Artificial intelligence biosensors for continuous glucose monitoring

Artificial intelligence (AI) algorithms in combination with continuous monitoring technologies have the potential to revolutionize chronic disease management. The recent innovations in both continuous glucose monitoring (CGM) and the closed‐loop highlight the far‐reaching potential of AI biosensors for individual healthcare. This review summarizes some of the most advanced progress made in CGM biosensing. We will focus on three main applications of AI algorithms in diabetes management: closed‐loop control algorithms, glucose predictions, and calibrations. The challenges and opportunities of AI technologies for CGM in individualized and proactive medicine will also be discussed.


| INTRODUCTION
Diabetes mellitus is a worldwide chronic disease caused by a disorder of glucose metabolism. [1,2] The global diabetic population is projected to increase to about 439 million adults, with nearly 490 billion dollars in financial burden by 2030. [3] Glucose metabolism abnormality is the main cause of diabetes and downstream complications. Therefore, it is important to control glucose levels in diabetic patients. [4] Traditional blood glucose monitoring is based on standard capillary blood glucose testing through a finger prick. However, such monitoring is unable to identify the roles of glycemic in a timely manner. [5] Expansion to continuous glucose monitoring (CGM) is an area of active interest and will, to a great extent, depend on technical innovation, which took many decades to optimize. Recently, the popularity of wearable CGM biosensors has risen enormously, exceeding nearly $1 billion in the market. [6] Compared with fingerstick blood glucose monitoring, CGM has many advantages in diabetes management. First, CGM avoids many inconveniences such as physical and psychological pain that are normally associated with the standard method of capillary blood glucose testing. [7] Furthermore, CGM can be used to capture a holistic model of glycemic control to obtain "time in range" targets, which is recommended by multiple international consensus reports. [8] Additionally, the CGM biosensor is vital to enable a closed-loop that relies on realtime glucose data. [9] Ultimately, CGM is an important tool for obtaining patient-specific big data sets to help better lifestyle behavior adherence and personalized therapy. [10] The use of CGM has been demonstrated to have the potential in improving HbA1c levels as well as reducing the time spent in hypoglycemic events. [5] In the future, with increased clinical experience and ongoing research on biosensors, CGM will be used in a variety of circumstances.
Despite the rapid growth in CGM technologies, there are still several challenges to the goal of widespread acceptance, such as cost, time lag, need for calibration, and so on. [11] The topics of CGM have been covered in excellent reviews. [5][6][7][12][13][14][15][16][17][18][19][20][21][22][23][24][25] In this paper, we review the evolution of CGM, as well as wearable CGM biosensors. Our emphasis is placed on three main artificial intelligence (AI) applications for CGM biosensors: closed-loop control algorithms, glucose predictions, and calibrations ( Figure 1). The challenges and limitations of the next technological wave in glucose sensing and feedback loop will also be discussed.

