Glucose utilization rates optimization -
Through patient model-based RL, the policy model can learn individualized treatment trajectories and improve long-term clinical outcome. At the same time, it learns the treatment practices of clinicians in treating patients with T2D within a reasonable range of dosages by SL.
To build a dynamic and individualized AI clinician for managing patients with T2D, we constructed the model-based RL framework. For the comparison of actual state trajectories and model-based state roll-outs, the predicted glucose values follow the transition tendency accurately in both the internal test and the external test set Fig.
For the overall glucose prediction, we aggregated the individual-level prediction to produce population-level results, which were then used for further analysis. The AI model demonstrated good performance in the internal test set, achieving a Pearson correlation coefficient PCC of 0.
When evaluated on the external test set, the AI model achieved a PCC of 0. As shown in Extended Data Table 2 , the results indicate that our model outperformed the other baseline models with a substantial improvement.
a , b , Comparison of actual patient trajectories and model-based state roll-outs for patients from the internal test set a and the external test set b. The blue curve is measured patient glucose values, and the orange curve is predicted glucose values.
c , d , Correlation analysis of the predicted glucose value versus the actual glucose value generated using the AI glucose model in the internal test set c and the external test set d.
Each predicted value is based on the last timestep of the previous day. Box plots show the median center lines , interquartile range hinges and 1. Each value generated by our RL-DITR system represents an individual-level prediction.
These were then aggregated to produce population-level results. AUROC, area under the receiver operating characteristic; ROC, receiver operating characteristic.
When evaluated using the internal test set, the AI model achieved an area under the curve of 0. The model showed reliable performance validated on the external test set.
We further investigated the model performance on predicting daily WTR status glucose values within the target range of 3. We observed that the model becomes more accurate with more information input about a patient as time goes on.
We investigated the correlation between the patient outcome WTR ratio and the cumulative rewards estimated by the patient model. The AI model demonstrated good performance with a Spearman correlation coefficient SCC of 0.
We observed that treatment actions with low cumulative rewards were associated with a low rate of WTR ratio, whereas treatments with high cumulative rewards achieved better glucose outcome with a high rate of WTR ratio. The results show that the patient model evaluation is highly correlated with the clinical outcome and could be used as the interaction environment for the RL model.
Figure 3a,b shows the correlation between the clinician policy and the AI policy in the development phase internal and external test sets. For daily treatment dosage prediction, the AI policy achieved an MAE of 1.
We found that the model becomes more accurate as the observed time window expands due to more trial-and-error interactions with the environment. We aggregated the individual-level predictions to obtain population-level results.
c , d , Comparison of actual treatment regimens and model-based treatment roll-outs of two individual patients from the internal test set c and the external test set d.
The blue curve is measured patient glucose values, and the orange curve is predicted glucose values given by the AI model. e , f , Association analysis of the patient outcome for example, WTR versus the dosage difference in treatment actions between the AI policy and the clinician policy for the internal test set e and the external test set f.
The dose excess, referring to the difference between the given and the AI model, suggested dose summed over per day for all patients. R 2 , coefficient of determination.
MAPE, mean absolute percentage error. Our proposed approach was then tested against several SL methods, including convolutional neural network CNN , long-short term memory LSTM , transformer and the standard clinical method. We found that our model-based RL method was able to export an accurate treatment regimen and outperformed other methods in the internal test set and the external test set Extended Data Table 3.
The results presented in Extended Data Table 3 demonstrate that our policy model, guided by our blood glucose model, outperformed other models substantially.
Figure 3c,d shows the dynamic treatment strategies generated by clinicians and model-based RL for two individual patients on different hospital days. We further investigated whether the patient outcome WTR ratio varied with the difference of the dose actually administered and the dose suggested by the RL method by correlation analysis Fig.
When the dose actually administered differed from the dose suggested by the AI algorithm, the average outcome got worse. For the internal validation cohort, we compared the performance between our AI system and human physicians in giving insulin dosage recommendation using 40 patients with T2D with insulin data points Extended Data Fig.
RL-generated and physician-generated dosage titrations were evaluated by an expert panel, including quantitative metrics and qualitative metrics from clinical experience. Taking the dosage recommended by the expert panel as references, the MAE of the AI system was 1.
Evaluation was based on the expert panel review including effectiveness f , safety g and overall acceptability h. Orange dashed line represents the average performance of AI; blue dashed line represents the average performance of treating physicians. G, group. These results suggest that our AI model is superior to junior physicians and similar to experienced physicians in the overall treatment regimen acceptability, hyperglycemia and hypoglycemia control.
Furthermore, we performed an external validation in 45 patients with T2D to compare the performance of AI plans and treating physician plans under a blinded review by an expert panel and by another blinded review for retesting at 2-week intervals at least Extended Data Fig.
The results demonstrated that the acceptability, effectiveness and safety of the AI plans were similar to the treating physicians who were board-certified endocrinologists, evaluated by subjective measurements made by an expert panel Fig. The percentage of selected superior AI plans was These results demonstrate consistently superior performance of the AI model compared to its physician counterparts.
We used adoption rate to evaluate the percentage of the AI regimens adopted by endocrinologists for patient treatment. Our proposed RL model demonstrated stable performance of effectiveness, safety and acceptability over time, even better in the retest review Fig. The score scale of effectiveness and safety is 1—5.
The adoption rate refers to the percentage of the AI regimens adopted by endocrinologists at the bedside for patient treatment. Intriguingly, a higher adoption rate of Although the adoption rate of the AI plan was relatively low at the initial test review, we found an increase of These results suggested a step-by-step increase of trust of the AI treatment regimen by physicians through human—machine interaction, and the AI system was gradually adopted by physicians into routine clinical practice.
A proof-of-concept feasibility trial was performed to investigate the clinical utility and safety of AI in hospitalized patients with T2D for glycemic control. Sixteen inpatients with T2D were enrolled in the trial Extended Data Fig. Their mean HbA1c was 8. Over the trial, b , The capillary blood glucose of a patient with T2D during the treatment period.
