ChronoSynthNet: a dual-task deep learning model development and validation study for predicting real-time norepinephrine dosage and the early detection of hypotension in patients with septic shock

Introduction

Sepsis, a critical health condition caused by an overwhelming immune response to infection, often progresses to septic shock, a state marked by profound hypotension and multiple organ failure. This complex (1) syndrome is particularly dangerous for patients in intensive care units (ICUs), where mortality rates exceed 40% (2).

Effective management of sepsis and septic shock relies on a comprehensive approach aimed at restoring perfusion, controlling infection, and supporting failing organs. A primary goal is to maintain a mean arterial pressure (MAP ≥65 mmHg) to ensure tissue oxygenation and prevent organ dysfunction (3). Initial resuscitation begins with the prompt administration of intravenous fluids to correct intravascular volume depletion; this fluid therapy serves as the cornerstone of early hemodynamic support. When hypotension persists despite adequate fluid resuscitation, norepinephrine is introduced to achieve and sustain target MAP. However, in patients with prolonged or severe septic shock, norepinephrine resistance may emerge, often requiring the escalation to second-line agents such as vasopressin to manage refractory hypotension (4). Beyond hemodynamic stabilization, optimal sepsis care also demands timely administration of broad-spectrum antibiotics, prompt source control (e.g., abscess drainage or device removal), and adjunctive organ support measures including mechanical ventilation or renal replacement therapy when clinically indicated.

Research into the management of septic shock highlights the importance of timely and appropriate vasopressor administration. Recent studies suggest that the initial choice of vasopressor and the timing of additional support significantly affect patient outcomes. For instance, clinical evidence indicates that an incremental rise in norepinephrine dosage up to 60 µg/min at the onset of vasopressin therapy is associated with a 20.7% increase in in-hospital mortality rates, whereas doses exceeding this threshold do not show this association (5). Furthermore, a retrospective study with Medical Information Mart for Intensive Care (MIMIC) data has provided insights into the optimal timing and dosage of vasopressors. These studies suggest that starting with lower doses of norepinephrine and introducing vasopressin early may help reduce the 28-day mortality rates in patients with septic shock (5,6).

The 2021 Surviving Sepsis Campaign guidelines provide a weak recommendation to initiate vasopressin (the second-line vasopressor) when the norepinephrine dose reaches 0.25–0.5 µg/kg/min (6). Nevertheless, the management of septic shock remains challenging, largely because of inter-patient variability in hemodynamic responses and the limited evidence supporting the concurrent use of multiple vasopressors (7).

This variability underscores the need for a personalized approach, which requires the consideration of specific patient factors such as baseline health status, the severity of sepsis, and dynamic clinical indicators (8). Furthermore, integrating advanced monitoring techniques and leveraging big data could enhance the precision of vasopressor therapies. This comprehensive approach aims to bridge the gap between current clinical practice and the evolving landscape of personalized medicine in the ICU. By focusing on the strategic use of vasopressors, integrating real-time data, and adapting to individual patient needs, healthcare providers can enhance the efficacy of sepsis management and improve outcomes for this high-risk population.

Artificial intelligence (AI) has shown increasing promise in supporting clinical decision-making for sepsis and septic shock, particularly in tasks such as early detection of hypotension and prediction of vasopressor requirements. Prior studies have explored classical machine learning methods—including logistic regression and random forests, as well as deep learning approaches like recurrent neural networks (RNNs) and long short-term memory (LSTM) networks to predict hypotension or hemodynamic instability from ICU data streams (9-11). However, many of these models are designed for single-task prediction (e.g., only risk stratification or only drug dosage recommendation), limiting their applicability in real-time ICU care. Moreover, traditional LSTM-based models often focus solely on temporal trends and may overlook complex cross-variable dependencies that are critical for clinical reasoning in dynamic conditions such as sepsis (12).

To address these gaps, we propose ChronoSythNet, a novel multitask deep learning model that incorporates Transformer multi-head attention (13) and LSTM (14) mechanisms with dynamic feature weighting. This model is designed to perform two critical tasks: predicting the temporal changes in individual norepinephrine administration levels and determining the optimal timing for initiating second-line vasopressors following norepinephrine use to address hypotension. These tasks share a common network structure, and features are learned using two separate loss functions. The contributions of this study are twofold: (I) first, we have developed a model that can continuously predict the levels of norepinephrine administered and optimize dosage control to effectively manage hypotension and customize treatment plans. Furthermore, the model assesses the risk of hypotension after norepinephrine administration, helping to inform the timing of second-line vasopressor deployment. (II) Our model provides feature importance ranking, enhancing the interpretability of the deep learning model in clinical settings. We present this article in accordance with the TRIPOD reporting checklist (available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-265/rc).