| EVOLUTION OF CGM
A typical CGM system usually consists of a glucose recognition element, a physical or chemical transducer element, a wireless transmitter element, and a receiver. [17,24] The glucose recognition element provides high selectivity towards glucose and plays a key role in the success of CGM. The function of the transducer element is to convert the glucose concentration to a measurable analytical signal. Ideally, the signal would be transmitted wirelessly to a receiver, which may be a mobile app or a dedicated handheld device that integrates specific algorithms to follow glucose measurements or an insulin pump. On the basis of the type of transducer element, CGM can be classified into the category of electrochemical sensors and optical sensors.
Over the past few decades, there have been many efforts to explore new technologies to improve the performance of CGM ( Figure 2). Beginning in the 1960s, Weller et al. [26] first reported continuous monitoring of blood glucose concentration in human subjects. Inspired by the quest for an artificial pancreas, numerous efforts have been devoted to the development of CGM. For example, in 1963, Kadish used continuous real-time glucose monitoring as the initial approach to test out the closed-loop glucose control. [27,28] Moreover, in 1974, Albisser et al. [29] proposed the concept of an artificial pancreas, and the combination of CGM with algorithms to automate insulin delivery has shown its promise in subjects with Type 1 diabetes (T1D). In 1977, Miles Laboratory launched a commercial CGM product called the Biostator. [28] However, the CGM systems that were mentioned in the latter are considered bulky devices by today's standards. Owing to the development of enzymatic electrochemical glucose sensing, the miniaturization of CGM biosensors has been gradually realized. For instance, in 1982, Shichiri et al. first proposed a needle-type glucose sensor (2 cm long and 0.4-1 mm in diameter), which continuously monitors blood glucose in dogs for up to 7 days. [30] In 1999, the first professional CGM system was approved by the US Food and Drug Administration (FDA), which ushered a new era in diabetes management. [18] Since then, upcoming developments in CGM technologies are moving towards real-time monitoring, miniaturization of the devices, and sustainability for the long term. In 2005, real-time CGM systems began to be applied in clinics. In 2016, the first commercial hybrid closed-loop system (MiniMed 670G) was approved by the FDA. [9,31] Evidently, the current market for CGM is everexpanding. The major players in the market are Dexcom, Abbott, Medtronic, and Senseonics. For example, Dexcom G6 CGM is a fully integrated system made available in 2018. The wearable G6 sensor can be used for up to 10 days without calibration. Due to customizable alerts, the G6 System also aids in alarming blood glucose levels. [32] Moreover, Abbott's FreeStyle Libre CGM sensor is used by nearly 800,000 people in over 43 countries worldwide. [16] The sensor can be worn for up to 14 days without fingersticks. Furthermore, FreeStyle Libre 2 added a smart alarm function in 2018. [33] Medtronic Guardian sensor 3 can be worn for up to 7 days and alert hypoglycemic and hyperglycemic events before 60 min. [34] In contrast, the Eversense is a fluorescence-based sensor that can be used for up to 180 days and is composed of an implantable sensor, a smart transmitter, and a mobile application. [35] 3 | GLUCOSE SENSING TECHNOLOGY FOR CONTINUOUS MONITORING

| Electrochemical glucose sensing technology
Most commercialized blood glucose assays are based on electrochemical enzymatic reactions of glucose with glucose oxidase (GOx) or glucose dehydrogenase. [36] The first generation of electrochemical glucose biosensors is fabricated by the enzyme-catalyzed O 2 -oxidation of glucose. However, there are two major obstacles to the first-generation glucose biosensor. One of the main obstacles in the first-generation biosensor is its high operating potential that is required for measuring hydrogen peroxide which also oxidizes biological interference and creates high current densities. The other obstacle is that without a redox mediator, the direct electron transfer from the FADH 2 center of GOx to the working electrode surface is too slow for it to take effect in time. To overcome the obstacles mentioned above, the second-generation glucose biosensor introduced the redox mediator which can decrease the dependence on ambient oxygen ( Figure 3). Therefore, the measurement can be carried out at a lower applied potential to avoid the effects of biological interferences. [37][38][39][40] There have been numerous publications describing the use of electrochemical sensing technology for CGM applications. For example, microneedle-based CGM biosensors are gaining attention because of their lower pain and tissue damage. [41][42][43] Electrochemical microneedle-based devices can realize simultaneous glucose monitoring and therapy. [44] The nonenzymatic glucose sensors have several advantages, such as stability of environmental variation, simplicity, and reproducibility of fabrication, low cost, and free from oxygen limitation. [39,45] For example, Yoon et al. reported a nonenzymatic glucose sensor-based nanoporous Pt as the catalytic material and a wireless module for CGM. [46] Compared with optical transducers, electrochemical biosensors have many advantages in CGM. Since there are no optical components, electrochemical sensors can be easily miniaturized for a wearable continuous monitoring platform. Moreover, because most electrochemical biosensors are label-free, the reaction time is very short, thus allowing electrochemical biosensors to easily achieve continuous monitoring. [47] Although commercialized electrochemical CGM sensors have reached 2-week lifespans, some key challenges remain before realizing long-term application, such as the degradation of enzymes, biocompatibility, and biofouling. For nonenzymatic electrochemical biosensors, poor selectivity is the main limitation for their application in CGM.