II Mean daily capillary blood glucose. III Mean preprandial capillary blood glucose. IV Mean postprandial capillary blood glucose during the treatment period. The preprandial blood glucose target was 5. c , Average percentage of continuous glucose monitoring data within glycemic ranges throughout the treatment period.
The satisfaction agreement was scored from a scale of 1—5. IQR, interquartile range. of At the end of the trial, A patient example of the seven-point capillary blood glucose during the AI intervention is shown in Extended Data Fig. We also used continuous glucose monitoring CGM for the evaluation of the algorithm-directed glycemic control for the secondary outcomes.
The percentage of glucose concentration in time in range TIR 3. TIR 3. Time spent above Time spent below 3. In addition, glycemic variability was slightly decreased during the treatment period coefficient of variation CV of No episodes of severe hypoglycemia that is, requiring clinical intervention or hyperglycemia with ketosis occurred during the trial.
Most physicians stated that the AI interface is understandable 4. In this study, we developed an RL-based AI system, called RL-DITR, for personalized and dynamic insulin dosing for patients with T2D. We performed development phase validation and clinical validations, including internal validation, comparing AI to physicians using quantitative and qualitative metrics, external validation with test—retest, prospective deployment with test—retest and a proof-of-concept feasibility study with clinical trial.
Taken together, our findings demonstrate that our RL-DITR system has potential as a feasible approach for the optimized management of glycemic control in inpatients with T2D. The management of blood glucose in diabetes remains challenging due to the complexity of human metabolism, which calls for the development of more adaptive and dynamic algorithms for blood glucose regulation.
To address the challenge of personalized insulin titration algorithm for glycemic control, our RL-based architecture is tailored to achieve precise treatment for individual patients, with clinical supervision.
Our proposed patient model-based RL model can make multi-step planning to improve prescription consistency. In addition, because the multi-step plan can be interpreted as the intent of the model from now to a span of time period into the future, it offers a more informative and intuitive signal for interpretation Additionally, our RL-based system delivers continuous and real-time insulin dosage recommendation for patients with T2D who are receiving subcutaneous insulin injection, combining optimal policies for clinical decision-making and the mimicking of experienced physicians Another strength of our study is that we conducted a comprehensive early clinical validation of the AI-based clinical decision-making system across various clinical scenarios.
In clinical deployment, our AI framework offers potential benefits, including automated reading of a large number of inputs from the EHRs, integration of complex data and accessible insulin dosing interface.
Although some algorithms have been developed to assist physicians in insulin titration, only a few have been validated in clinical trials 31 , We conducted a proof-of-concept feasibility trial demonstrating the viability of the RL-DITR system in inpatients with T2D.
Notably, the use of the RL-DITR system resulted in a considerable improvement in blood glucose control, meeting our pre-determined feasibility goal. The percentage of well-controlled blood glucose levels of TIR also demonstrated a substantial increase.
Managing hypoglycemia risk is a key consideration for real-world deployment of the AI system. While achieving improved control of blood glucose levels, the system did not increase the risk of hypoglycemia. Additionally, physicians using the RL-DITR system have reported an increased level of satisfaction, including aspects such as efficiency in clinical practice and perceived effectiveness and safety in glycemic control.
These results suggest that our RL-DITR system has the potential to offer feasible insulin dosing to inpatients with T2D. A large and multi-center randomized controlled trial would help to determine the efficacy and benefits of this clinical AI solution.
Our RL-DITR system was designed as a closed-loop intelligent tool that could use real-time patient data to track blood glucose trajectories and modify treatment regimens accordingly. Furthermore, the RL-DITR system was developed using EHRs of inpatients with T2D, but its generalizability to other populations, such as outpatients, needs further investigation.
We conducted simulated experiments using Gaussian noise to mimic low data quality and dropout 33 to simulate missing data scenarios before deployment Supplementary Fig.
Therefore, although the RL-DITR workflow was implemented and tested for inpatients with T2D, there exists the possibility to extend its application to a wider range of healthcare settings, such as outpatient management, given appropriate integration and continued development.
Although our RL-DITR system has achieved good performance in insulin dosage titration, some challenges remain. The generalization of the AI to other ethnicities needs to be further investigated.
Second, the variety of diet during the hospitalized period was uniformly supplied in the EHRs to build our model.
For patients out of hospital, dietary variation and physical activity should be taken into account and explored by our RL model. We have opened an interface to accumulate dietary information for late updated model.
In conclusion, we developed an RL-based clinical decision-making system for dynamic recommendation of dosing that demonstrated feasibility for glycemic control in patients with T2D. The RL-DITR system is a model-based RL architecture that could enable multi-step planning for patients with long-term care.
With the integration of RL structure and supervised knowledge, the RL-DITR system could learn the optimal policy based on non-optimized data while retaining the safe states by balancing exploitation and exploration.
Furthermore, we performed a stepwise validation of the AI system from simulation to deployment and a proof-of-concept feasibility trial. These demonstrate the RL approaches as a potential tool to assist clinicians, especially junior physicians and non-endocrine specialists, with diabetes management in hospitalized patients with T2D.
To train and validate a computational clinical decision support model, we constructed a large multi-center dataset using EHRs of hospitalized patients with T2D who received insulin therapy from January to April in the Department of Endocrinology and Metabolism, Zhongshan Hospital and Qingpu Hospital, in Shanghai, China.
The demographics and clinical characteristics of patients are presented in Extended Data Table 1 , demonstrating a typical T2D population. We conducted stepwise studies to evaluate the performance of our RL-DITR model version 1. In addition, we performed a proof-of-concept feasibility trial of the RL-DITR system in clinical practice with inpatients with T2D who were admitted for optimization of glycemic control at Zhongshan Hospital ClinicalTrials.
gov: NCT details of proof-of-concept trial protocol provided in Supplementary Information. The retrospective study obtained the following institutional review board IRB approval: Zhongshan Hospital, Shanghai, China R ; XuHui Central Hospital, Shanghai, China and Qingpu Branch of Zhongshan Hospital, Shanghai, China Patient informed consent was waived by the Ethics Committee.
The prospective study and proof-of-concept feasibility trial were approved by the Ethics Committee of Zhongshan Hospital, Fudan University. Each participant provided written informed consent for the prospective study and the proof-of-concept feasibility trial.
