.1. Chemical Relationships between Acetone and Blood Glucose
In spite of the lack of reliable non-invasive techniques for diabetic blood glucose monitoring and lifestyle management decision support, there is evidence to suggest potential for non-invasively extracting and mapping acetone levels from human fluid to a blood glucose level, which may then be automatically and intelligently analyzed to support more appropriate lifestyle choices, such as regulating sugar intake. This section focuses on the chemical relationship between acetone, present in human breath, and blood glucose. The next section introduces intelligent decision support systems to show potential for customized algorithm development for diabetic management. Finally, this paper concludes with the proposal of a non-invasive device for extraction and classification of blood glucose levels based on acetone, as extracted from the human breath.
Ketosis is the process by which the body burns fat to make energy. During this process, Β-hydroxybutyrate (β-HB), commonly known as 3-hydroxybutyric acid, a ketone body, is created by the liver typically from oxidation of unsaturated fats, and is transferred to peripheral tissues for use as an energy source [
13]. A ketone body essentially comprises three molecules: acetoacetate, β-HB, and acetone. β-HB and acetoacetate transport energy from the liver to alternate tissues and acetone is produced by unconstrained decarboxylation of acetoacetate [
13]. Ordinarily, ketosis can demonstrate that lipid metabolism has been enacted and the pathway of lipid degradation is in place. Normal ketosis is prevalent in many circumstances, for example, during fasting, after delayed activity or after a high-fat eating routine [
13]. Pathological reasons for ketosis include organ failure, diabetes, youth hypoglycemia, corticosteroid or growth hormone deficiency, intoxication with liquor or salicylates, and a few intrinsic mistakes of metabolism [
13].
Diabetic ketoacidosis (DKA) is a dangerous internal condition that most commonly happens in those recently diagnosed with diabetes. This condition is a result of insufficient insulin production by the pancreas [
8]. This condition can also develop in type 2 diabetics who have poor diabetic control and management [
8]. Due to a lack of insulin in the bloodstream, a rise in blood sugar concentration occurs, which in turns prompts fat, sugar, acid unevenness, water, and protein build-up. Currently, a diagnostic test for DKA evaluates patient urine. However, the test is more sensitive to acetoacetic acid than acetone, rendering this test ineffective for the successful extraction of acetone for patients with DKA.
Acetone may be extracted from other bodily fluids such as in blood analysis. In a study of forty patients, researchers compared the measurements of ketones, specifically β-HB, in the blood with acetone concentrations extracted from human breath exhaled from the same human subjects [
13]. Results indicate a positive correlation: that acetone concentrations increase with an increase in β-HB in the blood [
13]. This indicates an association between levels of acetone extracted from human breath and β-HB, which is formed when blood glucose increases. In a separate study, researchers analyzed acetone concentration in the breath to determine whether there was an association with blood glucose concentration in type 1 diabetes during hypoglycemia [
14]. Researchers applied ion flow tube mass spectrometry with eight subjects with a mean age of 28 years and a mean HbA1c level of 8.8 (glycemic control). An insulin clamp technique, a controlled stepwise reduction in plasma glucose levels, was performed over the duration of a morning [
14]. Insulin clamps have been widely utilized as a metabolic tool for manipulating blood glucose levels [
14]. Sample bags were used to collect breath samples from each subject. Test results indicate a positive linear correlation between plasma glucose and acetone in the breath [
14].
Further research is required to determine the direct relationship between acetone and blood glucose changes during ketosis. The correlation between acetone and blood glucose raises several research questions, such as whether acetone can be adequately extracted and accurately mapped from the human breath to blood glucose levels in the presence of environmental variables, whether a sensor device can precisely support this extraction and quantify the result, and whether an associated software for decision support based on this measurement may be developed that intelligently adapts in real-time to both patient and environmental dynamic stimuli.
3.2. Intelligent Information Processing for Diabetic Decision Support
This section evaluates intelligent classifiers toward their integration into a dynamic system that can input the results of the blood glucose level extracted from the human breath in addition to other indicators from a patient profile, to provide decision support for diabetic lifestyle management.
Fuzzy logic relies on the principle of dynamic classification of data based on a membership function that changes its parameters according to incoming, changing data. The boundaries of classification of the data change; they are fuzzy and account for variation in data characteristics, in contrast to hard limits. Fuzzy logic uses linguistic terms embedded within membership functions for intelligent classification of dynamic data, mimicking the process of human cognitive processing. Fuzzy sets are analyzed and evaluated according to if–then rules of the fuzzy system. The development of a reliable and accurate fuzzy expert system depends on the experience of the expert [
15,
16]. Fuzzy logic controllers (FLCs) may prove useful for blood glucose level evaluation, given a blood glucose measurement, user input, activity selection, and other patient-specific parameters that change dynamically. Researchers have applied FLCs to classify and offer decision support for diabetic disease diagnosis [
17,
18,
19,
20], and for progression and management [
18]. The FLC typically comprises a fuzzifier, which converts numeric values into fuzzy sets in the process of fuzzification, an inference engine that performs logical controls in the FLC, and a rule base, which comprises the control rules and membership functions. The results of the inference engine are then defuzzified with conversion back to numeric values.
