Edge computing with federated learning for early detection of citric acid overdose and adjustment of regional citrate anticoagulation

Abstract

Regional citrate anticoagulation (RCA) is critical for extracorporeal anticoagulation in continuous renal replacement therapy done at the bedside. To make patients’ data more secure and to help with computer-based monitoring of dosages, we suggest a system that uses machine learning. This system will give early alerts about citric acid overdose and advise changes to how much citrate and calcium gluconate are infused into the patient’s body. Citric acid overdose causes significant clinical risks, emphasizing the need for better adaptable anticoagulation procedures that can respond quickly. The study puts forward a new structure that uses edge computing and federated learning to make better citrate anticoagulation procedures. We proposed the resource-aware Federated Learning with Dynamic Client Selection (RAFL-Fed) algorithm in our method. In this setup, every client takes part by training a local model locally and then sending its outcome to a main server. The algorithm chooses clients for each training session depending on their computing resources, which keeps things efficient and scalable. The server collects the client inputs using weighted averages to update the global model. This step is performed repeatedly across many communication cycles, letting the system adjust to changing data trends from different locations. We put RAFL-Fed to the test on the MIMIC-IV dataset, and it outperformed other methods, getting a high accuracy of 0.9615 (IID) and 0.9571 (Non-IID), also with the lowest loss values being 0.2625 and 0.2469 in that order. It also noted the best MAE at 0.1731 (Non-IID) and a bit higher at 0.2081 (IID). Along with the high sensitivity at 0.9968, specificity stood strong as well, measuring 0.9449, plus latency was only 0.123s, which shows how effective it is for early detection of citric acid overdose as well as adjusting in real-time in the regional citrate anticoagulation process. The proposed method shows a promising solution for the real-time monitoring and adjustment of citrate anticoagulation regimens, greatly enhancing patient data security and treatment effectiveness in clinical settings. This method signifies a significant advancement in handling anticoagulation therapy.

Introduction

Edge computing allows data processing and analysis closer to the point of care, which decreases delay time and enhances response rate [1]. Federated learning is a machine learning technique that enables multiple institutions or devices (such as bedside monitors or ICU systems) to collaboratively train a shared predictive model, without exchanging raw patient data. Rather than transmitting confidential information to a main server, every device processes its data and only shares updates for the model (like gradients or weights). These are then gathered together to refine the global model. This method keeps the privacy of data, decreases communication burden, and boosts adherence with regulatory norms. It is particularly fitting for healthcare settings where the sensitivity of data is crucial [2]. By merging edge computing with federated learning in our study, we have created a real-time early warning system that preserves privacy. It can detect citrate overdose and modify RCA dosing during CRRT procedures.

The growing adoption of continuous renal replacement therapy (CRRT) in critically ill patients underscores the critical need for safe, effective, and sustainable anticoagulation strategies in intensive care units (ICUs). Critically ill populations often present with multi-organ dysfunction, surgical trauma, active bleeding, or disseminated intravascular coagulation (DIC), complicating anticoagulant selection. While systemic heparin remains a cost-effective option with established efficacy, its use is increasingly limited by risks such as heparin-induced thrombocytopenia (HIT), bleeding complications, and delayed monitoring of anticoagulation levels, which can compromise filter longevity and escalate treatment costs [3,4,5,6,7]. Regional citrate anticoagulation (RCA) has emerged as the preferred anticoagulation method for CRRT, supported by robust evidence demonstrating superior safety and efficacy. Unlike heparin, RCA minimizes systemic anticoagulation by selectively chelating ionized calcium within the extracorporeal circuit, thereby reducing bleeding risks and enhancing filter biocompatibility [8]. Recent randomized controlled trials (RCTs) confirm that RCA prolongs circuit lifespan, lowers transfusion requirements, and improves hemodynamic stability in patients with bleeding diatheses or hepatic dysfunction [9, 10]. These advantages have solidified RCAs position as the first-line anticoagulation strategy in CRRT, as endorsed by the Kidney Disease Improving Global Outcomes (KDIGO) guidelines [11]. Too much use of citric acid is a serious problem in taking care of patients who are going through continuous treatment for kidney problems (CRRT) in intensive care units (ICUs) [12, 13]. CRRT often helps to treat Acute kidney injury (AKI) and other renal diseases, often using citric acid as an anticoagulant during the procedure [14]. Nevertheless, if not correctly measured or tracked, there can be a buildup of this same citric acid, leading to metabolic irregularities such as alkalosis and hypocalcemia - all very dangerous for the patient’s health [15].