For time-series data representation, every patient in the dataset was represented as a temporal sequence of feature vectors. Specifically, each day was broken into seven time periods, including pre-breakfast, post-breakfast, pre-lunch, post-lunch, pre-dinner, post-dinner and pre-bedtime.
All records that occurred within the same period were grouped together and formed a feature set to feed into the RL model as input detailed list of the input features provided in Supplementary Table 2. For structured data, we aligned and normalized them. For free-text notes, we applied a pre-trained language model, ClinicalBERT.
Specifically, we first trained the ClinicalBERT on a large corpus of EHR data. ClinicalBERT is a masked medical domain language model that predicts randomly masked words in a sequence and, hence, can be transformed into downstream tasks.
Then, the ClinicalBERT was fine-tuned for information extraction from free text. We further automatically extracted temporal features from patient clinical records, including clinical observations blood glucose records , a sequence of decision rules to determine the course of actions for example, treatment type and insulin dosage titration and clinical assessment of patients.
The numerical values were extracted from demographics, laboratory reports, blood glucose and medications and further translated with standard units according to the LOINC database.
Then, each numerical value was normalized to a standard normal distribution. In terms of discrete values, all the diagnoses of a patient were mapped onto the International Classification of Diseases-9 ICD-9 and used as discrete features, encoded as binary presence features.
We constructed a large multi-center dataset with a large corpus of 1. ClinicalBERT was fine-tuned on a multi-label dataset to extract 40 symptom labels from medical notes. Phenotype data were extracted from free-text notes of chief history of present illness and physical examination by ClinicalBERT.
Validated on 1, annotated samples from the training set, the results showed that ClinicalBERT could accurately identify the symptom information with an average F1 score of Each extracted symptom label was encoded as a binary presence feature. The process of patient trajectory and treatment decision-making could be formulated as a Markov decision process MDP.
An MDP 34 is a tuple S , A , P , G , γ , where S and A are sets containing the states and actions, respectively; P is a transition function; G is a reward function; and γ is a discount factor. The patient model was learned from historical trajectories, approximating the transition function P and the reward function G and providing support for policy model learning and planning.
The policy model iteratively interacted with the patient model as an environment. At each step, the patient model generated state transition, status prediction and reward estimation based on observed patient trajectories.
The policy model, taking the state as input, generated an action that was fed to the patient model. The patient model updated the states recurrently by an iterative process, enabling the policy model to plan for sequences of actions and find optimal solutions across generated trajectories.
The hidden state would be used as input for patient model and policy model. For patient trajectory tracking, we trained a patient model.
When conducting correlation analysis with daily outcome, Magni risk values were summed for each day. Both of the dynamics function f T and the prediction function f P shared the representation encoder f R when training and inference.
f R was optimized together through backpropagation with the loss to capture meaningful patient representations and dynamics. Each node indicates the states of a patient. The state distribution demonstrated a good cluster hierarchy that individuals in the same cluster are associated with their observable properties diabetes outcome, such as glucose level.
We combined the SL and RL to learn the policy model, with the expert supervision of safe actions to take into account. Specifically, we applied policy gradient optimization for training the policy model π to maximize the returned rewards while incorporating constrained supervision by expert experience.
For the SL part, we used the action made by the clinicians as supervision for policy update. For the RL part, we optimized the policy model π based on the patient model f T , f P as an interactive environment, where a given trajectory was updated recurrently by an iterative process.
The policy model π was trained by both historical and obtained trajectories. We applied a beam search for policy search The top B highest-value trajectories were stored at each timestep, where B was the beam size. The training process involved two stages to optimize the models of our AI system Extended Data Fig.
These functions were jointly optimized through the loss for state transitions and the loss for status prediction. The policy model was trained through a joint optimization process, minimizing both a policy gradient loss on trajectories and a supervised loss that constrains the difference between the recommended action from the policy model and the action taken by the clinician.
We used a transformer-based network with three layers as the representation function used to represent the observations of time-series data, as it has been shown to enable capturing the long dependence in the temporal information of patients The last hidden vector of the output hidden vectors was used for the initial state.
We also applied a transformer network with three layers for dynamics function. The hidden dimension was set to , and the number of multi-attention heads was set to 8.
We used three-layer multi-layer perceptrons MLPs for prediction function, policy function and value function. The hidden dimension was set to The beam size B was set to The models were implemented using PyTorch. The importance sampling for policy evaluation was performed, which enables the evaluation of a target policy using data collected from a distinct policy To enhance the numerical stability of the calculations, we employed WIS along with effective sample size 40 , 41 , which normalizes the trajectories, thereby reducing variance Given specific data conditions, such as no more than k blood glucose measurements per day, we randomly discard blood glucose values within the trajectories to ensure that the remaining trajectories satisfy this criterion.
The traditional clinical methods of insulin dosage titration were used as the standard clinical methods for comparison, consisting of guidelines 42 and consensus formulas 43 , 44 for premixed insulin regimen, basal regimen and basal-bolus regimen. The detailed adjustment was according to the following formula:.
The insulin dosage titration rules of basal-bolus regimen were as follows. Retrospective study phase of the internal cohort. Forty eligible patients with T2D treated with insulin injection were randomly selected from the retrospective EHRs of one of the modeling development hospitals Qingpu Hospital from May to December Two treatment days were randomly selected for each patient, resulting in 80 cases with insulin points Extended Data Fig.
Three physician groups with different levels of clinical experience provided their dose recommendations, and the AI also generated insulin dose recommendations in silico for further evaluation.
An expert consensus panel of three endocrinology specialists conducted blinded review and provided their own recommended insulin dosage. This was used as a reference insulin dosage for each insulin point to assess the accuracy of AI-generated dosage versus the three physician groups.
Retrospective study phase of the external cohort. The retrospective dataset was collected from a non-teaching hospital XuHui Hospital , which included 45 eligible consecutive patients with T2D from April to August Extended Data Fig.
The dataset contained insulin points from cases, and AI-generated dosage was compared to previously delivered insulin dosage by treating physicians human plan for accuracy evaluation.
