Artificial Intelligence–Driven Norepinephrine Control Using Adaptive Cardiac Digital Twin for Critical Care Patients

Abstract

Background

Infusion pumps are used in critical care settings to deliver medicines such as norepinephrine (NE), insulin, and certain nutrients. Currently, clinicians preset the infusion rate in the device, usually for 1 hour, resulting in a constant drug delivery. However, a patient's response to a preset infusion can vary over time depending on their vital changes. In the case of NE, dosage errors are reported due to delayed infusion and interpatient variability in hemodynamics. To automate the infusion process, fuzzy logic, and proportional integral derivative controllers have been used, but they have fixed algorithms and lack clinical validation.

Objectives

This study proposes a data-driven artificial intelligence (AI)–digital twin adaptive framework for NE infusion to reduce the risks of hypotension and hypertension.

Methods

For personalized infusion, data were collected from patients supported with NE. A random forest decision model (RFDM) was used to predict the NE rates based on varying hemodynamics. The model performance at each time step is evaluated using a long short-term memory network that works as an adaptive digital twin cardiac model (DTCM). Virtual automation of the infusion process is performed using outputs of AI models.

Results

The RFDM achieved strong predictive accuracy, with a mean absolute percentage error (MAPE) of 5.3% and an R ² score of 0.97. MAPE values of DTCM for systolic, diastolic, and mean arterial blood pressure (BP) were 5.56, 5.79, and 5.36%, respectively. Compared with real-world infusion, the virtual automation reduced drug consumption by 13.82%.

Conclusion

The adaptive DTCM resembles real-world hemodynamics and has improved performance across various infusion rates. RFDM combined with DTCM provides an alternative method for NE infusion. The simulated evaluation shows that the proposed framework maintains BP, automates the infusion process, enabling patient-specific personalized dosing.

Introduction

Smart infusion pumps are medical devices used to deliver drugs in a controlled amount. However, the infusion rate in the device is preprogrammed for 1 hour, and the same quantity of drug is transferred to the patients irrespective of their current vital conditions. In intensive care units (ICUs), the instantaneous vital variations of the patients are not continuously monitored by the clinicians. During abnormal vital conditions, they manually adjust the infusion rate. However, there may be a delay in this process, as it is only intermittent. ^{1 2 3 4} The use of infusion pumps and associated clinical practices can lead to drug administration errors. ^{5 6} Among these errors, inotropic dosage errors are high due to the complex hemodynamics and intracohort variability in drug responses.

In this study, we use norepinephrine (NE), which is an inotrope used to increase the BP of hypotensive patients. It is administered during surgery time and also stays in the critical care unit to maintain the BP of patients at the target level. However, the target mean arterial pressure (MAP) of each patient varies based on their demographics and surgery. In the case of postsurgery patients, BP variations are too complex and unpredictable to handle. ^{7 8 9 10 11} To achieve and maintain stable hemodynamic conditions, NE is often administered automatically through syringe pumps.

To automate the infusion process, there are various closed-loop drug delivery systems developed for applications such as inotropic infusion, insulin administration, anesthesia, and cancer therapy. ^{12 13 14 15 16 17 18} The controllers, such as proportional integral derivative (PID), fuzzy logics (FLCs), and mathematical models, are used to compute drug dosages. However, due to inter- and intrapatient variability in hemodynamics, tuning these controllers for NE drug delivery remains challenging. ^{19 20 21} These control approaches show uncertainty while managing complex physiological systems and have limited capability to adapt rapidly to changing patient responses. ^{21 22 23 24}

Pharmacokinetic and pharmacodynamic (PK–PD) models are average patient models commonly used to analyze drug responses. These model-based approaches consist of certain assumptions about the human body and also have complex mathematical computations. In PK–PD modeling, the human body is designed as various compartments, with the assumption that drug distribution is uniform across all compartments. ^{25 26 27} Typically, two- to three-compartment models are used, which can lead to inaccuracies in predicting drug distribution and its physiological effects. Digital twin models for patient-specific modeling are an emerging area in precision medicine that adapts to changing medication rates. In this study, an artificial intelligence (AI)-enhanced infusion decision model with an adaptive cardiac digital twin is used to predict the NE infusion rate. The closed-loop virtual infusion control personalizes NE rates by continuously adapting to patient-specific and time-varying physiological responses.