Methods

Database and patient cohort selection

This retrospective study was conducted using a publicly available online database, MIMIC IV v. 2.1 (15). The database consists of approximately 40,000 records of patients admitted to the ICU of the Beth Israel Deaconess Medical Center (Boston, MA, USA) from 2008 to 2019. Access to the database was granted by the organizer. The database contains a variety of data including vital signs, laboratory measurements, medications, and survival data. The inclusion criteria for this study were as follows: (I) adults (age ≥18 years old) admitted to the ICU with a diagnosis of sepsis based on Sepsis-3 criteria, (II) an ICU stay ≥6 hours, (III) count of vital signs >1, (IV) administration of norepinephrine therapy, and (V) initial norepinephrine administration after ICU admission ≥6 hours. The patient inclusion flowchart is presented in detail in Figure 1. Hypotension was defined as an MAP of <65 mmHg. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Figure 1 Flowchart of patient inclusion in the study. ICU, intensive care units; MIMIC, Medical Information Mart for Intensive Care.

Feature extraction and data processing

A total of 24 variables were initially selected based on their clinical relevance to septic shock and vasopressor titration. To ensure data quality, a multi-stage preprocessing pipeline was applied. Variables with more than 50% missingness were excluded, reducing the candidate set to 15. Physiological plausibility was enforced using prespecified bounds from clinical literature and expert consensus (Tables S1,S2); values outside these ranges were recoded as null, while patients with excessive outliers in key predictors were excluded. Redundancy was assessed using a correlation matrix (threshold >0.9), though no additional features were removed.

All predictors were temporally aligned to the initiation of norepinephrine administration (≥6 hours post-ICU admission). Hourly time-varying data were extracted prior to this reference point. For vital signs and ventilator settings, the most recent measurement within a 2-hour window was used, whereas laboratory tests were aligned using the most recent value within a 24-hour window. Measurements outside these windows were treated as missing, and missing values were handled according to expiration filter rules (Table S1). When noninvasive blood pressure was unavailable, invasive measurements were used for imputation.

This procedure ensured physiologic validity, standardized temporal alignment, and robust handling of missingness, yielding a high-quality cohort of 1,968 patients from the original 3,834 for model development and validation (Figure S1).

Multitask learning model

As shown in Figure 2, the model that called ChronoSynthNet was developed to support early clinical decisions by performing two tasks at the same time: predicting whether a patient is likely to develop low blood pressure (hypotension), and estimating how much norepinephrine they may need. To achieve this, the model uses a shared core that processes patient data in a way that benefits both predictions, improving efficiency and reliability (16). This approach involves a shared Transformer and LSTM encoder, enabling the extraction of latent features that are pivotal for both predictive tasks, thus enhancing the model’s efficiency and efficacy in clinical settings.

Figure 2 Overview of model development. LSTM, long short-term memory.

The model works by first handling any missing values using a masking technique to avoid gaps in the data (17). It then assigns weights to each clinical feature based on its importance before passing the data through a series of processing layers. These include a Transformer module and two long short-term memory (LSTM) layers, which help the model understand complex patterns over time. The final steps involve separate output layers, one predicts future norepinephrine dosage based on past trends, and the other assesses the likelihood of upcoming hypotension. The model continually adjusts itself during training to ensure both predictions are as accurate and dependable as possible.

Hybrid neural network architecture (the ChronoSynthNet multitask learning model)

The model combines two advanced technologies, which are Transformer and LSTM to better understand changes in patient data over time. This design helps the system recognize both long-term patterns and short-term changes, which are important for making accurate predictions in critical care settings.

• Transformer encoder: this part of the model looks at all time points in the patient’s data simultaneously, using a mechanism called “multi-head attention”. This enables the model to focus on different aspects of the clinical data at once. For example, it can detect how changes in lactate levels may relate to fluctuations in blood pressure or heart rate, even when these relationships are not immediately obvious.
• LSTM layer: after the Transformer, the data passes through an LSTM layer, which is especially good at capturing how patient conditions develop hour by hour. It helps refine the model’s understanding of trends in sequential data, such as whether a rising lactate trend over several hours is associated with worsening hemodynamic status.
• Feature weight network: a novel addition to this model is the feature-weighting mechanism, which dynamically assigns importance to different input features through a dedicated neural network. This adaptability allows the model to prioritize the most clinically relevant indicators during the learning process, thereby improving the precision of its predictions.

To train the model to handle both prediction tasks, which are norepinephrine rate estimation and hypotension risk detection, and this model uses two separate loss functions:

Norepinephrine rate forecasting loss (Lr): for the norepinephrine rate forecasting, a smooth L1 loss (Huber Loss) (18) is used, which is less sensitive to outliers than is the traditional mean squared error (MSE). This loss function helps provide a robust prediction of medication rates, which is crucial for maintaining patient stability.