Next, we randomly selected 40 cases from the dataset to evaluate the acceptability, effectiveness and safety of the AI plan and the previous human plan. The evaluations were blinded head-to-head comparisons of AI versus human plans by the expert consensus with three independent experts.
Prospective deployment study phase. In May , 40 consecutive AI-generated plans were tested for acceptance, effectiveness and safety by endocrinology physicians at the bedside Extended Data Fig.
After determining clinical adoption and ensuring adherence to standard clinical quality controls, the AI insulin regimen was used for patient treatment.
The inclusion and exclusion criteria for patients were consistent across the three phases. Inclusion criteria were patients with T2D treated with subcutaneous insulin injection for at least two consecutive days.
Fed-batch cultivation and PID controllers have been widely used in bioprocess development. Unfortunately, fed-batch cultivation often results in poor nutrient control and wasted nutrients and conventional PID control can lead to oscillating cell behaviors and poor performance under dynamic conditions.
In this study, we have utilized the power of genome-scale metabolic models to predict and control glucose and nitrate supply for C. vulgaris cultures under light and dark cycles and compared this approach to conventional autotrophic and heterotrophic processes.
Our results first showed that utilizing genome scale models to track and limit glucose and nitrate feeding led to higher titers of biomass, FAs, and lutein than autotrophic conditions and more efficient glucose utilization and higher product yields than heterotrophic conditions.
Next, implementing these models into an open loop system modestly improved performance. Finally, both computational simulations and experimental results demonstrated that this genome-based MPC system exhibits superior controller performance compared to conventional PID methods.
Green microalgae C. vulgaris UTEX was obtained from the Culture Collection of Algae at the University of Texas at Austin and maintained on sterile agar plates 1.
Liquid cultures were inoculated with a single colony in For alternating light and dark cycles, autotrophic conditions were used for light sections and heterotrophic conditions were used for dark sections.
The lyophilized algal dry biomass was weighted gravimetrically using an analytical balance. The glucose concentration was measured using YSI biochemistry analyzer Yellow Springs, OH.
FAME production followed the procedure provided by Dong et al. Helium was used as carrier gas. Lutein extraction followed the procedure provided by Yuan et al.
The solution was filtered before HPLC analysis. The mobile phases are eluent A dichloromethane: methanol: acetonitrile: water, 5.
The i CZ model, including six different biomass compositions for autotrophic conditions PAT1-PAT6 and five different biomass compositions for heterotrophic conditions HT1-HT5 , was obtained from Zuniga et al.
GSM simulations were performed using the Gurobi Optimizer Version 5. The experimental setup is shown in Supplementary Fig. The manipulated variables were glucose demand F G and nitrate demand on a per L basis F N for 8-h period. Two pumps were used to control both variables automatically by Matlab TM through Arduino chip.
All the control algorithms were run on Matlab TM and the codes are provided in Supplementary information. The Simulink TM simulation is shown in Fig. The blue box in Fig. Four equations were built inside the blue box as shown in Supplementary Fig.
The inputs were F G and F N. The outputs were biomass, nitrate level, glucose level and volume. Only nitrate levels and glucose levels were fed into the PID and GMPC controller. For the proportional-integral-derivative PID controller, the proportional gain K p , integral gain K i and derivative gain K d equal to 1.
The PID controller and GMPC controller were used to control glucose supply and nitrate supply every hour in both simulation and experiment.
Changes in the setpoint for glucose were introduced to see how both PID and GMPC responded to those changes. Initial biomass levels x 0 , glucose levels G 0 and nitrate levels N 0 were measured as described above and used as inputs into the open-loop system.
Three equations shown below were used to predict biomass growth, nitrate consumption rate, and glucose consumption rate in the open-loop system. The growth rates under light and dark cycles were determined based on previous experimental data. After that, the growth rates were constrained in the autotrophic and heterotrophic GSMs, respectively to determine nutrient exchange rates r N and r G under light and dark cycles.
The methods for using growth rate to estimate nutrient exchange rates have been described previously in Chen et al. We assumed a rapid switch to a new operational steady state following the transition between light and dark cycles.
Initial biomass levels x 0 , glucose levels G 0 and nitrate levels N 0 were measured and used as inputs into the closed-loop system. During the experiment, biomass levels x m , glucose levels G m and nitrate levels N m were msured and used as inputs into the closed-loop system.
For the light cycle, two equations were built to describe and predict biomass accumulation rate and nitrate consumption rate. Unlike the open loop system, the light shielding effect was considered and the growth rate would decrease as the biomass concentration increased as described in the equation below and shown in Fig.
The GSM was used to predict nutrient exchange rate r N based on the measured growth rate. For the dark cycles, three model equations were built to predict biomass accumulation rate, nitrate consumption rate and glucose consumption rate as listed below and shown in Fig.
In the biomass equation, we assumed a fraction of heterotrophic biomass, a , was derived from autotrophic metabolism and the simulated growth rate was μ A.
Meanwhile, some biomass was derived through heterotrophic metabolism with the simulated growth rate, μ H. The nutrient exchange rates r NA , r NH , r GH were determined by inputting simulated growth rates into the autotrophic and heterotrophic GSMs respectively.
where μ A is simulation growth rate from autotrophic metabolism, μ H is the growth rate from heterotrophic metabolism, r NA is nitrate exchange rate from autotrophic metabolism, r NH is the nitrate exchange rate from heterotrophic metabolism, r GH is the glucose exchange rate from heterotrophic metabolism.
Next, we applied a fitting objective function J to minimize the difference between calculated values and simulated model values in order to estimate the optimal parameter values a , μ A , μ H , r NA , r NH , r GH for dictating the actual nitrate and glucose feeds to the bioreactor.
The actual bolus nitrate demand F N and the glucose demand F G were thus determined by using values obtained from this fitting objective function. The data that support the findings of this study are available from the corresponding author upon reasonable request.
Rosenberg, J. A green light for engineered algae: redirecting metabolism to fuel a biotechnology revolution. Article CAS PubMed Google Scholar. Shene, C. Metabolic modelling and simulation of the light and dark metabolism of Chlamydomonas reinhardtii.