The main objectives of this research work are:

• Design of an AI-enhanced closed-loop drug-vital system that can adapt to varying hemodynamics.

• Design of a data-driven random forest decision model (RFDM) NE infusion decision model that can capture intracohort variability in drug response.

• Design of adaptive digital twin cardiac model (DTCM) for analyzing the performance of each infusion decision from RFDM.

• Virtual automation of the infusion process using the AI model outputs.

• To generalize DTCM performance across various ranges of infusion rates.

• To personalize and maintain BP at the patient-specific target level in the virtual closed loop automation.

Methods

Our methodology starts with data collection from critical patients, those who were NE supported. Fig. 1 shows the architecture of the proposed closed-loop infusion system. BP parameters, such as systolic blood pressure (SBP), diastolic blood pressure (DBP), and MAP, were collected using a BP sensor attached to the patient. These BP values are recorded and stored on the bedside monitor. In current clinical practice, after diagnosis, doctors decide the target MAP of a patient based on the current blood pressure (BP) condition and adjust the infusion pump manually to set a suitable infusion rate. These infusion adjustments are recorded on the patient's chart to track medication history. These recorded infusion rates and BP data were used for training the AI decision models.

Fig. 1.

Architecture of closed-loop AI-enhanced NE infusion framework using RFDM and adaptive DTCM models. It consists of a time series data collection, which is shown as recorded BP and NE values, which are further used for the training and design of RFDM and DTCM. The outputs of these AI models trigger the control action for infusion, which is represented as a decision support system. AI, artificial intelligence; BP, blood pressure; ICU, intensive care unit; DTCM, digital twin cardiac model; NE, norepinephrine; RFDM, random forest decision model.

AI modeling consists of two parts: design of (1) RFDM and (2) adaptive DTCM. Fig. 2 shows the four key phases of the infusion control system. To design RFDM, we use the current BP parameters and NE rates as input for the model, and it predicts the next instant NE rate. The DTCM is used for clinical validation to forecast the drug response in the form of hemodynamic parameters. The decision support uses the outputs of both RFDM and DTCM to decide the final control action to transfer the predicted amount of drug to the patient. Closed-loop infusion process using the outputs of two AI models was performed to check the resemblance of the virtual infusion process with the real-world infusion. The DTCM performance was also evaluated in various ranges of infusion rates.

Fig. 2.

Workflow diagram of various phases of the AI-enhanced NE infusion process. It consists of four key phases: (1) data collection, where time series data of patients is collected, (2) RFDM design for NE prediction, (3) DTCM design for hemodynamics prediction, and finally, (4) virtual validation of AI models before clinical implementation. AI, artificial intelligence; DTCM, digital twin cardiac model; NE, norepinephrine; RFDM, random forest decision model.

Data-Driven Analysis

Data were collected from three ICUs of Amrita Hospital, India. It consists of 34 patients' details, containing their BP parameters and infusion rates. Fourteen cardiac patients, 12 neuro patients, and 8 gastrosurgery patients had abnormal hemodynamics. The time series data were collected at an interval of 5 minutes for a minimum of 1 hour and a maximum of 24 hours. Fig. 3 shows the duration of data collection for each patient, where P1, P2, …, P34 denote individual patients across the three ICUs. Patients above 40 years of age were considered and excluded children and those who were supported with multiple drugs. Due to the time required for drug distribution and metabolism, a visible physiological response to a given infusion rate typically occurs after 3 to 5 minutes. ²⁸ Accordingly, data were sampled at 5-minute intervals. The study included patients aged over 40 years, whereas pediatric patients and those receiving multiple concurrent drugs were excluded to reduce drug interaction effects. Time-series data were collected for each patient for a minimum of 1 hour and a maximum of 24 hours. For two patients, data were collected for only 1 hour, as NE infusion was administered for a limited duration and subsequently discontinued by the clinician.