In this situation, it is very important to have early alert systems for noticing an overdose of citric acid and making quick changes to the dosage of regional citrate anticoagulation (RCA) used in severe care settings [16]. At present, strategies for observing the levels of citrate and handling RCA dosing often depend on manual actions and periodic checks, which might not be sensitive or fast enough to prevent adverse incidents [17]. In addition, managing data at a central level and training models can cause problems regarding privacy of information, scalability potential, as well as real-time reaction abilities [18].

To overcome these restrictions, our study suggests the use of edge computing and federated learning methods for creating an early warning system that detects citric acid overdose and adjusts RCA in CRRT procedures while preserving privacy. Edge computing allows data processing and analysis closer to the point of care, which decreases delay time and enhances response rate [19, 20]. Federated learning provides collaborative model training across distributed edge devices without compromising data security. This helps create precise and scalable predictive models [21,22,23].

The main goal of this study is to create and assess an early warning mechanism for detecting citric acid overdose and adjusting RCA using edge computing coupled with federated learning. The specific targets incorporate:

1. 1.

To develop privacy-preserving-based federated learning methods for training forecast models to detect anomalies in the data and predict the adjustment value based on the constant observation of patients’ information, including important vital signs to their health, results from laboratory tests, and CRRT measurements.

2. 2.

The proposed federated learning-based method implements a hybrid approach of Bidirectional LSTM and GRUs in a distributed edge computing platform. The model possesses the ability to perform classification and regression tasks for detection and prediction.

3. 3.

To assess the performance of RAFL-Fed with two scenarios of IID and non-IID MIMIC IV dataset. The experimental results show the superior effectiveness of the proposed approach compared to the state-of-the-art.

Related works

This section provides the background and inspiration for the suggested framework by examining the current literature in the areas of edge computing, federated learning, regional citrate anticoagulation (RCA) management, and citric acid overdose detection in healthcare. McMahan et al. (2017) [24] have investigated effective communication learning of deep networks by means of federated learning frameworks. Their research has shown how distributed data systems can train machine learning models efficiently without putting at risk the privacy of data. But the research didn’t include a medical context, which confines its immediate use to clinical environments like monitoring citrate. Brisimi et al. (2018) [25] connect federated learning and healthcare by creating prediction models for real-time electrolyte monitoring in ICU patients. The research highlights the potential of federated learning to manage sensitive patient information while keeping their privacy intact. Nonetheless, it was limited by a small sample size of 50 ICU patients, which restricted its broad applicability.

The comparative analysis of the two approaches is summarized in Table 1, highlighting their respective methodologies, application domains, key contributions, performance gains, and limitations.

[IMAGE OMITTED: SEE PDF]