| Optical glucose sensing technology
Optical sensing technology is based on detecting changes in photons in the properties of light rather than of electrons. [23] It is a promising candidate method for CGM biosensors with advantages in both sensitivity and versatility, which enables faster and continuous monitoring. [48] On the basis of transducer systems, optical biosensors can be classified into different categories: fluorescence, surface plasmon resonance (SPR), Raman, near-infrared, Fourier transforms near-infrared, optical coherence tomography, and surface-enhanced Raman scattering biosensors. [49] Fluorescence and SPR have played an important role in optical glucose sensing technology for continuous monitoring because the principle of fluorescence sensing technology is based on glucose-induced changes in the intrinsic fluorescence of enzymes and the fluorescence resonance energy transfer or quenching between fluorescent donor and acceptor. [50] ( Figure 4A) For example, Schultz et al. presented a fluorescence affinity hollow fiber sensor that consisted of dyed beads and Concanavalin A for transdermal fluorescence-based glucose monitoring. The fabricated sensor can measure glucose in the physiological range with response times between 5 and 7 min. [51] For commercial application, the Senseonics CGM sensor is a fluorescence-based glucose sensing system consisting of a fully implantable glucose sensor and a wearable transmitter to measure glucose in the interstitial fluid (ISF). [52] Recently, Sawayama et al. developed a fluorescence-based implantable CGM with a glucoseresponsive fluorescence dye that continuously monitors blood glucose concentrations for 45 days. [53] SPR has attracted extensive attention in the field of biosensors due to its simplicity and low cost. The principle of the SPR sensing technology is based on the changes in the refractive index caused by glucose molecules interactions via a surface plasmon wave. [54][55][56] (Figure 4B) In 2020, Yang et al. reported an optical tapered fiber glucose sensor based on the localized surface plasmon resonance technique. The sensor displayed a linear measurement range for glucose from 0 to 10 mM, with a detection limit of 322 μM. [57] Despite the advancements in optical sensing technology for CGM, several key challenges still remain before reaching commercial maturity. Currently, most optical F I G U R E 4 A general scheme of optical glucose sensing technology (A) fluorescence and (B) surface plasmon resonance devices are costly and difficult to miniaturize. Moreover, because of the complexity of body fluids, optical glucose sensing technology has poor signal-to-noise and usually needs invasive calibration of the measurement. [24] Lastly, due to individual differences such as skin thickness, body fat, and blood volume, noninvasive optical glucose sensing technology lacks a universal algorithm model that meets the clinical requirements.

| Wearable technology in CGM
With wearable technologies swarming into the field of glucose monitoring, efforts have been devoted to the development of CGM biosensors for minimally/noninvasive detection of biomarkers in accessible biofluids, such as ISF, tear, sweat, and saliva ( Figure 5 and Table 1).