Fig. 3.

Data collection duration of ICU patients supported with NE infusion. The time series data of hemodynamics and NE rate of patients were collected from three surgical ICUs. For each patient, the duration of data varies based on the drug response and risk level. ICU, intensive care unit; NE, norepinephrine.

Fig. 4 presents a sample of data from a cardiac patient, recorded at 5-minute intervals. The primary y -axis denotes MAP (mm Hg), and the secondary y -axis denotes the NE rate (mL/h). From 2:20 to 4:00 a.m., a constant infusion rate of 4 mL/h was given, and then the infusion rate was manually adjusted at 4:00 a.m. However, we observed that for every 5 minutes, the MAP showed a marked change, and hence, there was a potential to dynamically adjust these constant infusion rates. Among the collected data, we have observed both overdose (nine patients) and underdose (three patients) cases. The reported total dosage error is 35%. These kinds of dosage errors due to delays from clinicians are also highlighted in other studies. ^{5 29 30}

Fig. 4.

Graphical representation of the time series MAP and NE rate of a sample cardiac patient in the ICU. It shows the MAP variations of the patient for a constant pre-programmed infusion. This data-driven analysis highlights the dosage errors and hypertension that occurred due to manual and delayed infusion adjustment. ICU, intensive care unit; MAP, mean arterial pressure; NE, norepinephrine.

Design of Artificial Intelligence Models for Closed-Loop Norepinephrine Infusion Control

Random Forest Decision Model for Norepinephrine Rate Prediction

The closed-loop infusion process uses a data-driven control approach to predict NE rates. In this case, NE rates need to be decided quickly and more frequently based on the recent patient data. Various regression analyses were performed using the data obtained from 34 postoperative patients. Among the regression models, the RFDM model was used to forecast the NE rates. This tree-based model captures the behavior of more recent BP and corresponding infusion rates and performs well for the instantaneous NE forecasting task.

As the next instant, NE is highly dependent on the recent BP and NE, the current BP parameters (SBP, DBP, MAP), and NE rates were given as inputs to the model. During the training process, the K-fold cross-validation technique and optuna hyperparameter tuning were performed to improve the performance of the models. The K-fold cross-validation method was performed patient-wise to avoid data leakage between temporally correlated windows from the same patient. Within each fold, all samples from an individual patient were assigned exclusively to either the training or testing data, ensuring that no patient appeared in both datasets simultaneously. The RFDM hyper-tuning parameters, such as the number of estimators, maximum depth of the tree, minimum sample split, and minimum leaf split, were used during the training process. The K-fold cross-validation technique and optuna hyperparameter tuning generalize the model performance. Model performance was evaluated using parameters such as mean absolute percentage error (MAPE), coefficient of determination (R ² value), mean absolute error (MAE), standard deviation (SD) of the error, and root mean square error (RMSE).

Digital Twin Cardiac Model for Forecasting Drug Response

To validate the effect of the predicted infusion rates, an adaptive DTCM was designed to forecast the BP values. This shows the drug response of a patient to a given infusion. A long short-term memory (LSTM) network was used for this DTCM design, as the LSTM memory cell mechanisms handle the temporal dependencies and long short-term patterns in the BP. ^{31 32 33} Additionally, these networks handle nonlinear data, making them well-suited for multivariate time series forecasting.

The time series data contains BP parameters (SBP, DBP, MAP) and their corresponding NE infusion values. The current BP of a patient is the combined effect of the previous infusion rates and BP values. Hence, the previous time series data of infusion rate, SBP, DBP, and MAP values were given as inputs. DTCM was tuned by changing the window size of inputs, number of neurons, number of hidden layers, activation function, batch size, and epochs. As the data were collected at an interval of 5 minutes, a sliding window approach with a stride of one time instant (5 minutes) was employed for sequential forecasting of outputs. The model forecasted the next time instant SBP, DBP, and MAP values by using the four input features. This approach allows us to simulate and validate the predicted infusion rate's impact on a patient's BP, ensuring safe and effective dosing. Hyperparameter tuning and the K-fold cross-validation technique were performed to improve model performance. The model's performance was evaluated using MAPE, R ² value, MAE, SD of error, and RMSE.

