GLUCOSE LEVEL

FIELD OF THE INVENTION

[001] The present invention relates to a system and method for estimating H bAlC and blood glucose levels in a subject in a non-invasive manner using m ultistage modeling that takes account of the subject's physiological state.

BACKGROUND OF THE INVENTION

[002] Current commonly utilized systems and methods for estimating H bAlC a nd blood gl ucose level require invasively obtaining blood samples. To avoid the pain a nd inconvenience associated with this invasive way of estimating HbAlC and blood glucose, there is a need for a non-invasive system and method for estimating H bAlC a nd blood glucose level.

[003] All known systems and methods for estimating HbAlC level require invasively obtaining blood sam ples. As for estimating blood glucose level, there does exist some non-invasive systems a nd methods. For exa mple, US13/128,205 patent a pplication by Ri poll et al . discloses a system for estimating blood glucose level usi ng an estimation model based on pulse wave a nd its energy using a fixed length vector in addition to other factors such as age, sex, height and weight.

However, Ri poll et al.'s invention does not teach taking account of physiological state of the subject when estimating blood glucose. Significantly, we have discovered that a subject's physiological state is an importa nt factor in estimating H bAlC a nd blood gl ucose.

[004] There exist non-invasive systems and methods for estimating blood glucose that do ta ke in account of physiological state of the subject. For example,

US6,968,221 patent by Rosenthal discloses a method a nd apparatus for

non-invasively measuring blood glucose level using estimation model built based on linea r regression a nalysis of i ndicator variables calculated from measurement parameters derived from energy a bsorption measurement data taken from signa l output by two LED signal sources passing through a finger of a subject. With respect to physiological state of a subject, Rosenthal teaches adding a meal factor to its esti mation model to account for post-meal rise in blood glucose without otherwise changing the estimation model. However, there are several drawbacks to Rosenthal's invention and similar i nventions. Firstly, we have discovered that nonlinear models are substantially more accurate than linear models for modeling HbAlC and blood glucose levels. Secondly, multistage nonlinear models of the present invention also provide substa ntially higher accuracy compared to the single stage models. Thirdly, there are physiologica l factors other than intake of food and liquids that are importa nt to incorporate in the estimation of HbAlC and blood glucose levels. For example, we've discovered medical treatment such as administration of diabetes medicine influence the physiological state of the subject, ma ki ng it an important physiologica l factor for estimati ng HbAlC and blood glucose level .

[005] Therefore, there is a need for a noninvasive system and method for estimating HbAlC and blood glucose level which uses multistage nonli nea r models that ta kes into account of importa nt physiological factors that determine the physiological state of the subject. SUMMARY OF THE INVENTION

[006] A H bAlC and blood glucose level estimation system comprising a signa l reader configured to read one or more signa ls from a subject, a state selector configured to assign a physiological state of the subject wherein a physiological state is determined by a combination of physiological factors com prising a fi rst

physiological factor related to fasting or post-mea l condition of the subject and a second physiological factor related to whether the su bject is under medical treatment and a processor configured to perform signal processing on the one or more signals read by the signal reader, wherein the processor comprises a plurality of multistage non linear models each trained to estimate HbAlC or blood glucose level based on data comprising the signal read by the signal reader, the signal processing results a nd the physiological state of the subject. In a n em bodiment, the processor further comprises a feature extractor configured to extract a plurality of extracted features from the one or more signals read by the signal reader a nd results of signal processing by the signal processor.

[007] In a n em bodiment, stage 1 of the multistage nonlinea r model comprises one or more state models that each estimates HbAlC or blood glucose level for a particular physiological state. I n another embodiment, the processor further comprises one or more state model extracted featu re sets that each corresponds to a particular state model. I n a n embodiment, the processor is further configured to categorize a subject into positive bias, centered and negative bias split groups by categorizing the subject in the positive bias group if the state model estimation is higher than a threshold percentage from the corresponding reference value, in the negative bias group if the state model estimation is lower than a th reshold percentage from the corresponding reference value or in the centered group if

B neither of the thresholds are exceeded wherei n the reference val ue is obtai ned using invasive methods requiring blood samples and wherein the th reshold percentage is about 1%, a bout 5% or about 10%.

[008] In a n em bodiment, stage 2 of the multistage nonlinea r model comprises a plu rality of split models which comprises one or more positive bias models, one or more centered models and one or more negative bias models wherein the one or more positive bias models each estimates HbAlC or blood glucose level for subjects who fall within the positive bias group and a particular physiological state, the one or more centered models each estimates HbAlC or blood glucose levels for subjects who fall within the centered group and a particular physiological state and the one or more negative bias models that each estimates HbAlC or blood glucose levels for subjects who fall within the negative bias group and a particular physiological state.

I n another embodiment, the processor further com prises one or more positive bias model extracted feature sets that each corresponds to a particular positive bias model, one or more centered model extracted feature sets that each corresponds to a particular centered model and one or more negative bias model extracted feature sets that each corresponds to a particular negative bias model. In yet another em bodiment, each of the extracted features sets comprises one or more static features, dynamic features, PPG morphology features and/or ECG morphology features.

[009] In a n em bodiment, stage 3 of the multistage nonlinea r model comprises the split model corresponding to the subject's categorization that is retrained using the most recent reference data and state model HbAlC or blood glucose estimation for the subject wherein the HbAlC or blood glucose estimation obtained from the retrained split model is the fi nal estimation. I n a nother embodiment, stage 3 of the multistage non linear model com prises determining that HbAlC level is the negative bias model esti mation if the negative bias model estimation is higher than about 7 %; however if the negative bias model esti mation is not higher than about 7 % and if the positive bias model estimation is lower tha n about 6 %, the HbAlC level is the positive bias model estimation; however if the negative bias model estimation is not higher than a bout 7 % a nd the positive bias model estimation is not lower than a bout 6 %, the HbAlC level is the center model estimation. In yet another embodiment, wherein stage 3 of the multistage nonlinea r model comprises determining blood glucose level is the negative bias model estimation if the negative bias model estimation is higher tha n about 160 mg/dl; however, if the negative bias model estimation is not higher than a bout 160 mg/dl a nd the positive bias model estimation is lower than about 120 mg/d l, the blood glucose level is the positive bias model estimation; however, if the negative bias model estimation is not higher tha n a bout 160 mg/dl and the positive bias model estimation is not lower than a bout 120 mg/dl, the blood glucose level is the center model estimation.

[010] In a n em bodiment, each state model and each split model comprises a n XGBoost prediction model. In another embodiment, each XGBoost regression is set to hyper para meters of max_depth = 3, number of boosted trees to fit = 100 and L2 regularization term lambda = 1.