Wang et al. (2022) [28] have pushed forward the use of federated learning by proposing frameworks that preserve privacy for electrolyte monitoring. Their research underlined how critical it is to protect private medical information. But the study’s single-center data source limited its scope, posing challenges for wider clinical implementation. Sahu et al. (2020) [29] put forward the FedProx optimization algorithm for federated learning in heterogeneous networks. This work makes an important contribution to tackling difficulties in distributed medical settings. However, tuning the proximal term (\(\mu\)) was found to be difficult. Besides, this research did not take edge computing integration under consideration, which is a key aspect of our investigation. Ahmed et al. (2023) [26] put forward a strategy of federated learning with reinforcement, combined with algorithms for local search (FLRLS), aimed at data in the medical field. The result they found was an increase in prediction accuracy when it comes to identifying urinary diseases. But this method brought about extra computational overhead, which raised concerns regarding its scalability for clinical uses in real-time. Jaladanki et al. (2021) [30] created a federated learning framework for analyzing data of COVID-19 patients from multiple healthcare institutions. The findings revealed that this method worked better than local models, and it also provided superior privacy safeguard measures. However, the data of this research is only concentrated in specific regions. This presents issues for wider use, especially in varying clinical environments. Bukhari et al. (2024) [27] have proposed a system of detection for intrusions based on federated learning, using SCNN-BiLSTM models. This is intended especially for wireless sensor networks. The model has been improved in terms of accuracy when detecting, and also reduced the delay time. But, its concentration on cybersecurity restricts direct pertinence to medical uses, and changing for healthcare settings remains unexplored. Rahmani et al. (2018) [31] have introduced a healthcare monitoring system based on fog computing that can process biomedical signals in real time. The research shows a clear reduction of delay with the help of edge computing, but it doesn’t include federated learning. As such, this method lacks a privacy-protective approach for delicate medical information. Khan et al. (2020) [32] explored hybrid deep learning structures for classifying multivariate time series, which is important for modeling changing physiological signals. But their attention to central learning fails to consider privacy issues, and it doesn’t incorporate edge computing for real-time analysis. Dang et al. (2022) [33] have applied federated learning methods to electronic health records (EHRs). This resulted in better model performance as well as preserving patient data privacy across many hospitals in the U.S. However, its focus on EHR data lacks specificity for real-time monitoring or citrate-related complications. Huang et al. (2023) [34] explored a federated learning (FL) model for predicting acute kidney injury (AKI) in ICU patients. Their study emphasized real-time health data analysis and privacy-preserving features. However, the research was limited narrow focus on a particular geographic location, which limits in generalizability to different populations. The model’s performance decreased in hospitals with different AKI prevalence rates. Pan et al. (2023) [35] developed a decentralized federated learning framework that allows multiple hospitals to train a shared predictive model without sharing raw patient data. While the system effectively demonstrated timely prediction and data privacy, it did not incorporate edge computing for real-time responses, and its focus on sepsis detection limits applicability to citric acid overdose monitoring.

While these studies collectively offer valuable contributions to federated learning and medical data analysis, several limitations hinder their application to our research focus:

1. 1.

None of the reviewed studies fully integrates edge computing for real-time data processing, which is critical for the early detection of citric acid overdose.

2. 2.

Despite research on electrolyte monitoring and patient data analysis, there is a noticeable gap in studies addressing regional citrate anticoagulation, particularly in the context of federated learning.

3. 3.

Computational overhead remains a common challenge, particularly for reinforcement learning-based federated systems, affecting their feasibility for large-scale, real-time medical applications.

4. 4.

Many studies relied on limited or single-center datasets, reducing the broader applicability of their findings across diverse patient populations.

5. 5.

Techniques such as FedProx show promise but require complex parameter tuning, which may hinder their deployment in dynamic clinical environments.

6. 6.

Several studies investigated either federated learning or edge computing independently. However, the synergistic combination of both technologies, vital for real-time and privacy-preserving healthcare applications, remains underexplored.

These gaps highlight the necessity of developing a framework that integrates edge computing and federated learning for real-time monitoring and adjustment of regional citrate anticoagulation, aiming for scalable, privacy-preserving, and clinically effective solutions.

Problem formulation

The MIMIC-IV dataset is a publicly available, de-identified electronic health record database containing high-resolution ICU and hospital data [36]. For this study, adult patients with acute kidney injury (AKI) admitted to the ICU for the first time were extracted from MIMIC-IV to detect citric acid overdose. Since federated learning requires decentralized data settings, we simulated the dataset distribution across multiple clients to mimic a federated learning environment, enabling evaluation of the model under realistic privacy-preserving constraints.

We aim to detect citric acid overdose and adjust RCA accordingly. Our approach employs federated learning (FL) to minimize a loss function \(L\), representing the prediction error of citrate levels and other CRRT parameters.

Mathematical formulation

Citric acid overdose detection

We define a dataset as shown in Eq. (1):

$$ D = \{(x_i, y_i)\}_{i=1}^{N}$$

(1)

where:

*

\(x_i\): Input features including vital signs, CRRT parameters, and biochemical markers.

*

\(y_i\): Citrate concentration (outcome variable).

The goal is to learn a predictive model \(f(x_i; w)\) parameterized by \(w\) as shown in Eq. (2):

$$ \hat{y}_i = f(x_i; w)$$

(2)

where \(\hat{y}_i\) represents the predicted outcome.

We define an empirical risk minimization objective Eq. (3):

$$ L(w) = \frac{1}{N} \sum_{i=1}^{N} \ell(f(x_i; w), y_i)$$

(3)

where \(\ell\) is an appropriate loss function:

*

Mean Squared Error (MSE) for regression.