Results

Performance Analysis of Artificial Intelligence Models for Norepinephrine Prediction

A suitable NE rate prediction at the required time is achieved by personalizing the dosage using the patient's recent BP and NE values. For the performance analysis of the infusion rate decision model, various regression models such as RFDM, XGBoost, decision tree, artificial neural networks, and support vector regression were used. Table 1 shows the performance analysis of these regression models. Among these models, RFDM showed high performance in predicting NE rates. During the RFDM training, the 5-fold cross-validation method and the optuna hyperparameter tuning were applied for each fold to get the best combination of tuning parameters. The average performance metrics of each fold were used to evaluate the performance of this model. Among the 5 folds, fold 1 showed higher performance with respect to statistical parameters, and this model was selected for further closed-loop automation experiments.

Table 1. Performance analysis of regression models for norepinephrine rate prediction.

AI models	MAPE	R 2 score	MAE	SD of error	RMSE
RFDM	5.39%	0.97	0.07	0.24	0.24
Decision Tree	7.36%	0.91	0.11	0.39	0.39
XGBoost	8.99%	0.96	0.12	0.26	0.26
ANN	9.05%	0.96	0.10	0.24	0.24
SVR	11.51%	0.95	0.13	0.29	0.29

Abbreviations: AI, artificial intelligence; MAE, mean absolute error; MAPE, mean absolute percentage error; RFDM, random forest decision model; RMSE, root mean square error; SD, standard error.

Performance Analysis of Adaptive Digital Twin Cardiac Model for Forecasting Blood Pressure

NE rate prediction from the RFDM model was validated using the DTCM, which forecasts BP (SBP, DBP, and MAP) values. The DTCM consists of an input layer, two LSTM layers, and an output layer. There were 32 neurons in the LSTM layers, and the softplus activation function was used in these layers. An Adam optimizer was used to train the model with mean square error as a loss function. During the training process, an early stopping regularization method was applied with a patience value of 10 to reduce overfitting. The model was tuned by using 200 epochs and a batch size of 30. Better performance was obtained by using an input window size of the previous 15 minutes. To generalize the model performance, K-fold cross-validation with 5 folds was used. Among the 5 folds, fold 3 showed higher performance, and this model was selected for further closed-loop automation experiments. Table 2 shows the average of performance metrics across all 5 folds. The average MAPE values across 5 folds for SBP, DBP, and MAP values are less than 10%, which shows high model performance. This trained DTCM can be further used for the validation of a new patient's predicted NE rate.

Table 2. Performance analysis of adaptive digital twin cardiac model for forecasting blood pressure values.

Performance metrics	SBP	DBP	MAP
MAPE (%)	5.56	5.79	5.36
MAE (mm Hg)	6.71	3.76	4.43
R 2 score	0.80	0.70	0.74
SD of error	10.27	6.77	6.77
RMSE	10.33	6.84	7.12

Abbreviations: DBP, diastolic blood pressure; MAE, mean absolute error; MAP, mean arterial pressure; MAPE, mean absolute percentage error; RMSE, root mean square error; SBP, systolic blood pressure; SD, standard error.