[Oil] In a n em bodiment, the static extracted features com prise age, sex, height a nd/or waist size of the subject. I n another embodiment, the signa l read by the signal reader comprises photoplethysmography (PPG) pulsatile signa l. I n yet a nother em bodiment, the signal read by the signal reader comprises

electrocardiogram (ECG) signal . I n a n em bodiment, the one or more signa ls comprises one or more of optica l, mecha nical, electrical, acoustic or thermal signa ls. In another embodiment, esti mation of HbAlC is limited to fasting state.

[012] A method for measuring HbAlC and blood gl ucose level comprising the steps of reading one or more signals ema nating from the subject, processing the one or more signals, assigning physiological state of a subject wherein the one or more physiological states is determined by a combination of physiologica l factors comprising a first physiological factor related to fasting or post-meal condition of the subject and a second physiological factor related to whether the subject is under medical treatment, training a plurality of multistage nonlinear models and estimating H bAlC or blood glucose level using the trained plurality of multistage non linear models based on data comprising the signal read by the signa l reader, the signal processing results and the physiological state of the subject.

[013] In a n em bodiment, the method further comprising the step of extracting one or more extracted features from the one or more signals read by the signal reader, results of signal processing and static data of the subject. I n another em bodiment, the extracted features may comprise static features, dynamic features, PPG morphology features a nd/or ECG morphology features.

[014] In a n em bodiment, stage 1 of the multistage nonlinea r models com prises one or more state models wherein each state model is trained to estimate HbAlC or blood gl ucose level for a pa rticular physiological state. I n an em bodiment, the method further com prising the step of creating one or more state model extracted feature sets each com prising a subset of the extracted features wherein each state model extracted feature set corresponds to a particular state model. In yet another em bodiment, the method further comprises the step of categorizing the subject into positive bias, centered or negative bias group by categorizing the subject in the positive bias group if state model estimation is higher than a threshold percentage from the corresponding reference value, in the negative bias group if the state model estimation is lower than a threshold percentage from the corresponding reference value or in the centered grou p of neither of the thresholds are exceeded whe rein the reference value is obtained usi ng i nvasive method requiring blood sa mple a nd wherein the threshold percentage is about 1%, about 5% or about 10%;

[015] In a n em bodiment, stage 2 of the nonlinear multistage models com prises a plu rality of split models which comprises one or more positive bias models, one or more centered models and one or more negative bias models wherein each positive bias model is trained to estimate HbAlC or blood gl ucose level for subjects categorized within the positive bias group and a pa rticular physiological state, each centered model is trained to esti mate HbAlC or blood glucose level for subjects categorized within the centered group and a particular physiological state and each negative bias model is trained to estimate HbAlC or blood glucose level for subjects categorized within the negative bias group and a pa rticular physiological state. In a nother em bodiment, the method further comprises the step of creating one or more positive bias model extracted feature sets that each corresponds to a particular positive bias model, one or more centered model extracted feature sets that each corresponds to a particular centered model and one or more negative bias model extracted feature sets that each corresponds to a particular negative bias model.

[016] In a n em bodiment, stage 3 of the multistage nonlinea r model comprises retraining a split model corresponding to the categorization of the subject using the most recent reference data for the su bject wherein the HbAlC or blood glucose estimation of the retrained split model is the final estimation. In another

em bodiment, wherein stage 3 of the m ultistage nonli near model comprises determining that HbAlC level is negative bias model estimation if the negative bias model estimation is higher tha n about 7 %; however, if the negative bias model estimation is not higher than a bout 7% and if the positive bias model estimation is lower than about 6 %, the HbAlC level is the positive bias model estimation;

however if the negative bias model esti mation is not higher than about 7 % and the positive bias model estimation is not lower than a bout 6 %, the H bAlC level is the center model estimation. In yet another embodiment, stage 3 of the mu ltistage nonlinear model comprises determi ning that blood glucose level is the negative bias model estimation if the negative bias model estimation is higher than about 160 mg/dl; howeve r, if the negative bias model estimation is not higher than a bout 160 mg/dl and the positive bias model estimation is lower than about 120 mg/dl, the blood gl ucose level is the positive bias model estimation; however if the negative bias model esti mation is not higher than about 160 mg/dl and the positive bias model estimation is not lower than about 120 mg/dl, the blood glucose level is the center model estimation.

[017] In a n em bodiment, each multistage nonlinear model comprises an XGBoost regression. I n another embodiment, each XGBoost regression is set to hyper parameters of max_depth = 3, number of boosted trees to fit = 100 and L2 regularization term lambda = 1.

[018] In a n em bodiment, the static featu re com prises age, sex, height and waist size of the subject. I n another embodiment, the signal read by the signa l reader comprises photoplethysmography (PPG) pulsatile signal. In yet another embodiment, the signal read by the signa l reader comprises electrocardiogram (ECG) signal . I n an em bodiment the one or more signa ls comprises one or more of optica l, mecha nical, electrical, acoustic or therma l signals. I n another embodiment, the measurement of HbAlC is performed only for the fasting physiological state. BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1A depicts a general process flow overview of an embodiment of the method and system for estimating HbAlC and blood glucose level of the present invention using multistage modeli ng.

FIG. IB depicts a general system overview of an em bodiment of the method and system for estimating H bAlC a nd blood glucose level of the present i nvention.

FIG. 2A i llustrates a detailed view of a n em bodiment of the processor of the method and system for estimating HbAlC and blood glucose level of the present invention. FIG. 2B i llustrates a detailed view of a n embodiment of the multistage models of the processor of the method and system for estimating H bAlC and blood glucose level of the present invention.

FIG. 3 depicts examples of extracted features that can be extracted from PPG signals. FIG. 4 depicts the methodology for deriving pulse wave velocity as an extracted feature of the system and method for estimati ng HbAlC and blood glucose level of the present invention.

FIGs. 5A a nd 5B depict an embodiment of the process for training the m ultistage models of the system and method for estimating HbAlC and blood glucose level of the present invention.

FIGs. 6A, 6B, 6C and 6D ill ustrate an em bodiment of the process for creating and training split models of the system and method for estimating HbAlC and blood glucose level of the present invention .

FIGs. 7A a nd 7B depict the process for estimating H bAlC a nd blood glucose levels of an embodiment of the method of the present invention with application of persona lized artificial intelligence learning (PAI learning).

FIG. 7C depicts a fully non-invasive embodiment of the process for estimating HbAlC and blood glucose of the method of the present invention with application of merged split model esti mation logic.