| Wearable ISF CGM
ISF is a kind of extracellular fluid, with glucose coming from the blood through continuous capillaries. The glucose concentration ranges of blood in people afflicted without and with diabetes are usually 2-30 and 4-8 mM, respectively. [89] The glucose concentration of ISF is nearly identical to that of the blood. [6] The main interferences for ISF glucose sensors include endogenous molecules, like, ascorbic acid (AA), dopamine (DA), and uric acid (UA). [90] The ISF has two main advantages for CGM. The main advantage it has in comparison to other noninvasive biofluids (saliva and tear) is that the glucose concentration in ISF is relatively high and has better detection accuracy and reliability. Furthermore, compared with blood, it is minimally invasive to realize CGM, thus CGM technology in ISF is now a mature field. [24] It is estimated that more than 44% of people with T1D have used an ISF glucose  [58] Copyright 2020, Elsevier B.V. (E) A wearable mouthguard CGM biosensor. Reproduced from Arakawa et al. [59] Copyright 2022, American Chemical Society. (F) A smart wristband for monitoring glucose in sweat. Reproduced from Gao et al. [60] Copyright 2022, Springer Nature Limited. CGM, continuous glucose monitoring; ISF, interstitial fluid; LED, light-emitting diode; PEG, polyethylene glycol.
sensor. [6] The use of microneedles for continuous sampling of the transdermal ISF provides an intermediate promise between implantable and noninvasive methods. [91,92] Pu et al. [68] presented a three-electrode electrochemical sensor integrated into a microfluidic chip for CGM. The authors used micromolding techniques to fabricate layers of a microfluidic chip Including a vacuum generator, valve, microchannel, electrode, and bottom five polydimethylsiloxane layers for transdermally extracting and collecting subcutaneous ISF. [93] Recently, Baghelani et al. reported microwave resonator-based sensors for continuous monitoring of glucose from ISF. The sensor consists of a substrateless split ring resonator tag which is electromagnetically coupled with a reader. [94] Due to the lack of power consumption of the sensing element, microwave resonatorbased implantable glucose biosensors have the ultimate potential to be coupled in the artificial pancreas. [95] Although CGM in ISF is widely accepted and rapidly growing, there are also several key challenges involving the closed-loop. First, there is an approximately 5-10 min lag time between the blood and ISF glucose due to the time it takes for glucose from plasma to ISF. [6] Second, as a subcutaneous glucose sensor, the loss of electrode materials and biocompatibility are essential issues in the development phase of ISF biosensors. [96] 3.3.2 | Wearable tear CGM Tear, also known as ocular fluid, is an aqueous humor that surrounds the eye. [97,98] The glucose concentration of tears for a healthy person is between 0.1 and 0.6 mM, while for a diabetic person, the glucose concentrations are between 0.5 and 5 mM. The main interferences for tear glucose sensors are lactate and pH fluctuation. [99,100] Due to the constant contact between the lens and ocular fluid, a contact lens could serve as an alternative sensing platform for CGM. [97] For example, Keum et al. [82] described a smart contact lens in a manner that allows both CGM and diabetic therapy. The flexible electrical interfaces formed in this manner are eventually integrated into the microcontroller chip with a power management unit, an electrochemical glucose biosensor, an on-demand controlled drug delivery, a resonant inductive wireless energy transfer system, and a remote radio frequency communication system located on the biocompatible polymer contact lens platform. [82] Kim et al. developed a wearable multifunctional contact lens sensor to monitor glucose in tears and intraocular pressure, respectively. In their study, field-effect transistors (FET) were fabricated using graphene-AgNW hybrid as S/D electrodes for real-time glucose sensing through the use of the resistance and capacitance of the electronic device for intraocular pressure. The demonstration of this smart contact lens has made continuous and wireless monitoring of ocular and overall health conditions possible. [83] Several other groups are pursuing similar approaches. [84,85,[101][102][103] While contact-lens biosensors have attracted great commercial attention for CGM among various wearable devices (such as Google lens), some key challenges such as security and privacy, clinically compatible testing, and so on still remain before the commercial maturity of tear CGM biosensors. [82] 3.3.3 | Wearable sweat CGM Sweat, a colorless biofluid, is secreted from the eccrine and apocrine sweat glands. [6,83,104] The concentration of glucose in sweat is in the range of 10-200 μM, which is measured to be about 1%-2% of that value in blood. Sweat glucose and blood glucose are highly correlated in subjects with diabetes. [105] Electrochemical sweat sensors are susceptible to interference from other electrochemically active compounds in sweat, such as AA (10-50 μM) and UA (2-10 mM). In addition, the sampling methods of the sweat and contamination from the skin may corrupt the concentration of glucose in sweat. [106] In recent years, sweat biosensing has received considerable attention for noninvasively CGM, since sweat can be conveniently collected and continuously monitored. [107][108][109] For example, Lee et al. developed a closed-loop solution for noninvasive sweat glucose monitoring and microneedlebased point-of-care therapy, as well as pH, temperature, and humidity measurements. The simultaneous use of a series of integrated sensors can enhance the accuracy of glucose sensing by correcting pH-dependent deviation of the enzyme-based glucose sensor. Moreover, the thermal actuation of hyaluronic acid hydrogel microneedles can thus lead to multistep controlled drug delivery. [70] Gao et al. presented a fully integrated wearable flexible sensing array for real-time simultaneous measurement of glucose, lactate, and electrolytes (sodium and potassium ions) in human sweat while skin temperature was also measured to calibrate the response of the glucose sensor. [60] Despite the recent advancements in sweat biosensing for CGM, there are still several major obstacles such as the dilution of glucose in sweat continued secretion, [106,110] as well as the varying rates of sweat production from different individuals. [19]