Comparison of Artificial Intelligence–Enhanced Infusion with Real-World Preprogrammed Infusion

Virtual automation of the infusion process is performed by using the RFDM and DTCM together in a closed-loop manner. We compared the real-world manual control of infusion with the AI-enhanced automated process. For this comparative analysis, the real-world BP and NE rate details of a test patient were collected for a continuous 10 hours (9 a.m. to 7 p.m.). Data recorded between 9:00 a.m. and 9:15 a.m. were used as the initial input window to start BP forecasting. Fig. 5 illustrates the closed-loop automation of AI models over a continuous 10-hour period, forecasting MAP and NE values from 9:20 a.m. to 7:20 p.m. In this figure, the MAP is plotted on the primary y -axis and NE rates on the secondary y -axis. The solid red line (MAP) and black line (NE) show the ground truth data of the test patient. The dotted red lines (forecasted MAP) and the black line (predicted noradrenaline) are the outputs of AI models. Forecasted MAP is denoted as (Pred_MAP), and the NE rate is denoted as (Pred_Norad). The target MAP range for the patient was maintained between 60 and 70 mm Hg.

Fig. 5.

Graphical representation of the comparison of AI-enhanced NE infusion control with preprogrammed infusion. It shows the real-world infusion process and the corresponding virtual automation to analyze the hemodynamic variations. This comparison highlights the role of AI-enhanced infusion to personalize and optimize drug consumption through adaptive digital twin control. AI, artificial intelligence; MAP, mean arterial pressure; NE, norepinephrine.

Initially, the patient was in a hypotensive condition with an MAP of 47 mm Hg. Hence, the RFDM model increased the NE rate and raised the MAP value to the target range of 60 to 70 mm Hg. After 12:00 p.m., the Pred_MAP value was slightly increased above 70 mm Hg, and then RFDM reduced the NE rate to avoid hypertension. This reduction in Pred_Norad avoided dosage errors and maintained the patient's condition in the normal MAP range. In this closed-loop approach, we predict the infusion rates frequently (every 5 minutes) based on the current patient condition and avoid overdose and underdose conditions.

However, in the real-world scenario, we have observed hypertension (after 12:25 p.m.) for a prolonged duration due to continuous preset infusion and delayed manual infusion adjustment. With the help of AI models, we personalized the NE rate by considering the current condition of the patient, and we also reduced total drug consumption. The quantity of drug delivered in real time was 22.79 mL, and the quantity delivered using AI models was 19.64 mL. Hence, the percentage of reduction in drug consumption from 9:20 a.m. to 7:20 p.m. is 13.82%. This shows the potential of fine-grained tuning of infusion rates while controlling the MAP.

Performance Analysis of Adaptive Digital Twin Cardiac Model for Various Ranges of Norepinephrine Rates

To generalize the performance of the DTCM for various ranges of NE rates, we selected two test patients having initial hypotension (case 1) and hypertension (case 2). For both cases, the initial 15 minutes of NE infusion and BP data were used as the starting point to forecast BP values at the next time steps. We continuously used the same NE rate for 1 hour to understand the BP trend for a constant infusion. After 1 hour, we slowly adjusted the infusion rate by a unit of 1 mL/h. Fig. 6 illustrates the variation of MAP (red line) corresponding to different ranges of NE rates (black line) for the case 1 test patient. The blue line shows the trend line of MAP variations. In this case, the patient was in a hypotensive condition with an MAP value of 73 mm Hg. The target MAP value for this patient is 90 to 100 mm Hg due to the vasodilated condition. The initial infusion rate of this patient was 2 mL/h. We analyzed MAP by varying the infusion rate from 2 to 6 mL/h and from 6 to 1 mL/h, as the AI models were trained using this range of infusion rates. The performance of the DTCM was evaluated for a continuous 9 hours from 8 a.m. to 5 p.m. In this figure, the MAP was increasing while increasing the NE rates, and it was decreasing while reducing the infusion rates. The MAP does not show a rapid decrease after 2:45 a.m. At this instant, the given infusion rate tries to maintain MAP, avoiding the sudden drop of MAP and the risk of hypotension.

Fig. 6.

Graphical illustration of the performance analysis of DTCM for a cardiac patient initially in hypotension. The infusion rate was adjusted manually at an interval of 1 hour to analyze the trend of MAP variations. It shows the adaptive DTCM property by varying the infusion rate from lower to higher and then reducing further to maintain the MAP level. DTCM, digital twin cardiac model; MAP, mean arterial pressure.