*

Binary Cross-Entropy (BCE) for classification (overdose detection as a binary problem).

Federated learning-based optimization

Federated learning updates the global model \(w_t\) iteratively based on local models trained on edge devices. At each communication round \(t\), a subset \(S_t\) of edge devices is dynamically selected based on computational resources and proximity.

Each client \(i\) minimizes its local loss function as shown in Eq. (4):

$$ L_i(w_t^i) = \frac{1}{|D_i|} \sum_{(x_j, y_j) \in D_i} \ell(f(x_j; w_t^i), y_j)$$

(4)

where \(D_i\) is the local dataset on client \(i\).

The global model is updated via weighted aggregation as shown in Eq. (5):

$$ w_{t+1} = \sum_{i \in S_t} \frac{|D_i|}{\sum_{j \in S_t}|D_j|} w_t^i$$

(5)

where clients with larger datasets contribute proportionally more to the global model.

Dynamic client selection

To ensure efficient training, we dynamically select clients \(S_t\) based on:

*

Computational capacity: Clients must satisfy \(R_i(w_{t+1}^i) \leq B_i\), where \(R_i\) is the resource cost and \(B_i\) is the available budget.

*

Data quality: Clients with higher data relevance are prioritized.

*

Proximity: Closer clients reduce communication overhead.

Selected clients participate in training while maintaining differential privacy by adding controlled Laplacian noise to local updates, which is shown as a formula (6):

$$ \hat{w}_{t+1}^i = \text{AddNoise}(w_{t+1}^i, \epsilon)$$

(6)

where \(\epsilon\) is the privacy budget.

To detect citric acid overdose and adjust RCA using federated learning, we aim to minimize a loss function \( \mathcal{L} \) that represents the error between the predicted and actual outcomes related to patient conditions (e.g., citrate levels and other CRRT parameters).

System design

Federated learning architecture

Figure 1 shows the system structure that mixes edge computing and federated learning. It’s used for early detection of citric acid overdose when using regional citrate anticoagulation (RCA). Many devices at the edge gather local clinical information and train separate models without moving raw data, keeping patient details safe. These models are forwarded to nodes on the edge, which then pass them onto a server located at an edge point, where they are combined in a federated way into one global model. The updated global model is sent back to the edge devices, allowing ongoing enhancement in real-time overdose forecasting while keeping data secrecy. The system uses a single global model to serve as a common aggregation point for all local learning updates from edge devices.

[IMAGE OMITTED: SEE PDF]

The proposed system uses a Resource-Aware Federated Learning with Dynamic Client Selection (RAFL-Fed) algorithm. The global model \( w^t \) is updated iteratively based on client contributions. The system selects clients dynamically based on their computational resources, represented by a capacity constraint \( R_i(w^{t+1}_i) \leq B_i \), where \( B_i \) is the computational budget of client \( i \).

Client selection

The set of clients \( S_t \) is selected dynamically at each communication round. Client \( i \in S_t \) that meet the computational budget \( B_i \), have sufficient capacity \( c_i \), and are within acceptable proximity \( p_i \) are selected for participation. This can be formulated as Eq.(7):

$$ S_t = \{i \mid R_i(w^{t+1}_i) \leq B_i \land c_i \geq \theta \land p_i \leq \delta \}$$

(7)

This indicates that only clients that meet the computational budget \( B_i \), have sufficient capacity \( c_i \), and are within acceptable proximity \( p_i \) are selected for participation.

where \( \theta \) and \( \delta \) are threshold values for capacity and proximity, respectively.

Differential privacy

To ensure patient data privacy, we introduce differential privacy in the federated learning process. After training the local model, each client adds noise to the model parameters before sending them to the central server. This noise is generated using the Laplace mechanism, which is shown in Eq. (8):

$$ \hat{w}_i^t = w_i^t + \text{Lap}\left(\frac{\Delta f}{\epsilon}\right)$$

(8)

where \( \Delta f \) is the sensitivity of the model update, and \( \epsilon \) is the privacy budget. The noise ensures that individual data points cannot be inferred from the model updates.

BiLSTM-GRU

The core of the predictive model is a hybrid BiLSTM-GRU model as shown in the Fig. 2.