Lr=1N∑i=1NSmooth∑L1(yi−y^l)[1]

Hypotension classification loss (Lc): the classification task involves binary cross-entropy loss (19) to effectively model the probability of hypotension occurrences. This loss function is ideal for binary outcomes and provides a probabilistic understanding of hypotension risks, which is vital for clinical decision-making.

Lc=1N∑i=1N[yilog(p^l)+(1−yi)log(1−p^l)][2]

By jointly optimizing these two loss functions, the model learns to provide both reliable rate suggestions and early warnings for hemodynamic instability, enhancing its clinical utility in critical care settings.

Model development

The ChronoSynthNet multitask learning model was meticulously designed for time-series forecasting in healthcare. The dataset is partitioned into 80% for training and 20% for testing, with strict separation to ensure no overlap, preserving the integrity and independence of the test data. Temporal windowing segments data into overlapping windows of size two with a 50% overlap, optimizing input structure for the Transformer and LSTM components to efficiently capture temporal patterns. Optimizer selection is critical, and adaptive optimizers such as adaptive moment estimation (Adam) (20) or root mean squared propagation (RMSprop) (21) are preferred due to their ability to automatically adjust the learning process based on how well the model is improving. The optimizers help the model learn more efficiently by fine-tuning how quickly it updates itself during training. A learning rate of 0.001 and a batch size of 32 were chosen to strike a balance between fast computation and reliable performance. Cross-validation employs a five-fold stratified approach within the training set, ensuring proportional representation of classes, which is vital for managing imbalanced medical datasets (22).

Statistics analysis

Patients’ characteristics and baseline characteristics including demographics, admission type of ICU, ICU unit type, Acute Physiology and Chronic Health Evaluation II (APACHEII) score, and Sequential Organ Failure Assessment (SOFA) score were compared between a nonhypotension group and a hypotension group. Nonparametric tests were applied since the population data did not have a normal distribution. The Kruskal-Wallis test (23) was used to test the significance of continuous variables expressed as medians and quartiles across both groups. The Fisher exact test (24) was used to test the categorical variables. The evaluation metrics, including area under the receiver operating characteristic curve (AUROC) for classification and MSE for regression, supplemented by precision-recall and calibration curves and other evaluation metrics, were employed to rigorously assess the model’s accuracy and reliability in predicting various medical outcomes.

Results

Of the 6,306 patients in the ICU that met theSepsis-3 criteria (25), 2,472 were excluded in our retrospective study. Among the patients, 8% (n=310) were considered to have hypotension and 92% (n=3,524 patients) to not have hypotension (Figure 1). In Table 1, the median age was similar across groups, at approximately 66 years, with no significant difference in age being noted (P=0.69). The gender distribution was nearly equal, although females were slightly more prevalent in the hypotension group. Both groups shared similar median SOFA scores of 4, while the median APACHE II score was higher in the hypotension group, a difference that was statistically significant (P=0.002). The duration of ICU stay was comparable between the groups (P=0.49). Admission types varied slightly, with a nonsignificant predominance of emergency admissions among patients with hypotension (P=0.098). Notably, there was a significant variation in the distribution of ICU types, particularly with a lower representation in the cardiovascular ICU and a higher representation in the medical ICU for patients with hypotension. Mortality rates differed significantly, with a lower rate observed in the hypotension group (28.1%) than in the nonhypotension group (39.2%).

Table S3 presents the critical variables that differed between the hypotension and nonhypotension group, including systolic blood pressure (98.3 vs. 107.5 mmHg; P<0.001), diastolic blood pressure (46.2 vs. 57.0 mmHg; P<0.001), and mean blood pressure (61.8 vs. 72.9 mmHg; P<0.001). Compared with the nonhypotension group, the hypotension group also had higher creatinine (2.0 vs. 1.6 mg/dL; P<0.001) and lactate levels (2.5 vs. 2.1 mmol/L; P<0.001), lower oxygen pressure (93.0 vs. 102.6 mmHg; P<0.001), and fewer platelets (135.0 vs. 151.6 ×10⁹/L; P=0.001). Heart rate and SpO2 were not significantly different, with P values of 0.63 and 0.026, respectively.

Model performance

The diagnostic capabilities of a binary classification model in the classification task were assessed via the AUROC and area under the precision-recall curve (AUPRC) across multiple validation folds and an internal testing dataset. The AUROCs (Figure 3A) for the five folds ranged from 0.93 to 0.89, indicating good effectiveness in minimizing false positives while maximizing true positives. The corresponding AUPRCs (Figure 3B) further confirmed the model’s precision across various recall levels, with values between 0.88 and 0.84, suggesting the model maintains high precision even under stringent recall thresholds. The internal validation, depicted in Figure 3C,3D, confirmed the model’s robust generalizability, with an internal AUROC of 0.89 and an AUPRC of 0.85. Table S4 summarizes performance metrics for a binary classification model across five training folds and an internal test. It highlights the model’s consistent ability to balance recall, precision, and specificity effectively. The recall values ranged from 0.74 to 0.81 across the five folds, indicating a consistent ability to identify relevant cases; meanwhile, the precision and specificity consistently hovered around 0.90 and 0.96 respectively, reflecting the model’s accuracy in predicting true positives and correctly identifying negatives. The internal testing set had a relatively lower recall (0.74), high precision (0.92), and high specificity (0.97).