Fig. 7 shows the performance analysis of DTCM for the case 2 test patient. In this case, initially, the patient showed hypertension, having an MAP value of 109 mm Hg. The infusion rate of this patient was 6 mL/h. Then, we gradually reduced it from 6 to 1 mL/h and analyzed the MAP variations for reduced NE rates. The blue trend line shows that the MAP decreased with a decrease in infusion rates. Then, we increased the infusion rate slowly to raise MAP and maintain MAP in the target range of 60 to 70 mm Hg. The increase in infusion rates at this stage further maintained the MAP in the normal range and avoided the risk of hypotension. In both of the above two conditions, the MAP varies according to infusion changes, and it does not fall below or above a certain limit to maintain the MAP in the patient-specific target range.

Fig. 7.

Graphical illustration of the performance analysis of DTCM for a cardiac patient initially in hypertension. The infusion rate was adjusted manually at an interval of 1 hour to analyze the trend of MAP variations. It shows the adaptive DTCM property by varying the infusion rate from higher to lower and then increasing further to maintain the MAP level. DTCM, digital twin cardiac model; MAP, mean arterial pressure.

Discussion

The major challenge related to personalized medicine is the lack of proper controllers for adjusting the infusion rate on time. In the case of inotropic infusion, the control algorithms need to consider intracohort variability and work rapidly based on the instantaneous change in BP. In the case of NE, there are dosage errors reported due to preprogrammed infusion and delay from the clinician's side to adjust the infusion rate. The use of conventional controllers having fixed control algorithms does not perform well for time-variant hemodynamics. Advanced FLC, PID, and mathematical patient models were used in various drug delivery studies to personalize the infusion process. However, due to the unpredictable drug response and lack of adaptive control algorithms, this remains an unsolved clinical problem. The AI-enhanced infusion process has been tested only for insulin and anesthesia drug control.

The proposed AI-enhanced approach combining two AI models shows improved performance in the field of inotropic infusion. It maintained the MAP of critical care patients in their target range. The experimental analysis across various test conditions involving hypotension and hypertension shows that the adaptive DTCM effectively simulates the real patient response. By systematically varying the infusion rate from lower to higher and from higher to lower, BP trends in patients were analyzed. This closed-loop control approach automates the infusion process, thereby reducing the risk of hypotension and hypertension. Using the adaptive DTCM responses, the NE rate of test patients was optimized, preventing overdose and underdose cases. While further validation is needed to conduct a prospective study using more patient samples in various categories, this research represents a promising step toward personalized medicine in health care.

Tables 3 and 4 show the comparative performance analysis of AI-enhanced infusion with conventional methods. In both of these infusion decision and patient model analyses, we observed that the virtually simulated AI-enhanced results showed promising results in the field of precision medicine, automated drug control, and personalized infusion, which further shows the possibility of using these methods in real ICU settings.

Table 3. Comparison between conventional methods and random forest decision model for infusion rate prediction.

Performance metrics	PID and FLC	RFDM method
Drug consumption	Overdose or underdose cases 34	Controlled consumption (observed 5% reduction)
Dosage error duration	High (delay in infusion change) 35	Low (personalized drug control)
Personalized infusion	Low (same control logic) 21 34	High (patient-specific data)
Clinical performance	Manual infusion adjustments 21 34 36	Controlled dosing
Patient data requirement	Very low (initial requirement)	High (model training data)
Workload of clinicians	High (manual tuning) 34 35	Low (automated control)
Frequency of dosage change	Low (fixed design) 34 37	High (Automated control)

Abbreviations: FLC, fuzzy logic; PID, proportional integral derivative; RFDM, random forest decision model.

Table 4. Comparison between conventional patient models and adaptive digital twin cardiac model.