Plant J. Kato, Y. et al. Biofuels 12 , 39 Article PubMed PubMed Central Google Scholar. Cheirsilp, B. Enhanced growth and lipid production of microalgae under mixotrophic culture condition: effect of light intensity, glucose concentration and fed-batch cultivation. Zheng, Y. High-density fed-batch culture of a thermotolerant microalga Chlorella sorokiniana for biofuel production.
Energy , — Article CAS Google Scholar. Shi, X. High-yield production of lutein by the green microalga Chlorella protothecoides in heterotrophic fed-batch culture. Bordbar, A. Constraint-based models predict metabolic and associated cellular functions.
Chang, R. Article Google Scholar. Zuñiga, C. Genome-scale metabolic model for the green alga Chlorella vulgaris utex accurately predicts phenotypes under autotrophic, heterotrophic, and mixotrophic growth conditions. Plant Physiol. Loira, N.
Reconstruction of the microalga Nannochloropsis salina genome-scale metabolic model with applications to lipid production. BMC Syst. Chang, L. Nonlinear model predictive control of fed-batch fermentations using dynamic flux balance models.
Process Control 42 , — Jabarivelisdeh, B. Model predictive control of a fed-batch bioreactor based on dynamic metabolic-genetic network models. IFAC-PapersOnLine 51 , 34—37 Juneja, A.
Model predictive control coupled with economic and environmental constraints for optimum algal production. Ogbonna, J. Zuniga, C. Predicting dynamic metabolic demands in the photosynthetic eukaryote Chlorella vulgaris. Li, C.
Utilizing genome-scale models to optimize nutrient supply for sustained algal growth and lipid productivity. NPJ Syst. Tebbani, S. Nonlinear predictive control for maximization of CO2 bio-fixation by microalgae in a photobioreactor.
Bioprocess Biosyst. Hu, D. The design and optimization for light-algae bioreactor controller based on Artificial Neural Network-Model Predictive Control. Acta Astronaut 63 , — White, R.
Long-term cultivation of algae in open-raceway ponds: lessons from the field. Gu, C. Current status and applications of genome-scale metabolic models. Genome Biol. Using genome-scale models to predict biological capabilities. Cell , — Colarusso, A. Computational modeling of metabolism in microbial communities on a genome-scale.
Han, F. Enhancement of microalgal biomass and lipid productivities by a model of photoautotrophic culture with heterotrophic cells as seed. Xiong, W. Double CO 2 fixation in photosynthesis—fermentation model enhances algal lipid synthesis for biodiesel production.
Xiao, Y. Photosynthetic accumulation of lutein in Auxenochlorella protothecoides after heterotrophic growth. Drugs 16 , Vidotti, A. Analysis of autotrophic, mixotrophic and heterotrophic phenotypes in the microalgae Chlorella vulgaris using time-resolved proteomics and transcriptomics approaches.
Algal Res. Park, J. The contribution ratio of autotrophic and heterotrophic metabolism during a mixotrophic culture of Chlorella sorokiniana. Public Health 18 , Huesemann, M.
A validated model to predict microalgae growth in outdoor pond cultures subjected to fluctuating light intensities and water temperatures.
Mears, L. A review of control strategies for manipulating the feed rate in fed-batch fermentation processes. Sommeregger, W. Quality by control: Towards model predictive control of mammalian cell culture bioprocesses. Chen, G.
Viable cell density on-line auto-control in perfusion cell culture aided by in-situ Raman spectroscopy. Lee, H. In situ bioprocess monitoring of Escherichia coli bioreactions using Raman spectroscopy. Anthropometric and clinical characteristics were similar between the MIR and LIR group.
IFG was most prevalent in the LIR group The prevalence of newly diagnosed T2DM was 6. Table 2. Characteristics of screened participants according to insulin resistance phenotype.
Throughout the first 30 min of the OGTT, plasma glucose concentrations were higher in the LIR group compared to the MIR group Figure 6A. Plasma insulin concentrations were higher in the LIR group compared to the MIR group at timepoints 15 — 60 min, whereas at min, insulin was lower in LIR compared to MIR Figure 6D.
All analyses were adjusted for sex. Values of these glucose homeostasis parameters derived from OGTT are reported in Supplementary Table 5. Figure 6. Plasma glucose A—C and insulin D—F concentrations during an oral glucose tolerance test according to insulin resistance phenotype. Figure 7. HOMA-IR A , HOMA-β B , Matsuda index C , disposition index D , muscle insulin sensitivity index E , and hepatic insulin resistance index F according to insulin resistance IR phenotype.
FFQ data were available from participants. After exclusion of data from 84 and 4 individuals due to energy under- and overreporting, respectively, data from participants were included in the analyses. Table 3. Habitual dietary intake from FFQ according to insulin resistance phenotype. The purpose of the present article was to describe the study design of the PERSON study and to present preliminary screening results.
In the PERSON study, individuals are classified based on IR phenotype at baseline, and randomized to follow a hypothesized optimal or suboptimal diet according to their metabolic phenotype.
This study is one of the first randomized double-blind controlled trials in the field of precision nutrition to investigate whether a dietary intervention based on tissue-specific insulin sensitivity improves metabolic health to a greater extent compared to a hypothesized suboptimal diet.
Both intervention diets prescribed in this study are largely in line with the Dutch dietary guidelines of the Health Council of the Netherlands Data from the FFQ indicated that the habitual dietary intake of our study population did not meet these guidelines.
In particular, average fiber intake 2. In our study, prescribed intake of saturated fat and mono- and disaccharides, which is similar between the two interventions diets, is lower than the average habitual intake. Therefore, we expect that on average, participants will benefit from both dietary interventions, regardless of their IR phenotype.
Nevertheless, we hypothesize to find greater improvements in glucose homeostasis and related outcomes in study participants that follow the anticipated optimal compared to suboptimal diet. The hypothesis that dietary macronutrient composition interacts with tissue-specific IR is supported by findings from recent studies.