[IMAGE OMITTED: SEE PDF]

BiLSTM layer

The bidirectional LSTM layer captures both forward and backward temporal dependencies in the patient’s data as shown in Eq. (9).

$$ \begin{aligned} \overrightarrow{h}_t &= \text{LSTM}(x_t, \overrightarrow{h}_{t-1}) \\ \overleftarrow{h}_t &= \text{LSTM}(x_t, \overleftarrow{h}_{t+1}) \end{aligned}$$

(9)

where \( \overrightarrow{h}_t \) and \( \overleftarrow{h}_t \) are the forward and backward hidden states at time \( t \), respectively.

The output of the BiLSTM at each time step can be formulated as in Eq. (10):

$$ h_t = \overrightarrow{h}_t + \overleftarrow{h}_t$$

(10)

GRU ayer

The GRU layer efficiently captures sequential patterns with fewer parameters which is shown in the Eq. (11):

$$ \begin{aligned} z_t &= \sigma(W_z x_t + U_z h_{t-1}) \\ r_t &= \sigma(W_r x_t + U_r h_{t-1}) \\ \tilde{h}_t &= \tanh(W_h x_t + U_h (r_t \odot h_{t-1})) \\ h_t &= (1 - z_t) \odot h_{t-1} + z_t \odot \tilde{h}_t \end{aligned}$$

(11)

where \( z_t \) is the update gate, \( r_t \) is the reset gate, and \( \tilde{h}_t \) is the candidate activation.

Final output

The final output \( y_t \) for binary classification and continuous regression is obtained by passing the combined output of both the BiLSTM and GRU layers through a dense layer, as shown in Eqs. (12) and (13) respectively:

For Binary Classification (using sigmoid activation):

$$ y_t = \sigma(W_f h_t + b_f)$$

(12)

where \( W_f \) is the weight matrix and \( b_f \) is the bias term.

For Continuous Regression (using linear activation):

$$ y_t = W_f h_t + b_f$$

(13)

where \( W_f \) and \( b_f \) are the learned parameters of the dense layer.

RAFL-Fed algorithm

In the Algorithm 1, calculation of importance weights \(p_{i,j}\) as shown in Eq. (14) for every edge device \(i\) is based on a combined measure that includes three main elements:

1. 1.

Data quality (\(Q_i\)): This represents the statistical usefulness and variety of the local dataset \(D_i\). It is usually evaluated using metrics like entropy, balance in class distribution, or representational completeness.

2. 2.

Proximity (\(L_{i,j}\)): This refers to the physical or network proximity between the edge device \(i\) and the data source or user node \(j\), which affects latency and reliability. Proximity can be quantified via signal strength, hop count, or communication delay.

3. 3.

Resource usage (\(R_i\)): This represents the computational effort and energy use at device \(i\) during local training. It is important for making sure resources are used efficiently, following the rule \(R_i(w_i^{(t+1)}) \leq B_i\).

The weight \(p_{i,j}\) is computed as a normalized weighted sum of these factors:

$$ p_{i,j} = \frac{\alpha \cdot Q_i + \beta \cdot \left(\frac{1}{L_{i,j}} \right) + \gamma \cdot \left(\frac{1}{R_i} \right)}{\sum_{k \in S_t} \left[ \alpha \cdot Q_k + \beta \cdot \left(\frac{1}{L_{k,j}} \right) + \gamma \cdot \left(\frac{1}{R_k} \right) \right]}$$

(14)

where \(\alpha\), \(\beta\), and \(\gamma\) are hyperparameters that control the relative importance of each criterion.

In practical medical settings, devices like CRRT machines and patient monitors are usually used within a specific hospital network. To calculate closeness (represented as \(L_{i,j}\)), we use real-time delays in the network signal (like round-trip time or ping latency) along with available physical location data (for instance, from device registration information or RSSI-based triangulation). These techniques are simple, don’t need big infrastructure changes, and can be regularly checked to keep them accurate. The usage of resources (\(R_i\)) is computed by observing the present computational load, battery condition, and memory consumption of each device. We collect these measurements using internal system monitors or light monitoring agents operating on the devices. This data enables us to apply the limitation:

$$ R_i\left(w_i^{(t+1)}\right) \leq B_i \quad \text{(Resource Constraints)}$$

(15)

Here, \(B_i\) stands for the utmost limit of resources allowed for device \(i\) when it is in the update stage The metrics are gathered by the edge server from time to time, and it chooses clients for each round \(t\) through the use of these two constraints as given in Eqs. (15) and (16):

$$ L_{i,j} \leq \tau \quad \text{(Proximity/latency threshold)}$$

(16)

In this context, \(\tau\) is an adjustable parameter that signifies the highest acceptable delay to make sure updates are done in time during federated aggregation.