| Wearable saliva CGM
Human saliva is a serous or mucous exocrine secretion from both the salivary glands and gingival crevicular fluid. [111,112] The main constituents of saliva include water (98%) and other compounds, depending on the type of gland, time of day, age, and gender. [6,113] Compared with other noninvasive biofluids (ISF, tear, and sweat), saliva has the advantage of easy sampling. [114] Some studies have revealed that the concentrations of glucose in saliva are positively correlated to the levels in blood plasma. [6,79] A comparison study of salivary and blood glucose levels revealed a significant correlation between salivary and blood glucose levels in both diabetic and nondiabetic subjects. The salivary glucose levels in the subjects with and without diabetes are in ranges of 10-32 and 4.3-12.9 mg dl −1 , respectively. [115] Considering that saliva glucose is usually based on the format of the mouthguard, the main interference substrates are the cross-contamination of food and drink. For example, de Castro et al. reported a microfluidic paper-based device through craft cutter printing as a wearable sensor for saliva glucose monitoring. Such devices expanded new possibilities of applications of paper-based platforms as low-cost and biodegradable wearable sensors. [77] García-Carmona et al. described a pacifier-based wearable biosensor that is based on electrochemical principles, as a wireless device for noninvasive glucose monitoring in the infant's saliva. For this, the electrodes were fabricated using screenprinting technology on a PET layer, which could be easily replaced. Although this design had achieved great advances in saliva collection, the final result in efficient saliva pumping and continuous monitoring was still limited by the stability of the chitosan layer that provides enzyme modification into the electrode. [78] Although saliva is an alternative source of biofluid for CGM, there are still several limitations for commercialized applications. For example, considering miniaturization, saliva biosensors have a higher risk of being swallowed by the user. [6] With a flood of AI technologies swarming into healthcare, we are facing the age of AI biosensors for CGM. AI algorithms can extract relevant information from the rich data sets recorded by CGM biosensors and generate valuable insights that could be applied to personalized diagnosis and patient adherence therapy ( Figure 6).

| AI TECHNOLOGIES FOR CGM BIOSENSORS
Over the past decade, AI has been emerging in an effort to improve the performance of CGM biosensors. The American Diabetes Association recommends the use of AI not only as an alternative to traditional screening approaches but also as the detection of mild diabetic retinopathy and diabetic macular edema. [116] The most popular AI algorithms applied in diabetes care are the support vector machine, artificial neural networks (ANNs), supervised machine learning (SML), and principal component analysis algorithms. The combination of machine learning in CGM has shown its promise in different application scenarios for diabetes management such as calibration, decision support systems, closed-loop control, patient self-management tools, and automated retinal screening (Figure 7 and Table 2). [117,118] However, our main attention is still focused on closed-loop control, decision support system, and calibration exploiting AI techniques.
F I G U R E 6 Illustration of artificial intelligence biosensors for continuous glucose monitoring. Patients attach a CGM sensor to their skin which continuous monitoring glucose concentration and transmits wirelessly to a smartphone. AI-data processing can be grouped into interface, data classification storage, data model, and analysis. AI, artificial intelligence; CGM, continuous glucose monitoring; ISF, interstitial fluid.