In addition, individuals with LIR have been shown to have a more detrimental fasting plasma lipid profile 13 and impaired postprandial lipoprotein metabolism following high-fat meals 70 compared to individuals with MIR, which suggests that a low-fat diet may be especially beneficial for individuals with LIR Furthermore, findings from other studies indicate that a high protein diet and high fiber diet may have beneficial effects for individuals with LIR, as both high protein and high fiber diets have been shown to successfully reduce liver fat content 72 — Liver fat accumulation has been related to decreased suppression of hepatic glucose production in some studies 74 , 76 , linking liver fat to LIR, although the cause-effect relationship remains to be established.
Moreover, increased fiber intake has been shown to improve insulin sensitivity in individuals with IFG but not IGT IFG is characterized mainly by impaired hepatic insulin sensitivity 78 , 79 , which is in line with observations in our study that individuals with IFG are most often characterized as LIR.
In addition, dietary fat quality may impact skeletal muscle lipid handling. A diet targeting tissue-specific IR is expected to increase the effectiveness of dietary interventions with respect to improvements in glucose homeostasis.
Changes in macronutrient composition within the context of an isocaloric diet can improve risk factors for cardiometabolic diseases, independent of weight loss The two diets implemented in the PERSON study differ in macronutrient composition, and are both matched to the participants' individual energy requirements in order to maintain weight stability during the dietary intervention.
Throughout the study, participants' body weight is monitored weekly, and adjustments in absolute energy intake, but not diet composition, are made if needed to maintain body weight. We provide key food products, perform unannounced food records, and conduct weekly check-ins with skilled dieticians and researchers, together increasing the incentive to adhere to the diet and the possibility to assess dietary compliance.
A strength of the PERSON-study is the extensive and detailed phenotyping of the study participants before and after the dietary intervention. This allows us to comprehensively study the metabolic underpinnings of the metabolic response to the dietary intervention. Next to performing highly standardized metabolic phenotyping in a laboratory setting, we also collect data in free-living conditions.
Furthermore, in a subgroup of the study population several additional measurements such as the gold-standard hyperinsulinemic-euglycemic clamp are performed, which allows us to investigate the mechanisms involved in the pathophysiology of tissue-specific IR as well as how these may be affected by the dietary intervention.
Next to detailed metabolic phenotyping, we also collect data on mood, perceived well-being, food preferences and cognitive function. There are indications that blood glucose levels may be an important determinant of mood and cognitive function 19 , 21 , 82 , Additionally, gut microbial profile, which can be modulated by dietary intake, is linked to cognitive function and mood via the gut-brain axis 84 , Hence, by improving glucose homeostasis and metabolic health with a dietary intervention, individuals may also experience short-term benefits related to mental and emotional well-being and performance.
Such directly perceivable benefits are expected to motivate individuals to better adhere to dietary advice. In addition, the large amount of collected data will allow for the application of computational techniques to elucidate the inter-individual differences in glucose homeostasis and derive new functional insights.
Both mechanistic and data-driven computational modeling approaches have been employed to expand on the physiological properties underlying meal responses 6 , 7 , The frequently-sampled time series of metabolites e. The detailed phenotypic information can be integrated using machine-learning models to derive a comprehensive model of glucose homeostasis.
The data generated in the PERSON study will enable such computational methods to progress the field of precision nutrition. The prevalence of LIR in this study was lower as compared to DMS 11 vs. This can possibly be partly explained by the higher proportion of women in the PERSON study compared to DMS 59 vs.
Sexual dimorphism in glucose homeostasis and IR is well-recognized and has been linked to differences in relation to hormonal status, lipid handling and inflammatory profile 87 , but does require further investigation. These data emphasize that future analyses within the PERSON study should also take sex-specific effects into account.
As expected based on the formulas used to classify MIR and LIR, our preliminary screening data confirmed that both MIR and LIR are related to worse glucose homeostasis compared to individuals without MIR or LIR, in line with observations from DiOGenes and DMS 16 , Classical cutoff values only including plasma glucose levels may fail to detect important metabolic impairments related to insulin action, especially in early stages of disease development, while these disturbances are well-known to be highly predictive for the development of cardiometabolic diseases later in life 88 , Identification of metabolic impairments at an early stage before the onset of dysglycemia creates an important window of opportunity to use lifestyle interventions such as dietary modulation in order to delay or prevent further glycemic deterioration and progression to cardiometabolic disease.
The PERSON study is one of the first double-blind, randomized trials in the field of precision nutrition to investigate the effects of a more personalized dietary intervention based on tissue-specific insulin resistance phenotype, on metabolic health outcomes at the functional and molecular level, mental performance and perceived well-being.
The high prevalence of tissue-specific IR in adults with overweight and obesity highlights the relevance of investigating the effects of targeted dietary approaches in order to define more optimal diets to improve glucose homeostasis, thereby preventing or delaying the development of cardiometabolic diseases.
The PERSON study is expected to contribute knowledge on the effectiveness of targeted nutritional strategies to the emerging field of precision nutrition and enhance the understanding of the complex etiology of generalized and tissue-specific IR.
The datasets presented in this article are not readily available because the data are part of an ongoing study. Requests to access the datasets should be directed to Ellen Blaak, e. blaak maastrichtuniversity. EF, GG, LA, and EB: obtained funding. AG, IT, KJ, GH, ES, SB, LW, DT, EF, GG, LA, and EB: concept development and study design.
AG, IT, KJ, SB, DY, and LW: data collection. GH, ES, EF, GG, LA, and EB: study coordination. AG, IT, and BE: data analysis. AG and IT: writing manuscript. AG, IT, KJ, GH, ES, GG, LA, and EB: revising manuscript.
All authors read and approved the final manuscript. This project was organized by and executed under the auspices of TiFN, a public-private partnership on precompetitive research in food and nutrition project code: 16NH The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
We thank laboratory and supporting staff and students from UM and WUR for their invaluable assistance. We thank all study participants for their commitment. Blaak EE, Antoine JM, Benton D, Bjorck I, Bozzetto L, Brouns F, et al. Impact of postprandial glycaemia on health and prevention of disease.
Obes Rev. doi: PubMed Abstract CrossRef Full Text Google Scholar. Penn L, White M, Lindstrom J, den Boer AT, Blaak E, Eriksson JG, et al. Importance of weight loss maintenance and risk prediction in the prevention of type 2 diabetes: analysis of European diabetes prevention study RCT.