Performance metrics	Conventional patient models	Adaptive DTCM
MAE of MAP (mm Hg)	5.68 ± 0.93 37	4.43
RMSE of MAP (mm Hg)	7.46 ± 1.32 37	7.12
Personalization of patient model	Low (average model) 25 38 39	High (learns from data)
Resemblance to real patient	Low (assumptions-based study)	High (data-driven study)
Human body compartments	Single or multiple compartments 25 38	No compartments
Drug distribution	Uniform distribution (assumption) 25	No assumptions in this study
Adaptability to BP variations	Moderate 38 39 40	High (trained data)

Abbreviations: BP, blood pressure; DTCM, digital twin cardiac model; MAE, mean absolute error; MAP, mean arterial pressure; RMSE, root mean square error.

Clinical or Public Health Implications

We demonstrate the following potential benefits of utilizing AI models to develop inotropic infusion pumps that further help doctors and nurses in clinical settings.

• Automation and MAP control: the proposed framework automates the infusion process to maintain MAP in the target range.

• Reduced drug consumption: the closed-loop demonstration of AI models for continuous 10-hour observation shows that the automation could optimize total drug consumption.

• Personalized infusion rate prediction: the RFDM considers individual BP and infusion conditions for personalizing the next instant infusion rate.

• Validation of predicted infusion: the adaptive DTCM model simulates the drug response of an ICU patient by forecasting BP changes for a given infusion.

• Improved performance: compared with existing control methods, our approach potentially achieved high performance in NE prediction and BP control.

• Virtual validation of AI models: demonstration of two models on a test patient shows promising results and paves the way for real-world clinical implementation.

Limitations and Future Works

Even though we have achieved an improved performance in infusion rate prediction and BP control, the proposed system has some limitations when compared with the clinical settings. The following are some of the limitations and the future directions of this research:

• The real-world patient data revealed greater fluctuations in BP during NE infusion. In this study, we did not test our model with real-world data due to ethical issues. Moving forward, we will leverage complex digital twin models to refine the BP predictions.

• During the infusion time, when the syringe pump is empty, the entire system will become open-loop, leading to underdose conditions. Therefore, in further study, we will try to integrate an automatic refilling system to avoid such dosage errors.

• Drug–drug interactions significantly impact individual medication rates, necessitating simultaneous consideration of multiple variables such as the quantity, type, and current infusion rates of interacting drugs. Depending upon the vital changes, we need to switch between the medicines, and the infusion rates need to be predicted simultaneously. This is a major area of future research that we believe would emerge as a strong beneficiary of advancements in AI and digital twins.

Overall, this study shows a promising step for personalized medicine and potentially improves patient outcomes in critical care settings. However, more advanced cardiac digital twin techniques and control algorithms are required to handle more patients having complex hemodynamics, multiple inotropes, and drug interaction studies. Moving forward, our research will be focused on the design and development of multiple infusion pumps using advanced digital twins and infusion decision models.

Conclusion

This study aims to automate the NE infusion process using data-driven AI control algorithms. The existing control methods, such as PID, FLC, and PK–PD models, have limitations in adapting to patient drug responses. In this proposed closed-loop adaptive framework, the AI models showed acceptable performance to virtually demonstrate the NE infusion process. Considering the intracohort variability in drug response, the RFDM personalized NE infusion rate, and adaptive DTCM forecasted BP parameters, to validate the infusion rates. By the combined use of these two models, the real-world infusion process was virtually simulated, and the total drug consumption in both cases was determined. Furthermore, the DTCM model performance was generalized across various patient conditions having initial hypotension and hypertension. This automated data-driven control approach provides multiple advantages compared with existing methods by automating the infusion process, maintaining BP, and controlling drug consumption.

Funding Statement

Funding Information This study was funded by the Amrita Vishwa Vidyapeetham (grant ID: ASG2022026) and the Department of Science and Technology, Government of India (DST/INSPIRE Fellowship/2024/IF240119).

Ethical Approval

This study was performed in line with the principles of the Declaration of Helsinki. The approval for data collection was granted by the Ethics Committee of Amrita Institute of Medical Sciences and Research Center (11–4-2023/No: ECASM-AIMS-2023–190).

Informed Consent

Consent was obtained from all participants involved in the study.