FIG. 8A, 8B and 8C each provides state model HbAlC estimations plotted agai nst reference data in pe rcentage units and correlation between the estimation and the reference data that i llustrate increase in accuracy of state model HbAlC estimations for subjects in fasting condition when diabetes medical treatment physiological factor is taken i nto account. FIG. 8A provides state model HbAlC state model estimations plotted against reference data in percentage units and correlation between the esti mation and the reference data when the datasets are a mix of subjects in fasting condition with as well as without medical treatments. FIG. 8B provides state model HbAlC state model estimations plotted against reference data in percentage units and correlation between the estimation and the reference data only for subjects in fasting condition with medical treatments. FIG. 8C provides state model HbAlC state model esti mations plotted against reference data in percentage units a nd correlation between the estimation and the reference data only for subjects in fasting condition without medical treatments.

FIGs. 9A, 9B, 9C and 9D provide HbAlC estimations plotted against reference data in percentage units a nd correlation between HbAlC estimations a nd the reference data that ill ustrate increase in accuracy of HbAlC estimations for subjects in fasti ng condition without diabetes medical treatment when split models a re used for estimation. FIG. 9A provides HbAlC state model estimations plotted against reference data in pe rcentage units and their correlation for subjects in fasting condition without medical treatments. FIG. 9B provides HbAlC positive bias model estimations plotted against reference data in percentage units and their correlation for subjects in fasting condition without medical treatments. FIG. 9C provides HbAlC negative bias model estimations plotted against reference data in percentage units a nd their correlation for su bjects in fasting condition without medical treatments. FIG. 9D provides HbAlC centered model estimations plotted against reference data in pe rcentage units and their correlation for subjects in fasting condition without medical treatments.

FIGs. 10A, 10B provide PAI applied split model HbAlC estimations plotted against reference data in pe rcentage units and correlation between the estimation and the reference data that i llustrate increase in accuracy of split model HbAlC estimations for subjects in fasting condition with and without diabetes medical treatment when PAI modeling is applied to the split models. FIG. 10A provides H bAlC split model estimations with application of PAI plotted against reference data in percentage units a nd thei r correlation for subjects in fasting condition without medical treatments. FIG. 10B provides HbAlC split model estimations with application of PAI plotted against reference data in percentage units and their correlation for subjects in fasting condition with medical treatments.

FIG. 10C provides merged split model estimation logic applied HbAlC estimations plotted against reference data in percentage units and correlation between the estimation and the reference data that ill ustrate increase in accuracy of split model HbAlC estimations for su bjects in fasting condition without diabetes medical treatment when merged split model estimation logic is a pplied.

FIG. 11A il lustrates a general overview of multistage modeling of the method and system of the present invention for estimating HbAlC leve l for subjects in fasting condition with and without diabetes medical treatment with application of PAI .

FIG. 11B il lustrates a general overview of mu ltistage modeling of the method and system of the present invention for estimating HbAlC leve l for subjects in fasting condition with and without diabetes medical treatment with application of merged split model esti mation logic.

FIGs. 12A, 12B and 12C each provides state model blood glucose estimations plotted against reference data in mg/d L units and correlation between the estimation and the reference data that ill ustrate increase in accuracy of state model blood glucose estimations for subjects not undergoing medical treatments when fasting and post-meal physiological factors are taken into account. FIG. 12A provides state model blood glucose state model esti mations plotted against reference data i n mg/dL units and correlation between the estimation and the reference data when the dataset is a mix of subjects not undergoi ng diabetes medical treatment but in both fasting condition and post-meal conditions. FIG. 12B provides state model blood gl ucose estimations plotted against reference data in mg/d L units and correlation between the estimation and the reference data for subjects in post-meal condition not undergoing dia betes medical treatment. FIG. 12C provides state model blood glucose estimations plotted against reference data in mg/d L units and correlation between the estimation and the reference data for subjects in fasting condition not undergoing dia betes medical treatment.

FIGs. 13A, 13B, 13C and 13D provide estimations plotted against reference data in mg/dL units and correlation between blood glucose estimations and the reference data that illustrate some increase in accu racy of blood glucose estimations for subjects in fasting condition without diabetes medical treatment when split models a re used for estimation. FIG. 13A provides state model blood glucose state model estimations plotted against reference data in mg/dL units and correlation between the estimation and the reference data for subjects in fasting condition without medical treatments. FIG. 13B provides positive bias model blood glucose estimations plotted against reference data in mg/dL units and correlation between the estimation and the reference data for subjects in fasting condition without medical treatments. FIG. 13C provides negative bias model blood glucose estimations plotted against reference data in mg/dL units and correlation between the estimation and the reference data for subjects in fasting condition without medical treatments. FIG. 13D provides centered model blood glucose estimations plotted against reference data in mg/dL units and correlation between the estimation a nd the reference data for subjects in fasting condition without medical treatments. FIGs. 13E, 13F, 13G and 13H provide estimations plotted against reference data in mg/dL units and correlation between blood glucose estimations and the reference data that illustrate some increase in accu racy of blood glucose estimations for subjects in post-meal condition without diabetes medica l treatment when split models are used for estimation. FIG. 13E provides state mode l blood glucose estimations plotted against reference data in mg/dL units and correlation between the estimation and the reference data for subjects in post-meal condition without medical treatments. FIG. 13F provides positive bias model blood glucose estimations plotted against reference data in mg/dL units and correlation between the estimation and the reference data for subjects in post-meal condition without medical treatments. FIG. 13G provides negative bias model blood glucose estimations plotted against reference data in mg/dL units and correlation between the estimation and the reference data for subjects in post-meal condition without medical treatments. FIG. 13H provides centered model blood glucose estimations plotted against reference data in mg/dL units and correlation between the estimation a nd the reference data for subjects in post-meal condition without medical treatments. FIGs. 14A and 14B provide PAI applied split model blood gl ucose estimations plotted against reference data in mg/d L units and correlation between the estimation and the reference data that ill ustrate increase in accuracy of split model blood glucose estimations for subjects not undergoing diabetes medical treatment in fasting and post-meal conditions, respectively, when PAI modeling is a pplied to split models.

FIG. 14A provides blood glucose split esti mations with application of PAI plotted against reference data in mg/d L units and thei r correlation for subjects in fasting condition without medical treatments. FIG. 14B provides blood glucose split model estimations with application of PAI plotted against reference data in mg/dL units and their correlation for subjects in post-mea l condition without medical treatments.