| Closed-loop control algorithms
A person with T1D is dependent on insulin therapy, and needing to make frequent dosing decisions to achieve glycemic targets. [9] Thus, closed-loop control systems have the potential to automate insulin therapy for diabetes management. A closed-loop control system, also known as an artificial pancreas, is a network of drug delivery feedback loops, the idea of which glucose control was first proposed in the 1960s. [131] The first fully integrated commercial artificial pancreas system was the MiniMed 670G, which was approved for T1D management in 2016. [31] The basic architecture of an insulin therapy closed-loop control system is composed of three main elements: a CGM biosensor to make glucose measurements, a controller based on an algorithm that directs the insulin pump's delivery, and an insulin infusion pump to continuously deliver insulin. [5,7,9] The control algorithm plays an important role in closed-loop systems. It can be located in a smartphone as well as incorporated into the insulin pump. [131] Three main traditional types of control algorithms are proportional-integral-derivative (PID), fuzzy logic, and model predictive control (MPC). The PID controller is an algorithm that calculates the difference between the target and the measured point based on the current point of time, past difference values, and the area under the curve to direct insulin doses. The fuzzy logic controller is a powerful algorithm that can synthesize expert knowledge by imitating the logic of diabetes practitioners or experts in adjusting insulin delivery. The MPC algorithm uses a dynamic model based on hypothetical output values to predict glucose levels and adjust insulin delivery. [7,9,132] Due to recent advances in CGM biosensors and AI algorithms, the application of closed-loop technologies for diabetes management and precision medicine is ever-expanding. For example, Bahremand et al. developed ANNs-based MPC for blood glucose level prediction for T1DM rats. Blood glucose data obtained from the CGM biosensor was used for training the ANN prediction model which maintained the blood glucose level within the normal range for 90% of the time, suggesting that it can be used in conjunction with a closed-loop artificial pancreas system. [133] A fuzzy logic controller was designed to regulate the blood glucose based on serious disturbances (mixed meal absorption, exercise, delay, and noise in the glucose sensor). The results revealed that the fuzzy logic controller was more reliable and safer than the PID controller. [134] Clarke et al. used a personalized MPC to adjust insulin doses for recognizing blood glucose levels between 70 and 140 mg dl −1 overnight and hypoglycemic events reduction. MPC algorithms eliminate the inherent time delays between CGM and insulin infusion. [135] Despite the growing progress in the field of closedloop control algorithms, there are still several key challenges such as dosing accuracy, patient intervention, unexpected glucose disturbances from meals and exercise, patient-specific prediction, and so forth remaining before improving the performance of the closed-loop artificial pancreas system. Although the closed-loop system automatically controls blood glucose has been commercially available, the use of an artificial pancreas could still be unreliable and potentially burdensome for some patients. The majority of individuals with T1D still rely on multiple daily injection therapy. [136] F I G U R E 7 Schematic representation of artificial intelligence in diabetes management. CGM, continuous glucose monitoring.

| Glucose prediction based on CGM biosensors and AI algorithms
A decision support system is generally defined as an analytical system that uses data collected by CGM biosensors to provide personalized recommendations to patients. [12] One important application of decision support systems for diabetes management is glucose prediction. Glycemic control is a balance between the complications of hyperglycemia and the short-term danger of hypoglycemia. CGM provides great help in glycemic control and gives rise to glucose prediction which can prevent hypoglycemic and hyperglycemic events. Integration of AI algorithms with CGM biosensors can bridge the gap between data acquisition and analysis and achieve improved therapeutic accuracy. For example, Marcus et al. presented an SML algorithm based on CGM biosensor data to analyze the glucose level of 11 T1D patients aged 18-39. Using the SML algorithm, the hypoglycemia prediction rate was up to 64%, which indicates that it can improve blood glucose control. [137] Georga et al. proposed a support vector regression (SVR) algorithm for the multivariate predictive analysis of subcutaneous glucose of T1D patients. The SVR algorithm is good at solving nonlinear regression problems and increases the accuracy of both short-term and longterm glucose prediction. [138] A real-time prediction ANN algorithm was proposed by training the data only from the CGM biosensor and based on the result of the rootmean-square error of the proposed model which thus shows that the ANN algorithm was evidently more accurate than the autoregressive model. [139] In addition to AI algorithms, the calibration performance of CGM biosensors is also crucial to improve glucose prediction accuracy.