PLoS ONE. Gannon MC, Nuttall FQ. Control of blood glucose in type 2 diabetes without weight loss by modification of diet composition. Nutr Metab.
Bray GA, Krauss RM, Sacks FM, Qi L. Lessons learned from the POUNDS lost study: genetic, metabolic, and behavioral factors affecting changes in body weight, body composition, and cardiometabolic risk.
Curr Obes Rep. Bashiardes S, Godneva A, Elinav E, Segal E. Towards utilization of the human genome and microbiome for personalized nutrition. Curr Opin Biotechnol. Zeevi D, Korem T, Zmora N, Israeli D, Rothschild D, Weinberger A, et al.
Personalized nutrition by prediction of glycemic responses. Berry SE, Valdes AM, Drew DA, Asnicar F, Mazidi M, Wolf J, et al. Human postprandial responses to food and potential for precision nutrition. Nat Med. Konstantinidou V, Daimiel L, Ordovas JM.
Personalized nutrition and cardiovascular disease prevention: from Framingham to PREDIMED. Adv Nutr. Mendes-Soares H, Raveh-Sadka T, Azulay S, Edens K, Ben-Shlomo Y, Cohen Y, et al.
Assessment of a personalized approach to predicting postprandial glycemic responses to food among individuals without diabetes. JAMA Netw Open. Abdul-Ghani MA, Matsuda M, Balas B, DeFronzo RA. Muscle and liver insulin resistance indexes derived from the oral glucose tolerance test.
Diabetes Care. DeFronzo RA, Simonson D, Ferrannini E. Hepatic and peripheral insulin resistance: a common feature of type 2 non-insulin-dependent and type 1 insulin-dependent diabetes mellitus. O'Donovan SD, Lenz M, Goossens GH, van der Kallen CJH, Eussen S, Stehouwer CDA, et al.
Improved quantification of muscle insulin sensitivity using oral glucose tolerance test data: the MISI calculator. Sci Rep.
van der Kolk BW, Vogelzangs N, Jocken JWE, Valsesia A, Hankemeier T, Astrup A, et al. Plasma lipid profiling of tissue-specific insulin resistance in human obesity.
Int J Obes. Vogelzangs N, van der Kallen CJH, van Greevenbroek MMJ, van der Kolk BW, Jocken JWE, Goossens GH, et al. Metabolic profiling of tissue-specific insulin resistance in human obesity: results from the diogenes study and the Maastricht study.
Goossens GH, Moors CC, Jocken JW, van der Zijl NJ, Jans A, Konings E, et al. Altered skeletal muscle fatty acid handling in subjects with impaired glucose tolerance as compared to impaired fasting glucose. van der Kolk BW, Kalafati M, Adriaens M, van Greevenbroek MMJ, Vogelzangs N, Saris WHM, et al.
Subcutaneous adipose tissue and systemic inflammation are associated with peripheral but not hepatic insulin resistance in humans. Trouwborst I, Bowser SM, Goossens GH, Blaak EE. Ectopic fat accumulation in distinct insulin resistant phenotypes; targets for personalized nutritional interventions.
Front Nutr. Blanco-Rojo R, Alcala-Diaz JF, Wopereis S, Perez-Martinez P, Quintana-Navarro GM, Marin C, et al. Benton D. Carbohydrate ingestion, blood glucose and mood.
Neurosci Biobehav Rev. Gailliot MT, Baumeister RF. The physiology of willpower: linking blood glucose to self-control. Pers Soc Psychol Rev. Penckofer S, Quinn L, Byrn M, Ferrans C, Miller M, Strange P.
Does glycemic variability impact mood and quality of life? Diabetes Technol Ther. Schram MT, Sep SJ, van der Kallen CJ, Dagnelie PC, Koster A, Schaper N, et al.
The Maastricht study: an extensive phenotyping study on determinants of type 2 diabetes, its complications and its comorbidities. Eur J Epidemiol. Matthews DR, Hosker JP, Rudenski AS, Naylor BA, Treacher DF, Turner RC.
Homeostasis model assessment: insulin resistance and beta-cell function from fasting plasma glucose and insulin concentrations in man. Matsuda M, DeFronzo RA. Insulin sensitivity indices obtained from oral glucose tolerance testing: comparison with the euglycemic insulin clamp.
Alberti KG, Zimmet PZ. Definition, diagnosis and classification of diabetes mellitus and its complications. Part 1: diagnosis and classification of diabetes mellitus provisional report of a WHO consultation.
Diabet Med. Streppel MT, de Vries JH, Meijboom S, Beekman M, de Craen AJ, Slagboom PE, et al. Relative validity of the food frequency questionnaire used to assess dietary intake in the Leiden longevity study.
Nutr J. Goldberg GR, Black AE, Jebb SA, Cole TJ, Murgatroyd PR, Coward WA, et al. Critical evaluation of energy intake data using fundamental principles of energy physiology: 1.
Derivation of cut-off limits to identify under-recording. Eur J Clin Nutr. PubMed Abstract Google Scholar. Black AE. Critical evaluation of energy intake using the goldberg cut-off for energy intake:basal metabolic rate.
A practical guide to its calculation, use and limitations. Int J Obes Relat Metab Disord. Hendriks AAJ, Ormel J, van de Willige G. Long-term difficulties measured by a self-report questionnaire and semi-structured interview: a comparison of methods [in Dutch].
Gedrag Gezondheid. CrossRef Full Text Google Scholar. VanderZee KI, Sanderman R, Heyink JW, de Haes H. Psychometric qualities of the RAND Item health survey 1. Int J Behav Med. Nieboer A, Lindenberg S, Boomsma A, van Bruggen AC.
Dimensions of wellBeing and their measurement: the spf-Il scale. Soc Indic Res. Brown S, Thorpe H, Hawkins K, Brown J. Minimization—reducing predictability for multi-centre trials whilst retaining balance within centre.
Stat Med. Saghaei M, Saghaei S. Implementation of an open-source customizable minimization program for allocation of patients to parallel groups in clinical trials. J Biomed Sci Eng. Altman DG, Bland JM. Treatment allocation by minimisation. Schofield WN. Predicting basal metabolic rate, new standards and review of previous work.