FIGs. 14C and 14D provide PAI applied split model blood gl ucose estimations plotted against reference data in mg/d L units and correlation between the estimation and the reference data that ill ustrate increase in accuracy blood gl ucose estimations for subjects currently undergoi ng diabetes medical treatment in fasting and post-meal conditions, respectively, when PAI modeling is applied to split models. FIG. 14C provides blood glucose split model estimations with application of PAI plotted against reference data in mg/dL units and their correlation for subjects in fasti ng condition with medical treatments. FIG. 14D provides blood glucose split model estimations with application of PAI plotted against reference data in mg/d L units and their correlation for subjects in post-mea l condition with medical treatments.

FIGs. 14E and 14F provide merged split model estimation logic applied split model blood gl ucose estimations plotted against reference data in mg/d L units and correlation between the estimation and the reference data that illustrate increase in accuracy of split model blood glucose estimations for subjects not undergoi ng dia betes medical treatment i n fasting a nd post-meal conditions, respectively, when merged split model estimation logic is applied to split models. FIG. 14E provides blood gl ucose split model estimations with a pplication of merged split model estimation logic plotted against reference data in mg/dL units and thei r correlation for subjects in fasting condition without medical treatments. FIG. 14F provides blood gl ucose split model estimations with a pplication of merged split model estimation logic plotted against reference data in mg/dL units and thei r correlation for subjects in post-meal condition without medical treatments.

FIG. 15A il lustrates a general overview of multistage modeling of the method and system of the present invention for estimating blood glucose level for subjects in both fasting and post-meal conditions as well as with a nd without diabetes medical treatment with PAI applied.

FIG. 15B il lustrates a general overview of mu ltistage modeling of the method and system of the present invention for estimating blood glucose level for subjects in both fasting and post-meal conditions as well as with a nd without diabetes medical treatment with merged split model esti mation logic applied.

FIG. 16 illustrates a sum mary of accuracy of va rious modeling stages of the method a nd system of the present i nvention.

DETAILED DESCRIPTION OF THE INVENTION

[019] It is to be understood that the following detailed description a re exemplary a nd explanatory only and a re not restrictive of the i nvention, as claimed .

[020] In general, the terms used in the following disclosure should not be construed to limit the technology to the specific em bodiments disclosed in the specification, u nless the detailed description explicitly defi nes such terms.

Accordingly, the actua l scope of the technology encompasses the disclosed em bodiments and all equivalent ways of practicing or im plementing the technology.

[021] As used herein, the term "about" as a modifier to a quantity is intended to mean + or - 10% inclusive of the quantity being modified.

[022] As used herein the term "physiologica l factor" is intended to refer to any factors that can affect the physiological state of the subject so as to influence estimation of HbAlC and/or blood glucose levels, including but not li mited to whether the subject is in a fasting or post mea l condition or whether or not the subject is undergoing medical treatment such as administration of dia betes medicine. A combi nation of one or more physiological factors determines a particular physiological state of the subject.

[023] Fig. 1A illustrates general flow of an embodiment of the system and method of the present invention. As shown in Fig. 1A, an embodiment of the present invention estimates HbAlC and blood glucose levels in a non-invasive ma nner based on the physiologica l state of a subject using multistage models.

Stage 1 of the multistage models com prises the state models 231 whereas stage 2 of the multistage models com prises split models 233, 234 and 235 which will be described in further detai l below in connection with Fig. 2B. Moreover, an additional stage 3 modeling may be applied. There are two embodiments of stage 3 modeling that may be used. The first embodiment of stage 3 modeling is

persona lized artificial intelligence learning (PAI Learning) that personalizes split models 233, 234 or 235 to the subject to improve accuracy of H bAlC a nd blood glucose level esti mation as will be described in further detail in connection with steps 1060, 1070 and 1080 of Fig. 7B and ill ustrations of Figs. 1A, 11A a nd 15A. The second e m bodiment of stage 3 modeling involves application of merged split model estimation logic to improve accu racy of HbAlC and blood glucose level estimation as wil l be described in fu rther detail in connection with step 1075 of Fig. 7C and illustrations of Figs. 1A, 11B and 15B.

[024] The system for estimating HbAlC and blood glucose levels of the present invention is illustrated in fu rther detail in Fig. IB. As shown i n Fig. IB, the system of the present invention comprises subject 100, signal device 110, connector 120 and processor 200. The subject 100 comprises a person whose HbAlC and/or blood glucose level is to be estimated using the system and method of the present invention.

[025] As discussed above, the subject 100 may be in one of various physiological states wherei n each physiological state is determined by a combination of one or more physiological factors such as fasting or post-meal condition and/or whether the subject is undergoing any medical treatment that affects HbAlC or blood glucose estimations. I n one embodiment, each physiological factor is defined by a unique combination of physiological factors. Fasting condition of the subject 100 may be defined as the condition in which subject 100 has not had any intake of food or liquids for at least about 4 hours, at least a bout 6 hours or at least about 8 hours.

The post-meal condition of the subject 100 may be defined as the condition in which the subject 100 has had a n intake of food and/or liquids at a bout 120 minutes prior to the use of the present system a nd method for estimating H bAlC a nd blood glucose level. In addition, the medical treatment physiological factor may be defined as whether the subject 100 is undergoing medical treatment that affects HbAlC and or blood glucose estimation such as admi nistration of diabetes medicine like oral administration of sulfonylureas etc... or injection of insulin for example.

[026] I mportantly, it is common ly understood in the field of HbAlC or blood glucose esti mation that intake of food and liquids can affect blood glucose level, but the com monly used finger pricking method or drawing of blood for measuring blood glucose level remains accurate since it is based upon electrochemica l detection principa ls. Therefore, it is widely accepted that accurate measurement of blood glucose level using the existing invasive method and system is not affected by fasting and post-meal conditions.

[027] In contrast, we have unexpectedly discovered that the above com monly accepted knowledge does not apply to non-i nvasive systems and methods for measurement of H bAlC and blood glucose level such as the present invention.

Specifically, we unexpectedly discovered that the subject's fasting condition or post-meal condition has substantial influence on accuracy of non-invasive system and method for estimating HbAlC and blood glucose level as il lustrated by Figs. 12A, 12B, a nd 12C. As seen in Fig. 12A, correlation of state model 231's blood glucose estimation with reference data obtained using commonly available finger prick system and method is 0.600 in Fig. 12A in which data from fasting condition and post-meal condition a re mixed together. However, if fasting and post-mea l data are sepa rated, the correlation increases to 0.680 and 0.796 as shown in Figs. 12B a nd 12C, respectively.