Figure 3 Plots of model performance for classification tasks. (A) ROC curves across validation folds. (B) Precision-recall curves across validation folds. (C) Internal ROC curve. (D) Internal precision-recall curve. AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; ROC, receiver operating characteristic.

Table S4 also presents the comprehensive performance metrics for the rate regression task evaluated across five data folds (loss values in https://cdn.amegroups.cn/static/public/cdt-2025-265-1.xlsx). Fold 1 had an MSE of 0.0868 (loss of 0.0145), indicating solid predictive accuracy with low error. Fold 2 had slightly higher loss and MSE values of 0.0165 and 0.0280, respectively. Fold 3 had the lowest loss rate and MSE among all folds, with values of 0.01 and 0.1633, respectively. Fold 4 had a midrange performance, with a loss rate of 0.015 and an MSE of 0.0569. Finally, fold 5 maintained a consistent performance with a loss rate of 0.013, and the highest MSE of 0.0838. The internal testing yielded a low MSE (0.0213). This level of precision in the model’s performance can be observed in the rate comparisons of patient 30097987 with 30436278 (samples) as presented in Figure 4. For patient 30097987, the predicted rate begins notably lower than the actual rate but converges closer over time. For patient 30436278, the actual rate, represented by a blue line, remains steady throughout the observed period without notable fluctuations or signs of low blood pressure. In contrast, the predicted rate, illustrated in red, is consistently lower than the actual rate, but the decrease is not significant, and the predicted trend remains consistent (additional patient information is available in https://cdn.amegroups.cn/static/public/cdt-2025-265-2.xlsx).

Figure 4 Comparison of actual versus predicted rates for patient ICU stays. (A) Plot for patient 30097987. (B) Plot for patient 30436278. ICU, intensive care units.

Evaluation of potential clinical benefit and lead time

Feature importance rankings (Figure 5A and Table S5) were determined using the feature weight network layer. Lactate emerged as the most crucial variable, achieving an importance score of approximately 1. Platelet and noninvasive blood pressure measurements also exhibited significant importance, with scores of 0.91 for platelets, and 0.89 for both diastolic and systolic pressure. Creatinine and PaCO2, indicative of kidney function and respiratory health, respectively, exhibited notable importance, with scores of 0.77 and 0.73, respectively. The features of heart rate and potassium demonstrated lower predictive power, scoring 0.65 and 0.48, respectively. Chlorine, ionized calcium, and peripheral oxygen saturation (SpO2) had lower scores of 0.35, 0.35, and 0.28, respectively, indicating their supportive but less critical roles in this model. Tracheostomy, the least important feature, scored 0.17.

Figure 5 Evaluation of the potential clinical benefit and lead time plots. (A) Feature importance ranking. (B) Lead time of early versus delayed hypotension warnings: the x-axis represents individual patients in the internal testing cohort who experienced at least one hypotension event while on norepinephrine therapy. The y-axis shows the difference, in hours, between the model-predicted alert time and the actual recorded hypotension onset time. Positive values (green bars) indicate early warnings, meaning the model predicted the event in advance; negative values (red bars) indicate delayed warnings, where the alert occurred after the hypotension event had already started. “Delay” refers to alerts issued after the onset time, and “Early warning” refers to alerts given before onset. (C) Model-recommended norepinephrine rate adjustments: each point on the x-axis corresponds to an individual patient in the external testing cohort. The y-axis represents the standardized recommendation score output by the model for norepinephrine rate change. Points in red (“Need to reduce administration”) indicate the model’s recommendation to decrease the norepinephrine infusion rate for that patient; points in green (“Need to add administration”) indicate a recommendation to increase the infusion rate. More than half of the patients were advised to reduce their norepinephrine rate. And when norepinephrine is administered by continuous intravenous infusion, the infusion rate (e.g., µg/kg/min) at any given moment is essentially the current dosage being delivered to the patient. In this context, “rate” and “dosage” refer to the same quantity, expressed as the amount of drug infused per unit body weight per unit time. BP, blood pressure.

Figure 5B presents a bar plot comparing the early and delayed warnings for hypotension events: there were a total 398 patients in the internal testing data, 73 patients had hypotension events, and 47 demonstrated a majority of early warnings occurring just before the event time, with a median early warning leading time of 3.5 hours and an average lead time of 8.1 hours. Figure 5C presents a bar graph detailing the recommendations for norepinephrine’s rate adjustments, with red bars indicating a need to decrease administration and green bars suggesting an increase. More than half of the patients required a reduction in norepinephrine administration.