| Calibration of CGM biosensors based on AI algorithms
Most commercialized CGM biosensors are sensitive to glucose substances and transform substance concentration into electrical signals for detection. This electrical signal is then converted to an estimated glucose concentration by a calibration process using previous blood glucose (selfmonitoring of blood glucose [SMBG]) data. [140] The firstgeneration CGM biosensors used the linear regression function as the calibration model; however, these simple techniques were unable to meet the requirements of the more complex time-dependent relationship between glucose concentration and electrical signal for diabetes management. Therefore, frequent calibration based on SMBG is necessary to ensure the accuracy of the CGM biosensor. Furthermore, calibration algorithms can amplify the inaccuracy of CGM biosensors and can cause dangerous hyper-/hypoglycemia prediction. To solve these problems, many machine-learning algorithms have been proposed during the last decade. [141,142] For example, Acciaroli et al. proposed a calibration algorithm based on a Bayesian multiple-day framework for subcutaneous glucose sensors. The performance of the algorithm reduces the calibration frequency from 2 to 0.25 day −1 with an accuracy improvement by decreasing the mean absolute relative difference (MARD) from 12.83% to 11.62%. [140] In addition, Vettoretti et al. presented a retrospective fitting algorithm based on constrained deconvolution for outcome metrics computation in clinics. Additionally, the proposed retrofitting algorithm can achieve more accurate estimates of CGM biosensors. [143] To improve accuracy, Lee et al. proposed a run-to-run strategy using data from previous weeks to personalize CGM biosensor calibration parameters such as manipulating the curve slope and intercept as well as changing the mean reference error and sensitivity drift curve. The performance of the CGM biosensor was significantly improved by 25% in the first week while the MARD also decreased by 27%. [144] The machine-learning approach not only improves the performance of electrochemical CGM biosensors but also suggests that it can be used for noninvasive optical CGM biosensors. For example, Anand et al. presented a data-driven machinelearning algorithm for noninvasive blood glucose measurement. The trained domain knowledge clustering technique and AdaBoost algorithm were used to undergo personalized calibration for the glucose monitoring system which achieved the final MARD under 7.3%. [145] With the rapid development of calibration algorithms, the next generation of CGM biosensors is geared towards factory-calibrated or calibration-free. For example, FreeStyle Libre is a factory-calibrated CGM biosensor that can be used for up to 14 days without fingersticks. The next-generation Dexcom CGM biosensor can register a calibration-free approach when applied to an online Bayesian calibration algorithm. [146]

| CONCLUSION AND OUTLOOK
In conclusion, this paper reviewed the evolution of CGM, particularly highlighting the combination of AI and CGM performance. Emphasis is placed on the three main application scenarios of AI technologies in diabetes management: closed-loop control algorithms, glucose predictions based on CGM biosensors and AI algorithms, and calibrations of the CGM biosensor based on AI algorithms. Essentially, CGM biosensors have the potential to revolutionize patient care in diabetes management and other disease treatments. The main barriers that limit the use of CGM biosensors are related to the cost of supplies (35.3%), inaccurate (30.1%), and disliking devices on the body (29.7%). [136] Nevertheless, the developments in CGM technologies are advancing towards miniaturization, flexibility, long-term, calibration-free, and closed-loop.
Despite the rise of CGM has been one of the most transformative developments in diabetes management over the past few decades, there are several remaining for realizing the goal of AI biosensors for CGM. First, it is well known that there is a lag time between blood glucose and ISF glucose, which is dependent on the physiological lag time and the performance of the CGM sensor. [147,148] Second, due to most CGM sensors can be worn for up to 14 days, a calibration algorithm is required for the insulin pump after each sensor change. Meanwhile, the cost is another factor to consider for T1D. Furthermore, it is not only necessary to take closedloop decisions depending on CGM sensors but also to learn from the data and adaption. Nonetheless, AI is already in development for biosensors, preparing and establishing advances in the medical field for the next technological wave. In addition to the three aspects of AI improving the performance of CGM mentioned above, the rich data sets generated by continuous monitoring help lead the way in personalized medicine in the near future. Ultimately, the closed-loop therapy technology is the perfect embodiment of CGM and AI, providing numerous clinical opportunities and technological advancements in both fields of artificial intelligent biosensors and medicine.