[028] In addition, we have also unexpectedly discovered that medical treatments a lso substantially influence accuracy of H bAlC and blood glucose estimations as illustrated by Figs. 8A, 8B and 8C. Fig. 8A shows correlation of 0.648 for state model estimations a nd reference data for a mix of data from subjects that are undergoing medical treatments as well as subjects that are not undergoing medical treatments. Whe n estimations of su bjects undergoing medica l treatments a re separated from estimations of subjects not undergoing medical treatments, the correlation increases to 0.707 a nd 0.772 as shown in Figs. 8B and 8C, respectively. [029] Therefore, the system and method of the present invention accounts for physiological state of a subject to substantially improve accuracy of blood glucose level and HbAlC level estimations.

[030] Fig. IB illustrates in further detai l an embodiment of the system for measuring HbAlC and blood glucose level of the present invention. As shown in Fig. IB, the signal device 110 comprises a signal reader 112, signal emitter 114, one or more ECG electrodes 116 a nd a signal module 118. The signal reader 112 is configured to read signals ema nating from the subject 100. The signal may comprise optical, mechanical, electrical, acoustic, thermal signa l or a combination thereof. I n one em bodiment, the signal reader 112 comprises a photo diode configured to read signal emanating from the finger of the subject 100. The signal device 110 may further comprise a signal emitter 114 for outputting signals that pass through the body of the subject 100, then emanate out from the subject 100 to be read by the signal reader 112. Since light with wavelengths in the near infra red and visible red spectrum pass through the human body more easily tha n other

wavelengths, in an embodiment, the signal emitter 114 com prises one or more LED capa ble of outputting signa ls with wavelength in the nea r infrared and/or the red visible spectrum. In one embodi ment, the wavelength of the signal is between a bout 400 nm to a bout 1mm or, more specifical ly, about 660nm, about 532nm or a bout 940n m . I n another embodiment, the signal emitter 114 com prises one or more LEDs ca pable of outputting signal with more than one wavelengths such as a combination of light with wavelengths of about 660n m and about 532nm.

[031] Si nce the signal that the signal reader 112 reads passes through the subject 100, it contains information about the subject 100. For example, in a n em bodiment of the system of the present invention, when signal output from the signa l emitter 114 passes through part of the subject 100's body such as the subject's finger, the signal emanating from the subject 100's finger wou ld vary in intensity due to periodic cycles of the ca rdiovascular/circulatory system of the subject 100 that cha nges the volume and density of blood flowi ng through bloodstrea m. This change in the intensity of the signal that ca n be read by the signal reader 112 represents photoplethysmography (PPG) pulsatile signa l of the subject 100. As presence of glycosylated hemoglobin or glucose in the blood may a lter the way blood flows through the cardiovascular/circulatory system by affecting properties of blood such as its density and viscosity, it follows that presence of glycosylated hemoglobin or glucose in the blood can influence the PPG signal, a llowing the system and method of the present invention to estimate HbAlC a nd blood glucose levels using information derived from the PPG.

[032] In a nother embodiment of the present invention, in addition to capturi ng PPG signa l of the subject 100, the signa l device 110 may further be configured to capture electrocardiogram (ECG) signal of the subject 100. ECG signa l of the subject 100 also provides usefu l information regarding subject 100 for estimating HbAlC a nd blood gl ucose levels by measuring electrical activity of the cardiovascu la r system of the subject 100. Moreover, pulse wave velocity (PWV) information that reflects intrinsic properties as well as contents of blood vessels of the subject 100 may be derived from PPG and ECG signa ls. PWV is discussed in further detail below in connection with Fig. 4. To obtain ECG signal, the signal device 110 may com prise one or more electrodes 116 for ca ptu ring the ECG signal of the subject 100. In one em bodiment, the electrodes 116 may be placed at each wrist of the subject 100.

[033] The signal device 110 further comprises a signal module 118 configured to comm unicate with a nd control the signal reader 112, signal emitted 114 and the one or more ECG electrodes 116. The signal mod ule 118 may also comprise one or more control panels that al low a user to control incomi ng and outgoing signals such as triggering signals and/or capturing signa ls. I n one em bodiment, the signal module 118 may comprise one single module configured to control and capture both PPG and ECG information from the subject 100. In a nother em bodi ment, the signa l module 118 may comprise a module for ca pturing PPG information and a separate module for capturing ECG information.

[034] Connector 120 is configured to allow com mu nication between the signal device 110 and the processor 200. I n an embodiment, the connector 120 may transmit the signal read by the signal device 110 to the processor 200 as well as transmit com mands from the processor 200 to the signa l device 110 to comma nd the signal module 110 to trigger and/or read signals. In one embodiment, the connector 120 may be a physical wire. I n another em bodiment, the connector 120 may be a wireless connection such as those using Wi-Fi or Bluetooth technology.

[035] Figs. 2A and 2B ill ustrate processor 200 in further detail. As shown in Fig. 2A, processor 200 comprises an ana logue to digital converter (A/D converter) 220, signal processor 222, feature extractor 224, state selector 226, a set of multistage models 230, display 240 and memory 250.

[036] Fig. 2B illustrates the set of m ultistage models 230 in fu rther detail. As shown in Fig. 2B, the set of stage 1 models 230 comprises one or more state models 231a, 231b, etc.... and one or more corresponding state model extracted features

232a, 232b, etc . The set of multistage models 230 fu rther comprise stage 2 models comprising a plurality of split models comprising positive bias models 233a, 233b, etc...., centered models 234a, 234b etc.... and negative bias models 235a, 235b etc.... as well as one or more correspondi ng positive bias model extracted features 236a, 236b, etc...., centered model extracted features 237a, 237b, etc.... a nd negative bias model extracted features 238a, 238b, etc....

[037] It shou ld be noted that com ponents of the processor 200 described may reside in one si ngle device as illustrated in Fig. 2 or reside in sepa rate devices or in a cloud. For example, the a nalogue to digital converter 220, the signal processor 222, the feature extractor 224 and/or memory 250 may each or in various com bi nations be a stand-alone device or part of a cloud. Furthe rmore, display 240 which is configured to a llow user input into processor 200 such as via onscreen input as well as display information such as physiologica l factor, physiological state and/or FlbAlC a nd blood glucose level estimation may be a sta nd-a lone device.

[038] The A/D converter 220 is configured to digitize the analogue signal transmitted to the processor 200 into digitized signal 252 which may be stored in memory 250. The signa l processor 222 is configured to process the digitized signal 252 to facilitate extraction of features from the signal as known in the art. For example, in a n embodiment, the signal processor 222 may be configured to decompose the signal into AC and DC components, perform Fourier transformation, etc... in order to facilitate analysis and further processing of PPG and/or ECG signals such as extraction of features from the digitized signal 252 as described in further detail below in connection with the feature extractor 224. The signal processi ng results 254 may be stored in memory 250.