The RAFL-Fed Algorithm 1 and 2 are developed to assist federated learning in a health-related setting. This is where various edge devices work together to enhance a global model while ensuring patients’ privacy remains intact. To start, the algorithm sets up global model parameters using a Hybrid BiLSTM-GRU architecture as shown in Table 2 for successfully capturing time-based dependencies and stressing key data features. The procedure follows an organized sequence of communication rounds marked by \( T \). Each round has numerous steps that are built towards refining the model collectively.

[IMAGE OMITTED: SEE PDF]

[IMAGE OMITTED: SEE PDF]

[IMAGE OMITTED: SEE PDF]

In every round of communication, a group of edge devices (denoted as \( S_t \)) is dynamically chosen, considering their geographic closeness and computing ability. This makes certain that the resource cost (\( R_i(w^{(t+1)}_i) \)) doesn’t go beyond the budget available (\( B_i \)), for each client denoted by \( i \in S_t \). After this step, we distribute global model parameters named \( w_t\) to these selected edge devices. Each device then proceeds with local training, employing a differential privacy method ensuring protection of patient data. Data from local patients is gathered through onboard sensors or Continuous Renal Replacement Therapy (CRRT) machines. After collection, this data is divided into small batches for training purposes.

During the local training, each client changes its model parameters using gradient descent throughout many epochs. After local training is done, another cycle takes place across all data to improve the model parameters. To ensure privacy and not expose any individual patient’s information, Laplacian noise is introduced when updating the local models, which guarantees Differential Privacy.

Each edge device returns two model updates to the server:

*

\(w_i^{(t+1)}\): the locally trained model without noise.

*

\(\hat{w}_i^{(t+1)}\): the differentially private (DP) version of the model, produced by applying calibrated Laplacian noise based on the privacy parameter \(\varepsilon\).

This dual-return strategy serves two purposes:

1. 1.

The private model (\(\hat{w}_i^{(t+1)}\)) gets utilized for aggregation so that global model updates adhere to the rules of \(\varepsilon\)-differential privacy.

2. 2.

The non-private model (\(w_i^{(t+1)}\)) is applied for internal observation or optional assessment (for instance, local accuracy verifications, anomaly detection, or analysis of client drift), and this does not affect the privacy-preserving learning cycle.

This separation gives flexibility for maintaining strict privacy rules, but also permits additional understanding of the basic local performance.

The model updates from those clients taking part are combined according to calculated importance weights \( p_{i,j} \), which show data quality, closeness, and use of resources. A top-K selection method finds the most relevant devices that can contribute to this combining process. Choosing the top-\(K_i\) devices in Algorithm 1 based on importance weights \(p_{i,j}\) is a crucial strategy to improve both stability and communication effectiveness when gathering data. This is because not all clients selected in \(S_t\) have equal contribution - this can be due to variations in the quality of their data, how close they are, and resource efficiency differences. So we sort every device in \(S_t\) using calculated \(p_{i,j}\) values and select the top-\(K_i\) subset that provides us with the most dependable and representative updates. Global model parameters are updated with the weighted average of these local updates. With this, we round off one cycle to improve our collaborative model while keeping some important rules regarding privacy protection. The algorithm presented here perfectly balances the requirement for precise training models and the essential and regulatory needs of maintaining patient confidentiality in a distributed healthcare setup.

Experimental setup

The experiment utilized a system with an Intel® Core™ Ultra 7 155 H processor, an NVIDIA® GeForce RTX™ 4070 Laptop GPU (8GB GDDR6 VRAM), and 16GB LPDDR5X-7467MHz RAM. TensorFlow 2.18.0 was used for implementation. In RAFL-Fed, a Deep Learning-based binary and continuous model was built using a BiLSTM architecture with Keras.

Dataset

We use the MIMIC IV dataset [36], which contains high-resolution ICU data, including vital signs and laboratory results for detecting citric acid overdose. The dataset contains time-series data with multiple features, including:

*

Vital signs: heart rate, respiratory rate, etc.