Discussion

The study aimed to improve the management of septic shock (26) in the ICU by developing a multitask deep learning model that could predict temporal changes in norepinephrine dosage and detect concurrent hypotension in patients receiving the norepinephrine. Using innovative Transformer multihead attention and LSTM mechanisms within a shared network structure, the model addresses critical timing and dosing challenges associated with vasopressor administration in septic shock (27). By generating targeted insights into norepinephrine rate fluctuations and hypotension onset, the model aids in timely decision-making for second-line vasopressor administration, optimizing treatment efficacy. Additionally, the interpretability of deep learning in clinical settings is enhanced via the inclusion of feature importance ranking, thus advancing personalized patient care and transparency in the management of this life-threatening condition (28).

For the task of prediction of norepinephrine rates, the MSE in the internal testing set was 0.0213, indicating a generally accurate model in terms of the average squared differences between the predicted and actual values. As demonstrated in Figures 4,5C, and table available at https://cdn.amegroups.cn/static/public/cdt-2025-265-2.xlsx, it may not be necessary to administer many doses to achieve therapeutic effects. Prolonged or high-dose administration of norepinephrine can lead to resistance, wherein patients exhibit reduced sensitivity to the drug, necessitating higher doses to achieve the equivalent therapeutic effect. The primary mechanisms behind this include receptor downregulation, neurotransmitter depletion (29), and metabolic pathway adaptation (30). The development of resistance can increase the complexity and risks of treatment, particularly in critically ill patients (31). To prevent resistance, medical teams should closely monitor patient responses and adjust treatment plans accordingly, considering alternative therapies or combination strategies when necessary (32). Therefore, precise dosing is essential. Our study results indicate that using multiphenotypic state predictions can determine the appropriate norepinephrine dosage before hypotensive events occur, maintaining vital signs effectively and reducing the likelihood of resistance.

For the early detection of hypotension events (the task of classification), when patients had already received norepinephrine, the model could effectively predict hypotension events at a median of 3.5 hours in advance (Figure 5B), providing valuable information for the timely addition of second-line vasopressors. The internal testing results underscore the model’s robust generalizability, as evidenced by an internal AUROC of 0.89 and an AUPRC of 0.85. These metrics confirm the model’s predictive accuracy and practical utility in critical clinical settings in which accurate class distinction is vital. However, the lower recall of 0.74 and higher precision of 0.92 indicate that while the model is adept at confirming true positives, it fails to detect 26% of actual positive cases (Table S4). This shortfall in recall may be worsened by an imbalanced dataset, with the underrepresentation of positive cases potentially leading to overly conservative predictions and thus some true positives being missed (33). Additionally, the high specificity of 0.97 demonstrates the model’s efficiency in identifying true negatives, but training on imbalanced datasets might have caused the model to predict more positives to enhance sensitivity, inadvertently increasing the rate of incorrect positive predictions. The calibration plot (Figure S2) indicates that the model tended to overestimate extreme values with positive cases at lower predictive probabilities (range 0.0–0.2).

This model provides a ranking of feature importance to enhance explanations of the deep-learning model, as shown in Figure 5A. In analyzing the variables prioritized by ChronoSynthNet, we observed that lactate, platelet count, and systolic/diastolic blood pressure (SBP/DBP) consistently emerged as the most important contributors to the model’s predictions. This ranking aligns with established pathophysiological mechanisms in septic shock. Elevated lactate is a well-recognized marker of tissue hypoperfusion and cellular oxygen deficit, reflecting the transition to anaerobic metabolism during circulatory failure. Persistently high lactate levels are associated with poor prognosis and are frequently used to guide resuscitation efforts (34). Thrombocytopenia, another top-ranked feature, often indicates consumptive coagulopathy, bone marrow suppression, or endothelial injury, all of which are hallmarks of sepsis-related organ dysfunction and systemic inflammation (35). Meanwhile, SBP and DBP are direct hemodynamic parameters central to the definition and clinical management of septic shock. Reductions in blood pressure reflect progressive vasoplegia, impaired vascular tone, and worsening perfusion, thereby signaling the need for vasopressor initiation (36).