[039] The feature extractor 224 is configured to extract features from the digitized signal 252 and/or results of the signal processing 254. In an embodiment, extracted features 256 may com prise features from multiple categories of

information such as static, dynamic, PPG morphology information and ECG

morphology information. The static extracted featu res may comprise the subject 100's static data such as age, height, waist circumference, weight, etc .

[040] Dyna mic extracted features may comprise a subject 100's dynamic data such as hea rt rate, heart rate variance, PWV, PWV variance etc.... PPG and ECG morphology extracted features may comprise information derived from PPG data measured from the subject 100 such as ascending time, plateau time variance,

AC/DC, AC/DC va riance etc.... Figs. 3 and 4 illustrate several exam ples of the many possible extracted features 256. As shown in Figs. 3, the extracted features 256 may comprise ascending time of the signal, the maximum slope of the signal, plateau time of the signa l, AC component, AC variation, DC component, etc.... Additional possible extracted features 256 that a re of the dicrotic sequence type features.

[041] Furthermore, Fig. 4 illustrates PWV feature that may be calculated using both PPG and ECG signals. As shown in Fig. 4, the PWV feature is calculated as the a pproximate distance between the hea rt a nd location of the ECG electrodes 116 such as the wrist divided by the time difference between the pea k of the ECG signal and the leading edge of the PPG signal . I n a n em bodiment, the location of the ECG electrodes 116 are assumed by the system to be located at the subject 100's wrists a nd the distance between the wrist and the hea rt may be estimated by the processor based on height of the su bject 100, which is one of the static extracted feature entered into the system discussed a bove. I n another embodiment, the distance between the subject 100's wrist and heart can be entered as a static extracted data entered ma nually into the system via display 240. In addition, PWV variance feature may also be derived from the PWV feature. Since PWV relates to the way in which contents of blood vessels flow th rough them, it provides information regarding intrinsic properties of the subject 100's blood vesse ls as well as contents of the blood vessels. All the extracted features 256 may be stored in memory 250. [042] The state selector 226 is configured to assign a particular physiological state to a particula r signa l and related dataset. I n an embodiment, a physiological state is defined by a combination of two physiologica l factors: the first physiological factor is based on whether the subject is in fasting or post-meal condition in combination with the second physiological factor based on whether or not the subject is under diabetes medical treatment. Therefore, in this particula r e m bodiment, there are 4 possi ble physiological states: 1. fasti ng condition with dia betes medical treatment, 2. fasting condition with no dia betes medical treatment, 3. post-meal condition with diabetes medica l treatment and 4. post-meal condition with no diabetes medical treatment. In a n embodiment, the state se lector 226 may be triggered manually by a user such as the subject 100. For example, the user may toggle one or more onscreen buttons displayed on the display 240 to indicate various physiological states. The selected state 258 of a dataset may be stored in memory 250.

[043] Each state models 231 as well as the positive bias models 233, centered models 234 a nd negative bias models 235 each comprises a model for estimating HbAlC or blood glucose level for a physiological state as indexed by suffix a, b, etc.... In an embodiment, the suffix "a" indicates a physiological state of fasting condition with dia betes medical treatment, the suffix "b" indicates fasting condition without dia betes medical treatment, the suffix "c" i ndicates post-meal condition with dia betes medical treatment and the suffix "d" indicates post-mea l condition without dia betes medical treatment. Each of the state model 231 as well as the positive bias models 233, centered models 234 and negative bias models 235 is trained for a particular physiological state of the subject 100 as indicated by the physiological state suffix so that each model is capa ble of estimating HbAlC or blood glucose level for a subject 100 in that particular physiological state as will be explained in further detail in connection with Figs. 5, 6 and 7. It should be noted that we have found that HbAlC measurements are most accurate in fasting condition. Therefore, in one em bodiment, HbAlC measurements are only made in fasting condition as shown in FIG IB a nd FIG. 11A and 11B.

[044] Si nce we have found that nonlinea r models model HbAlC and blood glucose concentration substa ntially more accurate than linear models, in an em bodiment, each state model 231 a nd split models 233, 234 and 235 com prises one or more non linear models such as support vector machine, random forest, XGBoost, neural network, etc.... Each state model 231 has a corresponding set of state model extracted features 232. A set of state model extracted features 232 is a subset of extracted features 256 a nd may com prise a combination of static, dyna mic, PPG morphology a nd/or ECG morphology extracted features that optimizes the state model 231 duri ng the training process of the state model 231. Simila rly, a set of split model extracted features 236, 237 a nd 238 is a subset of extracted features 256 and may com prise a combination of static, dyna mic, PPG morphology and/or ECG morphology extracted features that optimize the split models 233, 234 and 235 during the training process of those models. In an embodi ment, each state model 231 and split models 233, 234 and 235 comprises an XGBoost predictive model with max depth=3, nu mber of boosted trees to fit n_estimate =100 and L2 regularization term lambda=l. ¹ The training process for optimizing the models is described in further detail in connection with Fig. 5 below.

[045] Fig. 5A-B il lustrate the method for traini ng HbAlC and blood glucose level measurement of the present invention. As shown in Fig. 5A, the method of training

¹ The following website provides information and usage information for XGBoost:

http://xgboost.readthedocs.io/en/latest/model.html the present invention begins with obtai ning a set of training data set for optimizing prediction models 231, 233, 234 and 235. This com prises step 1000 of acquiring PPG and/or ECG signa ls from a training popu lation using signa l device 110 as wel l as step 1005 of inputting each member of the training population's static data such as BMI, age, waist size, height, weight, etc.... I n addition, step 1000 involves obtaining reference data. In one em bodi ment, the reference data set comprises HbAlC a nd blood gl ucose level estimates obtained from each member of the training population using cu rrently available invasive system and methods for obtaining HbAlC and/or blood gl ucose involving finger pricki ng or d rawing of blood to obtain blood sa mples.

In one embodiment, there are more than 100, 500 or 1000 people in the training population. It is prefera ble that the training population com prises a diversity of people of various sexes, ages a nd physical conditions.

[046] If the PPG and ECG signals are collected in analogue format, the signals a re digitized in step 1010 by A/D converter 220. The digitized signal may be stored in database 250 as part of step 1010. Next, the digitized signal is subjected to signa l processing in step 1020 by the signal processor 222. I n one embodiment, as mentioned above, the signal processing step 1020 may comprise decom posing the signal into high frequency pa rt a nd low frequency part, such as separating its AC component from the DC com ponent. In another embodi ment, the signa l processing step 1020 may comprise transforming the signal using transformation methods such as Fourier tra nsformation, wavelet transformation, Hilbert-Huang transformation or a ny other transformation related to any time-frequency ana lysis. Results of the signal processing 1020 may be stored in memory 250 as part of step 1020.