*

Laboratory results: calcium, citrate levels.

The non-IID data distribution in the proposed federated setup stems from label skew, where each client receives a differing ratio of overdose vs. non-overdose cases, thereby simulating real-world heterogeneity in class distributions across decentralized data sources.

In our study, the non-IID data distribution was deliberately introduced using label skew and quantity skew mechanisms. Specifically, we partitioned the dataset across clients such that each client received a different proportion of samples corresponding to the binary target class, “citric acid overdose” (i.e., \(y_{\text{binary}}\)). The create_non_iid_data() function separates the data into overdose (\(\text{label} = 1\)) and non-overdose (\(\text{label} = 0\)) categories, and shuffles them independently. These two sets are then unevenly split among the clients, resulting in some clients receiving a higher concentration of one label over the other. This simulates label distribution skew across clients. Also, since the data divisions are created by splitting label arrays of different lengths using np.array_split(), the number of samples for each client is not the same. This leads to quantity skew. No clear feature skew was brought in; after standardization, the feature distributions stay uniform among clients. Our heterogeneity design mainly concentrates on label skew and quantity skew, which match common non-IID data scenarios in federated learning.

Feature selection

The most relevant features for the intended goal of early detection of citric acid overdose and adjusting regional citrate anticoagulation are:

Kidney function markers (creatinine levels)

Creatinine is a key indicator of renal function. Acute kidney injury (AKI) is the primary reason for initiating CRRT and impaired kidney function affects citrate clearance. High levels of creatinine show that the glomerular filtration rate (GFR) is lower, meaning we need to change CRRT settings to reduce citrate build-up. Relevant features:creatinine_min, creatinine_max, creat.low_past_48hr.

Acid-base balance (bicarbonate, base excess, pH)

When the liver metabolizes citrate, it produces bicarbonate that can cause metabolic alkalosis if too much is taken. On the other hand, when citrate metabolism doesn’t work well (like in liver problems), it may result in acidosis because of leftover citrate. Relevant features:bicarbonate, baseexcess, pH_minExample: pH \( < \) 7.35 signals acidosis, whereas high levels of bicarbonate hint at alkalosis.

Electrolytes (calcium, potassium, sodium)

Citrate makes ionized calcium less active, which results in hypocalcemia. This can lead to instability in neuromuscular as well as cardiac arrhythmias. The balance of potassium and sodium might get disturbed because of the fluid changes during CRRT or due to metabolic issues caused by citrate. Relevant features:calcium_min, calcium_max, potassium_min, potassium_max, sodium_min, sodium_max.

Liver enzymes (ALT, AST)

The liver metabolizes citrate into bicarbonate. Elevated ALT/AST indicates hepatic dysfunction, impairing citrate clearance and increasing overdose risk. Relevant features:alt.min, ast.max.

Coagulation markers (INR, PTT)

Prolonged INR (\( > \)1.5) or PTT (\( > \)60s) signals over-anticoagulation, increasing bleeding risk. These markers guide RCA adjustments to balance circuit patency and patient safety. Relevant features:INR_max, ptt_max.

Lactate and hemodynamic stability

Raised lactate levels show that there might be a lack of oxygen in the tissues or problems with metabolism, often made worse by too much citrate. Key signs like heart rate and blood pressure give immediate understanding about hemodynamic stability – how well blood flow is being maintained, which can become unstable because of low calcium levels (hypocalcemia) or imbalances in acid and base substances in the body. Relevant features:lactate_max, heart_rate, sbp_min, sbp_max, resp.rate.

Urine output

Reduced urine output indicates worsening renal function, affecting fluid balance and citrate clearance. This informs decisions on CRRT intensity and citrate dosing. Relevant features:uo.rt_6hr, uo.rt_24hr.

Clinical alignment

The selected features align with established clinical guidelines for CRRT and RCA management (for example, KDIGO, ISECM). They represent criteria usually observed in ICU environments to avoid citrate toxicity. These include: Calcium levels for the prevention of hypocalcemia; pH and bicarbonate to identify metabolic disorders; Coagulation markers for preventing bleeding issues; Liver and kidney function tests evaluating citrate metabolism and clearance.

Exclusion of non-prioritized features

Citrate might also influence other electrolytes, like magnesium or phosphate. However, these may not be included, probably because they have less predictive value in initial tests, and there is limited data available about them.