An FLC was developed to classify a patient as diabetic or non-diabetic based on blood glucose reading input, incorporating expert knowledge with parameters including diet, medicine, and exercise [
18]. This type of fuzzy expert system framework constructs an extensive knowledge base that uses defuzzification to convert fuzzy values into hard limit values. In [
18], rules were constructed to implement decisions on the amount of insulin dosage required and the number of times a person would need to self-inject; affiliated semantics and corresponding quantitative correction included “increase dosage” (5%) and “decrease dosage” (9%). Results of the classifier in [
18] indicate that the system effectively offered decision support to the diabetic user for self-regulation of blood glucose levels, toward the prevention of symptoms and complications.
An FLC is developed in [
19] for classification of a diabetic candidate into a category denoting the stage of diabetes. Symptoms of the disease are the input parameters and fuzzy logic is applied to analyze and classify the patient into one of the four categories: type 1 diabetes, type 2 diabetes, pre-diabetes, and gestational diabetes [
19]. Current patient data is entered, in addition to glucose level and historical patient data. Fuzzy logic is applied to classify this information using dynamic grading as implemented through customized membership functions [
19]. While thresholds defined within the membership function may change, FLC fails to adapt its design structure based on the incoming information [
19,
21]. The proposed system requires adaptation in patient lifestyle for efficiency and the ability to provide precise information to the patient in critical situations. Precision is key in this area, as it can solve a situation where a patient is experiencing low blood glucose in a matter of minutes by providing the patient the required or adequate amount of glucose in order to resume proper functionality of the body during hypoglycemia.
Diabetic Decision Support Systems (DSSs) have incorporated FLCs for the analysis of diabetic blood glucose reaction to insulin infusion through the use of Mamdani fuzzy controller models [
20], with the application of swarm optimization for fuzzy controller tuning. Healthy subjects were presented with minimum infused insulin and their reaction tested, against diabetics [
20]. Parameter uncertainties were modeled in the FLC by degree of insensitivity to multiple meal disturbances, level of accuracy, and superiority of robustness [
20]. The Mamdani fuzzy logic architecture contained two input variables and one output variable [
20]. The input linguistic variables were the error signal between the measured blood glucose level and the reference glucose level, and its rate of change [
20]. The output linguistic variable was the exogenous infusion rate of the insulin into the blood stream [
20].
Neural networks (NNs) are composed of nodal elements and have the ability to adapt to incoming information [
21]. Commonly, NNs are adjusted and trained to ensure that a specific input leads to a particular output [
21]. NNs have the capability to train and process complex functions in many applications including identification, pattern recognition, speech processing, classification, and control systems [
21]. NNs are implemented for diabetic diagnosis [
22], for analyzing blood glucose range as a primary indicator of the diabetic condition, in addition to diabetic management and associated insulin administration [
22]. NNs reveal their utility and efficiency in solving problems that require prediction, clinical diagnosis, pattern recognition, and/or image analysis.
A system is designed in [
22] to provide testing of diabetic abnormalities, where NNs are used in the early stages of diabetic diagnosis. Input parameters to the NN classifier include the number of times a patient has been pregnant, their plasma glucose concentration, blood pressure, body mass index, and insulin production level [
22]. Results of the research revealed the NNs capable of learning patterns corresponding to diabetic symptoms of an individual, achieving an accuracy of 98% [
22]. Prolonged, constant NN training was a noted limitation of the method, and researchers determined that, after a two-month period of no training, system accuracy dropped to 91% [
22].
A hybrid approach utilizing knowledge-based expert systems, adaptive control, simulation, optimization, and NNs is also implemented for decision support for insulin administration, where the expert system mimics soft knowledge expressed by medical expert opinions, intuition, laboratory results, and clinician’s observations [
22]. Parameters of interest include biological functions of blood glucose for the prediction of blood glucose response to insulin intake, coupled with decision-making and intelligent knowledge based systems combining mathematical and knowledge-based techniques [
22]. The study revealed NNs to be more suitable than knowledge-based approaches in diabetic diagnosis and management applications. The options for the selection of activity for insulin administration such as age, BMI, and blood glucose levels are inflexible and exclude other important factors such as activity intensity level, dietary habits, and predictive blood glucose control [
22], in addition to the inflexibility of applied uniform decision making. Experts offer non-uniform guidance and personal experience as well as patient lifestyle, which was excluded [
22]. Shortcomings were identified during clinical evaluation [
22] and NNs effectively addressed these deficits for more flexible lifestyle guidance with insulin administration. NNs seemingly provide a superior design structure over knowledge-based systems with the potential to enable variability of user input, patient profiling, and adaptive learning for incoming data and blood glucose-guided management for recommendations of intake and activity level based on intelligent information processing.