[047] Next, step 1025 involves determining the physiological state of the subject

100 via state selector 226. In a n em bodiment, the step 1025 of assigni ng physiological state may be performed manually. For example, as mentioned a bove, step 1025 may be performed by the subject 100 toggling one or more on-screen buttons displayed on the display 240, each button related to a pa rticular

physiological factor which in combination indicate a pa rticular physiological state.

In an embodiment, an on-screen button may be toggled to i ndicate fasting or post-meal condition a nd a second button may be toggled to indicate whether or not the pa rticular member of the training population is undergoing diabetes medical treatment. Each of the collected a nd digitized traini ng data in steps 1000, 1010 a nd 1020 is tagged with a particula r physiologica l state assigned in step 1025 so that each estimation models 231, 233, 234 and 245 may be trai ned for a pa rticular

physiological state so that they are each specialized to esti mate HbAlC or blood glucose level for subjects in that particular physiological state.

[048] Next, in step 1030, the feature extractor 224 extracts extracted features 256 from the digitized signal 252 and/or the processed signal 254 such as PWV, heart rate, heart rate variance, etc.... The extracted features 256 may be stored in memory 250 as pa rt of step 1030.

[049] After extracting the extracted features 256, estimation models 230 may be trained in steps 1040 to estimate HbAlC and blood glucose level using the training data obtai ned and derived in prior steps and selecting a set of state model extracted features 232. Training methods are known in the art such as discussed in the reference cited in footnote 1. Training should result in high correlation between reference data and corresponding state model 231 HbAlC or blood glucose estimation. Each training data incl uding each reference value is associated with a particular physiological state. The result is that each state model 231a, 231b, etc.... is trained to estimate HbAlC or blood glucose level for a particular physiological state as indicated by the alphabetic suffix a, b, etc .

[050] Next, in step 1050, the training data including signals and data taken and processed and reference values collected in steps 1000-1040 is categorized into a particular split group: positive bias, center or negative bias group. The

categorization step 1050 is performed by comparing trained state model 231's HbAlC or blood glucose estimation to the corresponding reference data as illustrated i n Fig. 6A. Specifically, if the state model 231's HbAlC or blood gl ucose estimation is higher by more than a percentage threshold above the corresponding reference value as indicated by area 1 in Fig. 6A, then that training data is categorized within the positive bias group and positive bias training data set, respectively. Simi la rly, if the state model 231 HbAlC or blood glucose estimation lower than a threshold percentage below the corresponding reference value as indicated by area 3 in Fig. 6A, then that training data is categorized within the negative bias group and negative bias training data set, respectively. The remaining training data is categorized within the centered group and centered training data set, respectively. Therefore, each split group training data is associated with a particular type of estimation (HbAlC or blood glucose), a particular physiological state as well as a particular split group. I n an embodiment, the threshold percentage may be a bout 10%, a bout 5% or about 1% of the reference data. Next i n steps 1052, 1055 a nd 1057, each of the three split models are trained using the respective split group training data set so that each split model 233, 234 and 235 would be trained to specialize in estimating HbAlC or blood glucose for a subject in a particular physiological state a nd in a particular split group. Specifically, training split group models involves training each of the positive bias model 233 to estimate HbAlC or blood glucose using on ly the positive bias training data of a particula r physiologica l state, training each of the centered bias models 234 to estimate HbAlC or blood glucose using only the center bias training data of a particular physiological state and training each of the negative bias model 235 to esti mate HbAlC or blood glucose using only the negative bias training data of a pa rticular physiological state.

[051] Therefore, after training, each of the positive bias models 233 is specia lized in estimating H bAlC or blood glucose level for those subjects 100 categorized in the positive bias group in a particula r physiologica l state, each of the centered models 234 is specialized estimating H bAlC or blood glucose level for those subjects 100 categorized in the centered group in a pa rticular physiological state, and each of the negative bias model 235 is specialized in esti mating HbAlC or blood glucose level for those subjects 100 categorized in the negative bias group i n a particular physiological state . The result of training the split model using their respective training set is shown in Fig. 6B, 6C and 6D i llustrating that, after split group training, each split model provides HbAlC or blood glucose estimates that high ly correlate with the reference data. Training methods are known in the art, some of which are described in the reference cited in footnote 1.

[052] In one embodiment, each of the positive bias model extracted feature 236, centered model extracted feature 237 and negative bias model extracted feature 238 is the sa me as the corresponding state model extracted features 232. I n another e m bodiment, each of the positive bias model extracted feature 236, centered model extracted feature 237 and negative bias model extracted feature 238 can differ from each other as well as differ from the corresponding state model extracted features 232.

[053] After training each of the split models, the models are ready to estimate

H bAlC and blood glucose level as illustrated in Fig. 7. As shown in Fig. 7A, the method for estimating HbAlC and blood glucose level comprises steps 1000 data acquisition, 1005 user input, 1010 digitize signal, signal processing 1020, determine physiological state 1025 a nd featu re extraction 1030. These are the same steps as the training process except that they are performed for one particular user, subject 100, who is interested in non-invasively estimating HbAlC and/or blood glucose levels. Next step 1050 is also the same step as done in the training phase involving taking H bAlC a nd blood glucose reference values using i nvasive methods such as ones that require blood sample or finger prick so that subject 100 can be categorized within a split group. Importantly, after the subject 100 is properly categorized within the proper split group, no further invasive methods for obtaining H bAlC and blood gl ucose is required unless the occasiona l recalibration is required as discussed below in connection with step 1080. Therefore, step 1050 is only required for the initial estimation for a particular subject 100 or only when recalibration step 1080 is required so that the present invention works largely on a non-invasive basis.

[054] Once the subject 100 has been categorized into split groups in step 1050, persona lized artificial intelligence learning (PAI learning) may be performed in steps 1060, 1070 and 1080. As explained above, PAI comprises the first embodiment of stage 3 mode ling. Specifica lly, in step 1060, the split model 233, 234 or 235 that corresponds to the subject 100's split group determined in step 1050 is retrained using the finger prick reference data for HbAlC and blood glucose obtained in step 1050 and the corresponding state model 231 HbAlC and blood gl ucose estimations in addition to the split grou p training data sets. Therefore, if the subject 100 falls into the positive bias group, then the positive bias model 233 is retrained, if the subject 100 falls into the centered group, then the centered model 233 is retrained a nd if the subject 100 falls into the negative bias group, then the negative bias model 233 is retrained. This retraining step personalizes the split model 233, 234 or 235 for the subject 100, increasing accuracy of the estimation for that the subject 100. Next, in step 1070, HbAlC and/or blood glucose level estimation is then obtained from the retrained split model 233, 234 or 235 and displayed on display 240 as shown in Fig. IB.