Typically, norepinephrine administration raises blood pressure, reducing life-threatening risks associated with hypotension (37). However, in some patients admitted to the ICU, routine norepinephrine administration fails to effectively increase blood pressure (38), possibly due to drug sensitivity, dosage (39), shock, multiple organ dysfunction (40), fluid deficits, electrolyte disturbances, acid-base imbalances, severe inflammation or infections, and nutritional status. We found that patients with hypotension resistant to norepinephrine often experienced inadequate vascular responsiveness due to multiple organ dysfunctions including compromised pulmonary, cardiac, renal, and hepatic functions. These dysfunctions manifest as imbalances in oxygen supply and demand, cardiovascular and respiratory disturbances, electrolyte and acid-base imbalances, and insufficient blood volume, which collectively reduce vascular responsiveness and diminish norepinephrine’s hypertensive effect. Persistent hypoxia in some patients, despite continuous oxygen supply, may relate to respiratory diseases (41,42), heart failure (43,44), severe anemia (45,46), or renal dysfunction (47), further weakening norepinephrine’s efficacy (48,49). Additionally, infections and inflammation can lead to vasodilation and impaired myocardial reactivity (50) while renal dysfunction or electrolyte disorders may increase potassium and chloride levels, adversely affecting cardiac function (51-53). The selected variables, including those such as heart rate, respiration rate, and lactate, are rational indicators in predicting future hypotensive events by reflecting the underlying multi-organ dysfunctions and systemic imbalances, thereby validating the model’s predictive power in such critical scenarios.

Limitations

The primary limitation of this study lies in its single-center design, which may reduce the generalizability of findings. Although ChronoSynthNet demonstrated promising performance, its current scope is restricted to vital signs, laboratory tests, and ventilation-related variables. Incorporating additional parameters such as cardiac output, stroke volume, imaging modalities, and medication data could further improve predictive accuracy (54). Missing data, despite preprocessing and imputation, may have biased results by underrepresenting certain clinical patterns, and class imbalance may have affected sensitivity and specificity despite mitigation measures. Future iterations should integrate automated quality-control pipelines, including physiological plausibility checks, temporal consistency screening, and cross-variable validation to manage low-quality inputs and enhance robustness in real-world settings.

At present, the model remains in the development and validation stage. Prospective, multi-center studies will be essential to confirm its generalizability and reliability across diverse healthcare environments. Federated learning approaches may further improve robustness and reduce center-specific bias (55). Clinical implementation will additionally require regulatory approval as a medical device, followed by integration into hospital information systems for real-time data processing and prediction. In practice, clinicians would interact with the model’s outputs, comprising real-time risk estimates, confidence scores, and key contributing features, while interpretation and clinical decision-making would remain under the responsibility of critical care professionals. These steps are necessary to ensure safe, effective, and scalable deployment of ChronoSynthNet in routine practice.

Conclusions

This study confirmed the ability of ChronoSynthNet multitask learning model to support precision medicine for managing septic shock. The advanced deep learning framework excelled at two critical tasks: accurately predicting norepinephrine administration rates and monitoring the onset of hypotension post administration. Additionally, its feature importance ranking enhances model interpretability, facilitating personalized treatment decisions. These capabilities enable timely, precise interventions, such as the administration of second-line vasopressors, demonstrating the model’s potential to improve individualized critical care.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-265/rc

Peer Review File: Available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-265/prf