Decision tree (DT) approaches can be used to model a sequential decision problem under uncertainty. DTs graphically represent the decisions to be taken, events that could happen and the outcomes that may result. The structure of DTs incorporate nodes, branches, terminal values, strategy, payoff distribution, and rollback. A major objective of the system is to determine the best decision to take based on tree nodes that will constitute parametric values for consideration of a chosen path. It has previously been established that type 2 diabetes is age- and lifestyle-related, such that older persons (over the age of 40) and those that do not have healthier lifestyles are at greater risk, in comparison to younger, healthier counterparts. DTs have been utilized to identify individuals with impaired glucose metabolism or type 2 diabetes [
23]. DTs have been implemented due to their clear presentation of complex data that allows easy interpretation [
23].
In [
23], the objective was a diabetic diagnosis through automatic generation of decision trees on a cohort of 1737 individuals without previously known diabetes [
23]. Analysis included 1175 females and 562 males of the cross-sectional Metabolic Syndrome Berlin Potsdam study [
23]. This test consisted of test patients of ages greater than 18 years and all without known diabetes. Exclusion criteria for the used sample in this experiment were existing diseases such as liver disease, renal failure, and/or cancer. Tested DTs included well-established risk factors of diabetes as nodes, including age, blood pressure, and fasting glucose. Results revealed that age was the greatest discriminating factor for disease diagnosis with a threshold of 48.3 years. People younger than 48.3 years were further sub-arranged by a child node based on systolic blood pressure with a limit at 127 mmHg. It was determined that the proposed DT would avoid OGTTS in about 25% of individuals with a reasonable rate of false-negative results [
23]. The research determined that a sensitivity just over 30% was required to identify IFG, IGT, or T2DM [
23], which does not enable the identification and classification of pre-diabetic situations [
23].
Hybrid approaches utilize a combination of strategies including FLC, NNs, DTs, and/or rule-based expert systems in diabetic decision support. A hybrid system was developed that uses NN predictors with an FLC to regulate blood glucose in type 1 diabetics [
24]. The NN is added into this system to avoid errors of user input that include mealtime and size. The use of NNs in future blood glucose prediction assists with the proposed control technique to handle delays connected with insulin absorption and time [
24]. The FLC works alongside NNs and uses the predicted information concerning blood glucose to determine the required insulin dosage for the patient [
24]. A feed forward NN and a recurrent NN are implemented and assessed as nonlinear glucose prediction models [
24]. The RNN provides vastly improved expectation accuracy than the FFNN, particularly at longer prediction horizons [
24]. The RNN architecture consists of three layers, two hidden layers (20 neurons and 13 neurons) and a one-output layer of one neuron [
24]. The RNN is trained using the back-propagation algorithm [
24], which looks for the minimum of the error function in weight space using the method of gradient descent. The FFNN architecture comprises three layers. Neurons in distinctive layers are completely joined and information or processed data is fed forward [
24]. The activation function used in the two concealed layers is a sigmoid function.
The preparation and testing of the RNN and FNN prediction models are performed utilizing glucose measurements from nine subjects and patient data [
25]. Most RNNs and FFNNs are tested to advance the predicted glucose values [
25]. For RNNs, two hidden layers with 20 neurons and 13 neurons gave the best results [
24]. In addition, observations have shown that using a predictive control strategy with a neural system as an indicator and an FLC gives glucose peaks lower than that obtained by an FLC without predictor. The proposed system was best suited for daytime usage for better regulation of postprandial glucose concentration. The RNN gives lower postprandial (period after dinner or lunch) glucose peaks than the FFNN. The FLC without a predictor does not have the ability to stay within the extreme values/threshold and thus is unable to regulate the blood glucose level. The proposed system that uses NNs and FLC to deliver the required insulin dose works well for reducing the complications of hyperglycemia by predicting the patient’s future blood glucose levels and providing decision support for insulin dosage. Within the system, four features including age, BMI, blood glucose levels, and fasting blood were used in a hybrid approach with FLC and NNs as classifiers performing weight calculation, with embedded membership functions and network logic, for diabetic diagnosis [
24]. This system achieved a classification accuracy of 98.14% [
24]. The method [
24] may be adapted for varied classification weights, membership functions, and nodal network connections, for lifestyle management decision support including fluid recommended intake and suggested activity. Hybrid approaches that include NNs and FLCs offer greater system accuracy for diabetic management: NNs offer utility for adaptation and training of the DSS while the FLC implements the decision and classification logic [
25].