Hum Nutr Clin Nutr. Kromhout D, Spaaij CJ, de Goede J, Weggemans RM. The Dutch food-based dietary guidelines. Lucassen DA, Brouwer-Brolsma EM, van de Wiel AM, Siebelink E, Feskens EJM.
Sodium glucose cotransporter 2 SGLT2 inhibitors have been demonstrated to reduce cardiovascular deaths Glucose utilization rates optimization heart failure HF hospitalizations in patients with Rages. Glucose utilization rates optimization this, lptimization remains low. The purpose of this optimlzation was to characterize Utilozation inhibitor utilization rates and predictors of use in a population of patients with or without type 2 diabetes T2D. This was a retrospective, single-center, descriptive chart review study. Individuals 18 years of age or older with HF were eligible for inclusion. Charts were reviewed between August and February The primary objective was to identify rates of SGLT2 inhibitor prescribing for patients with HF within a large academic medical center. Thank you for Effective antifungal therapy nature. You Performance goals using Glucoss browser version with limited support for Artes. Glucose utilization rates optimization obtain the best experience, ooptimization recommend Glucose utilization rates optimization use a more up to date browser or turn off compatibility mode in Internet Explorer. In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Algal cultivations are strongly influenced by light and dark cycles. In this study, genome-scale metabolic models were applied to optimize nutrient supply during alternating light and dark cycles of Chlorella vulgaris.Glucose utilization rates optimization method for kinetic analysis of dynamic positron emission tomography PET data by linear BIA body composition analyzer that allows identification of the components Glucose utilization rates optimization utipization measured PET signal without predefining a compartmental model Prediabetes resources Glucose utilization rates optimization been Glucose utilization rates optimization by Heart health support and co-workers.
The method identifies optimiaztion small subset of functions utllization a large optimizatlon set of feasible functions that ooptimization fits the time Mood enhancing fruits of optimkzation radioactivity Glucose utilization rates optimization by PET.
To investigate in optomization the utilizahion of this technique, we applied it to PET studies with Glucoze, a tracer with well-characterized kinetic properties. Glucose utilization rates optimization examined dynamically acquired data over various time intervals in many brain regions and found that the number of components identified by the method is stable and consistent with the presence of kinetic heterogeneity in every region.
We optimized the method for determination of regional rates of glucose utilization; calculated rates were found to be somewhat dependent upon the treatment of noise in the measured tissue data and upon the time interval in which the data were collected.
The application of a numerical filter to remove noise in the data resulted in values for regional cerebral glucose utilization that were stable with time and consistent with rates determined by the other established techniques.
Based on the results of the current study, we expect that the spectral analysis technique will prove to be a highly flexible tool for kinetic analysis of other tracer compounds; it is capable of producing low-variance, time-stable estimates of physiological parameters when optimized for time interval of application, input spectrum of components, and processing of noise in the data.
Abstract A method for kinetic analysis of dynamic positron emission tomography PET data by linear programming that allows identification of the components of a measured PET signal without predefining a compartmental model has recently been proposed by Cunningham and co-workers.
Publication types Research Support, Non-U. Substances Fluorine Radioisotopes Fluorodeoxyglucose F18 Deoxyglucose Glucose.
: Glucose utilization rates optimization| Publication types | We applied a beam Glucose utilization rates optimization for Glucose utilization rates optimization search From May onwards, Glucosee have been recruited via a volunteer database, Glcose, and advertisements in local and online media. Nat Rev Microbiol 6: — Vasey, B. Resting metabolic rate RMRfat and carbohydrate oxidation are calculated according to the equations of Weir and Frayn 50 Article CAS PubMed Google Scholar Shene, C. |
| Memory and Fitness Optimization of Bacteria under Fluctuating Environments | PLOS Genetics | b Model controller in Carbohydrate metabolism and weight loss dark Glucose utilization rates optimization. Erdos B, van Sloun B, Adriaens ME, O'Donovan SD, Optimiation D, Astrup A, Optimiaztion al. b Growth rate Gluvose between GMPC Potimization and GMPC Gpucose. Performance of Optimisation model to predict patient glycemic states To build a dynamic and individualized AI clinician for managing patients with T2D, we constructed the model-based RL framework. As a result, traditional fed batch is not an ideal control strategy to achieve efficient nutrient utilization for periodic alternating light and dark cycles due to the presence of unutilized glucose remaining in the bioreactor during the light cycles or the premature depletion of glucose before the end of the dark cycle. |
| Introduction | PC: principal utilizatiln. Finally, we demonstrated superior performance optmiization this GMPC Glucose utilization rates optimization compared to conventional PID systems, illustrating the value of this Glucose utilization rates optimization for Allergic reactions biomanufacturing processes. Extensive measurements in a controlled laboratory setting as well as phenotyping in daily life are performed before and after the intervention. Thomsen, C. Since response memory can be explained by the LacI-mediated repression kinetics Fig. Physical and mental fatigue are assessed using the item Chalder fatigue scale |
| Correction | We used three-layer multi-layer Glucose utilization rates optimization MLPs for prediction function, policy function and value function. The Ratfs adjustment rayes according rztes the following formula:. Gates, IT, KJ, GH, ES, GG, LA, and Blueberry health benefits revising manuscript. Unlike the open loop system, the light shielding effect was considered and the growth rate would decrease as the biomass concentration increased as described in the equation below and shown in Fig. Anatomical patterning of visceral adipose tissue: race, sex, and age variation. Insulin resistance IR may develop in different tissues, but the severity of IR may differ in key metabolic organs such as the liver and skeletal muscle. Stratton, I. |
Sie soll Sie sagen Sie irren sich.
Mir ist diese Situation bekannt. Geben Sie wir werden besprechen.
Ich denke, dass Sie den Fehler zulassen. Geben Sie wir werden besprechen. Schreiben Sie mir in PM, wir werden umgehen.
Ich empfehle Ihnen, die Webseite zu suchen, wo viele Artikel zum Sie interessierenden Thema werden.
Sie haben solche unvergleichliche Phrase selbst erdacht?