Funding: This study was partly supported by a grant from the National Natural Science Foundation of China (No. 32100510 to Z.H.), the Shanghai Rising-Star Program (No. 21QB1400900 to Z.H.), the CAMS Innovation Fund for Medical Science (No. 2019-I2M-5-066) and the Shanghai Municipal Science and Technology Major Project (Nos. 2023SHZDZX02 and 2017SHZDZX01).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-265/coif). Z.J. and Y.Y. work for Philips Research China, a research department for a manufacturer of healthcare devices and an informatics solutions provider. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Dugar S, Choudhary C, Duggal A. Sepsis and septic shock: Guideline-based management. Cleve Clin J Med 2020;87:53-64. [Crossref] [PubMed]
2. Bauer M, Gerlach H, Vogelmann T, et al. Mortality in sepsis and septic shock in Europe, North America and Australia between 2009 and 2019- results from a systematic review and meta-analysis. Crit Care 2020;24:239. [Crossref] [PubMed]
3. Carlos Sanchez E, Pinsky MR, Sinha S, et al. Fluids and Early Vasopressors in the Management of Septic Shock: Do We Have the Right Answers Yet? J Crit Care Med (Targu Mures) 2023;9:138-47. [Crossref] [PubMed]
4. Shankar A, Gurumurthy G, Sridharan L, et al. A Clinical Update on Vasoactive Medication in the Management of Cardiogenic Shock. Clin Med Insights Cardiol 2022;16:11795468221075064. [Crossref] [PubMed]
5. Zaaqoq AM, Barnett AG, Griffee MJ, et al. Beneficial Effect of Prone Positioning During Venovenous Extracorporeal Membrane Oxygenation for Coronavirus Disease 2019. Crit Care Med 2022;50:275-85. [Crossref] [PubMed]
6. Xu J, Cai H, Zheng X. Timing of vasopressin initiation and mortality in patients with septic shock: analysis of the MIMIC-III and MIMIC-IV databases. BMC Infect Dis 2023;23:199. [Crossref] [PubMed]
7. Santacroce E, D'Angerio M, Ciobanu AL, et al. Advances and Challenges in Sepsis Management: Modern Tools and Future Directions. Cells 2024;13:439. [Crossref] [PubMed]
8. Iskander KN, Osuchowski MF, Stearns-Kurosawa DJ, et al. Sepsis: multiple abnormalities, heterogeneous responses, and evolving understanding. Physiol Rev 2013;93:1247-88. [Crossref] [PubMed]
9. Mao Q, Jay M, Hoffman JL, et al. Multicentre validation of a sepsis prediction algorithm using only vital sign data in the emergency department, general ward and ICU. BMJ Open 2018;8:e017833. [Crossref] [PubMed]
10. Shashikumar SP, Stanley MD, Sadiq I, et al. Early sepsis detection in critical care patients using multiscale blood pressure and heart rate dynamics. J Electrocardiol 2017;50:739-43. [Crossref] [PubMed]
11. Moor M, Rieck B, Horn M, et al. Early Prediction of Sepsis in the ICU Using Machine Learning: A Systematic Review. Front Med (Lausanne) 2021;8:607952. [Crossref] [PubMed]
12. Wu S, Xiao Y, Ma T, et al. Transformer-based time-series modeling for sepsis prediction in intensive care units. IEEE J Biomed Health Inform 2023;27:1394-404.
13. Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need. In: Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS'17). Curran Associates Inc., Red Hook, NY, USA, 2017:6000-10.
14. Sherstinsky A. Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network. Physica D: Nonlinear Phenomena 2020;404:132306.
15. Johnson AEW, Bulgarelli L, Shen L, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data 2023;10:1. [Crossref] [PubMed]
16. Zhang Y, Yang Q. A Survey on Multi-Task Learning. IEEE Transactions on Knowledge and Data Engineering 2022;34:5586-609. [Crossref] [PubMed]
17. Zhang Z, Zhang Y, Zeng A, et al. Time-Series Data Imputation via Realistic Masking-Guided Tri-Attention Bi-GRU. Frontiers in Artificial Intelligence and Applications 2023;372:3074-82.
18. Barron JT. A general and adaptive robust loss function. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, USA 2019:4331-9.
19. Ho Y, Wookey S. The real-world-weight cross-entropy loss function: Modeling the costs of mislabeling. IEEE Access 2019;8:4806-13.
20. Zeng K, Liu J, Jiang Z, et al. A decreasing scaling transition scheme from Adam to SGD. Adv Theory Simul 2021;4:2100599.
21. Zaheer M, Reddi SJ, Sachan D, et al. Adaptive methods for nonconvex optimization. In: Proceedings of the 32nd International Conference on Neural Information Processing Systems (NIPS'18). Curran Associates Inc., Red Hook, NY, USA, 2018:9815-25.
22. Szeghalmy S, Fazekas A. A Comparative Study of the Use of Stratified Cross-Validation and Distribution-Balanced Stratified Cross-Validation in Imbalanced Learning. Sensors (Basel) 2023;23:2333. [Crossref] [PubMed]
23. Ostertagova E, Ostertag O, Kováč J. Methodology and application of the Kruskal-Wallis test. Applied Mechanics and Materials 2014;611:115-20.
24. Kim HY. Statistical notes for clinical researchers: Chi-squared test and Fisher's exact test. Restor Dent Endod 2017;42:152-5. [Crossref] [PubMed]
25. Singer M, Deutschman CS, Seymour CW, et al. The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA 2016;315:801-10. [Crossref] [PubMed]