[055] Next, step 1080 determines whether the split models need to be

recalibrated as recalibration may be required from time to time due to changes such as changes in a subject 100's body. Specifically, if

.

greater than a

threshold recalibration percentage, then steps 1050, 1060, 1070 and 1080 are performed again to recategorize the subject 100 into the proper split group as well as to retrain the model 233, 234 or 235 corresponding to the selected split group for the subject 100. In an embodiment, the threshold recalibration percentage is 20%.

In another embodiment, the threshold recalibration percentage is 1%, 5%, or 10 %.

[056] In an alternative embodiment with application of merged split model logic that is fully noninvasive is depicted in Fig. 7C. In this embodiment as show in Fig. 7C, steps 1050, 1060 and 1080 are eliminated for a fully non-invasive system and method for estimating HbAlC and blood glucose. In this em bodiment, steps 1000 data acquisition, 1005 user input, 1010 digitize signal, signal processing 1020, determine physiological state 1025 and feature extraction 1030 are performed as described above in connection with Fig. 7A. However, in this fully non-invasive embodiment, since steps 1050, 1060 and 1070 are eliminated because PAI is not applied so that step 1075 immediately follows step 1030. Step 1075, an HbAlC or blood glucose level estimation is obtained from each of the three split models 233, 234 and 235 and merge split model estimation logic is applied. Specifically, the merged split model estimation logic is if the negative bias model 233 is higher tha n a

predetermined threshold value, then that negative bias model estimation is the fina l estimation. However, if the negative bias model 233 is not higher than a

predetermined threshold value but the positive bias model 235 is lower tha n a predetermined threshold value, then the positive bias model estimation is the fina l estimation. Estimation from the centered model 234 is taken as the fina l result if the estimation result of the negative bias model 233 is not higher than the

predetermined threshold value and if the estimation result of the positive bias model 235 is not lower than a predetermined threshold value . In a n embodiment, the thresholds for HbAlC are 7 % for the negative bias model and 6 % for the positive bias model, a nd the thresholds for the blood glucose estimation are 160 mg/dl for the negative bias model a nd 120 mg/dl for the positive bias model. I n this way, the subject 100 neve r has to undergo any invasive method and system for estimating HbAlC and blood glucose levels.

[057] Figs. 8-16 illustrate results of the system and method of the present invention using XG Boost predictive model with max depth=3, number of boosted tree to fit =100 a nd L2 regu la rization term la m bda.

[058] Figs. 8, 9 a nd 10 illustrate H bAlC estimation for subjects in fasting condition with or without undergoing diabetes medical treatment. Fig. 8 illustrates the first stage of the multistage modeling of the present invention for estimating HbAlC. As shown in Figs. 8A, 8B a nd 8C, the correlation between state model estimations a nd the reference data improves from 0.648 as shown in Fig. 8A to 0.707 a nd 0.772 as shown in Figs. 8B and 8C respectively if diabetes medical physiological factor is taken i nto account by separating data with a nd without diabetes medical treatment. Fig. 9B, 9C a nd 9E illustrate results of the second stage of the multistage modeling of the present invention. As shown in the figures, split models a re used for estimati ng HbAlC for subjects in fasting condition a nd without diabetes medical treatment. Fi nally, Figs. 10A, 10B and 10C illustrate estimation results of the third stage of the multistage modeling of the present invention. As shown in Figs. 10A and 10B, the correlation im proves to 0.971 a nd 0.949 when PAI modeling of steps 1050-1070 applied for subjects in fasting condition without diabetes medical treatment in FIG. 10 A a nd with medica l treatment in FIG. 10B. Alternatively, FIG. 10C shows results of the fu lly noninvasive embodime nt of the present invention with a pplication of merged split model estimation logic as illustrated in FIG. 7C. As illustrated in 10C, after application of the merged split model estimation logic to the split model esti mations, the correlation improves to 0.816.

[059] Figs. 11, 12 and 13 ill ustrate blood glucose estimation for subjects in fasting or post-meal condition and with or without undergoing diabetes medical treatment. Fig. 11 il lustrates estimation results of first stage of the multistage modeling of the present invention for estimating blood glucose. As shown in Figs. 12A, 12B and 12C, the correlation of estimations by the state model with respect to the reference data improves from 0.600 as shown in Fig. 12A to 0.680 and 0.796 as shown in Figs. 12B a nd 12C respectively if fasti ng (FIG. 12C) and post meal (FIG. 12B) conditions are taken into account when the subject is not u ndergoing dia betes medical treatment. Fig. 13 il lustrates estimation results of the second stage of the multistage modeling of the present invention. As shown in Figs. 13B, 13C and 13D, split models are used for estimati ng blood glucose for subjects in fasti ng condition without diabetes medical treatment. Similarly, as shown in Figs. 13F, 13G and 13H, split models are used for estimating blood glucose for subjects in post-meal condition and not u ndergoing dia betes medical treatment. Fi nally, Figs. 14A, 14B, 14C and 14D illustrate estimation results of the third stage of the multistage modeling of the present invention with application of PAI of steps 1050-1070. As shown in Figs. 14A and 14B, the correlation im proves to 0.961 a nd 0.989 when PAI process is used for subjects not undergoing diabetes medica l treatments in fasting or post-meal conditions, respectively. Simi la rly as shown in Figs. 14C and 14D, the correlation improves to 0.913 and 0.982 when PAI process is used for subjects undergoing dia betes medical treatments in fasting or post-meal conditions, respectively.

Alternatively, FIG. 14E a nd 14F show results of the fully noni nvasive embodiment of the present invention as i llustrated in FIG . 7C. As illustrated in 14E and 14F, after a pplication of the merged split model estimation logic, the correlation improves to 0.829 for fasting condition without medica l treatment and 0.872 for post-meal condition without medical treatment.

[060] Fig. 16 summarizes HbAlC and blood glucose estimation results for various physiological states at various stages of the multistage modeling of the present invention.

[061] Although the present i nvention has been described in terms of specific exempla ry embodi ments and examples, it ca n be appreciated by those skilled in the a rt that cha nges cou ld be made to the examples described above without departing from the broad inventive concept thereof. It is understood, therefore, that this invention is not limited to the particular examples disclosed, but it is intended to cover modifications within the spirit and scope of the present invention as defined by the appended claims.