26. Evans L, Rhodes A, Alhazzani W, et al. Surviving Sepsis Campaign: International Guidelines for Management of Sepsis and Septic Shock 2021. Crit Care Med 2021;49:e1063-143. [Crossref] [PubMed]
27. Shi R, Hamzaoui O, De Vita N, et al. Vasopressors in septic shock: which, when, and how much? Ann Transl Med 2020;8:794. [Crossref] [PubMed]
28. Teng Q, Liu Z, Song Y, et al. A survey on the interpretability of deep learning in medical diagnosis. Multimed Syst 2022;28:2335-55. [Crossref] [PubMed]
29. Mohanty PK, Sowers JR, Thames MD, et al. Myocardial norepinephrine, epinephrine and dopamine concentrations after cardiac autotransplantation in dogs. J Am Coll Cardiol 1986;7:419-24. [Crossref] [PubMed]
30. Kubista H, Boehm S. Molecular mechanisms underlying the modulation of exocytotic noradrenaline release via presynaptic receptors. Pharmacol Ther 2006;112:213-42. [Crossref] [PubMed]
31. Chinemerem Nwobodo D, Ugwu MC, Oliseloke Anie C, et al. Antibiotic resistance: The challenges and some emerging strategies for tackling a global menace. J Clin Lab Anal 2022;36:e24655. [Crossref] [PubMed]
32. Baryakova TH, Pogostin BH, Langer R, et al. Overcoming barriers to patient adherence: the case for developing innovative drug delivery systems. Nat Rev Drug Discov 2023;22:387-409. [Crossref] [PubMed]
33. Luque A, Carrasco A, Martín A, et al. The impact of class imbalance in classification performance metrics based on the binary confusion matrix. Pattern Recognition 2019;91:216-31.
34. Vincent JL, Bakker J. Blood lactate levels in sepsis: in 8 questions. Curr Opin Crit Care 2021;27:298-302. [Crossref] [PubMed]
35. Iba T, Levy JH, Wada H, et al. Differential diagnoses for sepsis-induced disseminated intravascular coagulation: communication from the SSC of the ISTH. J Thromb Haemost 2019;17:415-9. [Crossref] [PubMed]
36. VON EULER US. Noradrenaline in hypotensive states and shock; physiological aspects. Lancet 1955;269:151-4. [Crossref] [PubMed]
37. Cecconi M, Evans L, Levy M, et al. Sepsis and septic shock. Lancet 2018;392:75-87. [Crossref] [PubMed]
38. Russell JA. Vasopressor therapy in critically ill patients with shock. Intensive Care Med 2019;45:1503-17. [Crossref] [PubMed]
39. Segal DS, Geyer MA, Weiner BE. Strain differences during intraventricular infusion of norepinephrine: possible role of receptor sensitivity. Science 1975;189:301-3. [Crossref] [PubMed]
40. Annane D, Ouanes-Besbes L, de Backer D, et al. A global perspective on vasoactive agents in shock. Intensive Care Med 2018;44:833-46. [Crossref] [PubMed]
41. Cortés I, Peñuelas O, Esteban A. Acute respiratory distress syndrome: evaluation and management. Minerva Anestesiol 2012;78:343-57.
42. Gorman EA, O'Kane CM, McAuley DF. Acute respiratory distress syndrome in adults: diagnosis, outcomes, long-term sequelae, and management. Lancet 2022;400:1157-70. [Crossref] [PubMed]
43. Köhler D, Voshaar T, Stais P, et al. Hypoxic, anemic and cardiac hypoxemia: When does tissue hypoxia begin? Dtsch Med Wochenschr 2023;148:475-82. [Crossref] [PubMed]
44. Alviar CL, Miller PE, McAreavey D, et al. Positive Pressure Ventilation in the Cardiac Intensive Care Unit. J Am Coll Cardiol 2018;72:1532-53. [Crossref] [PubMed]
45. Jelkmann I, Jelkmann W. Impact of erythropoietin on intensive care unit patients. Transfus Med Hemother 2013;40:310-8. [Crossref] [PubMed]
46. Ouellette DR. The impact of anemia in patients with respiratory failure. Chest 2005;128:576S-82S. [Crossref] [PubMed]
47. Caldwell HG, Howe CA, Chalifoux CJ, et al. Arterial carbon dioxide and bicarbonate rather than pH regulate cerebral blood flow in the setting of acute experimental metabolic alkalosis. J Physiol 2021;599:1439-57. [Crossref] [PubMed]
48. Alshahrani S, Rapoport RM, Soleimani M. Vascular contractile reactivity in hypotension due to reduced renal reabsorption of Na(+) and restricted dietary Na Naunyn Schmiedebergs Arch Pharmacol 2017;390:321-6. [Crossref] [PubMed]
49. Yamazaki T, Ushikoshi-Nakayama R, Shakya S, et al. The effects of bathing in neutral bicarbonate ion water. Sci Rep 2021;11:21789. [Crossref] [PubMed]
50. Paulus WJ, Tschöpe C. A novel paradigm for heart failure with preserved ejection fraction: comorbidities drive myocardial dysfunction and remodeling through coronary microvascular endothelial inflammation. J Am Coll Cardiol 2013;62:263-71. [Crossref] [PubMed]
51. Ferreira JP, Butler J, Rossignol P, et al. Abnormalities of Potassium in Heart Failure: JACC State-of-the-Art Review. J Am Coll Cardiol 2020;75:2836-50. [Crossref] [PubMed]
52. Hunter RW, Bailey MA. Hyperkalemia: pathophysiology, risk factors and consequences. Nephrol Dial Transplant 2019;34:iii2-iii11. [Crossref] [PubMed]
53. Littmann L, Gibbs MA. Electrocardiographic manifestations of severe hyperkalemia. J Electrocardiol 2018;51:814-7. [Crossref] [PubMed]
54. Dung-Hung C, Cong T, Zeyu J, et al. External validation of a machine learning model to predict hemodynamic instability in intensive care unit. Crit Care 2022;26:215. [Crossref] [PubMed]
55. Bharati S, Mondal M, Podder P, et al. Federated learning: Applications, challenges and future directions. International Journal of Hybrid Intelligent Systems 2022;18:19-35.