Binary target for citric acid overdose detection

In the context of detecting citric acid overdose, the following markers are used to establish binary targets:

*

Lactate max: Lactic acidosis, which may be associated with citrate overdose, is indicated by a lactate level above 2.5 mmol/L. This threshold is derived from clinical observations.

*

pH min: A pH of less than 7.35 is taken to be acidosis, which accompanies overdose of citric acid. This value is generally accepted as the threshold for acidic blood pH.

*

Platelets min: A platelet count below 150 × 109/L indicates thrombocytopenia, which can occur due to citrate toxicity.

*

INR max: An INR of above 1.5 indicates the presence of coagulopathy or bleeding risk, and anticoagulants like citrate may increase this further. As the INR may change during continuous anticoagulation therapy, this threshold has to be put into perspective with the history of treatment of the patient.

Continuous target for anticoagulation adjustment

For adjustments in ongoing anticoagulation, the following continuous targets are followed:

*

Calcium min: A calcium level below 1.0 mmol/L is crucial for determining the need for anticoagulation adjustments. Citrate use can reduce calcium levels, impacting blood clotting.

*

PTT max: A Partial Thromboplastin Time (PTT) \( > 60 \) seconds is evidence of prolonged clotting times and serves as a cut-off value for adjusting the dosage of anticoagulation. This threshold level can be appropriate for detecting over-anticoagulation conditions.

*

Platelets and INR adjustments: Low platelet counts and high INR values suggest an increased bleeding risk. Adjusting anticoagulation based on these factors is consistent with standard medical practice.

Data distribution analysis

The histograms shown in Fig. 3 present the distribution of some numeric features crucial for the early detection of citric acid overdose and the adjustment of regional citrate anticoagulation. These were obtained from the data gathered on edge computing nodes within the federated learning framework. The upper-left plot shows the distribution for lactate_max. It is right-skewed, since most of the data points are toward the lower values, but there are fewer instances of high values. It would be a typical shape, suggesting that high lactate values (overdose or other complications) are rather seldom in this dataset. The distribution of pH_min in the upper-mid plot is more symmetrical and roughly bell-shaped, indicating that the pH values are normally distributed but also slightly skewed. It is important to determine which acid or base component deviates from its normal range, as an indication of acidosis or alkalosis, a problem that might occur in citrate metabolism. The distribution of platelets_min in the upper-right plot is highly right-skewed; most values are huddled towards the lower extreme of the scale, reflecting low counts of platelets in some patients. This pattern is important because low platelet counts are critical in anticoagulation therapy and might imply coagulation disorders or complications in citrate anticoagulation. The bottom-left plot shows the distribution of INR_max, also highly right-skewed. Most values are huddled near the lower end, indicating that extreme increases in INR, potentially problematic when it comes to coagulation)—are infrequent. The pattern now becomes important to consider when adjusting anticoagulation therapy in patients. The distribution of calcium_min in the bottom-mid plot is close to normal distribution. The bell-shaped curve suggests that the calcium levels are huddled around a mean value in most patients. Calcium levels are fundamental for the balance of citrate anticoagulation and directly affect citrate metabolism. Lastly, the bottom-right plot represents ptt_max with a distribution slightly right-skewed. Though there is a notable peak toward lower values, there is an accumulation of values toward the top, which might point out that some patients have been recording raised levels of PTT. This could indicate problems in achieving blood clotting and could be an important aspect in managing anticoagulation and identifying complications.

[IMAGE OMITTED: SEE PDF]

Training procedure

Local training

Each edge device collects patient data and trains the local model using real-time CRRT parameters. Data is split into batches of size \( B \), and the model is updated using gradient descent:

$$w \leftarrow w - \eta \nabla \mathcal{L}(w, D_i)$$

where \( \eta \) is the learning rate, and \( D_i \) is the local dataset.

Global aggregation

After local training, each client sends the differentially private model updates \( \hat{w}_i^t \) to the central server. The global model is updated using weighted aggregation as discussed in Section Results and discussion.

Evaluation metrics

The system is evaluated based on:

*

Sensitivity (True Positive Rate):

$$ \text{Sensitivity} = \frac{\text{True Positives}}{\text{True Positives} + \text{False Negatives}} $$

*

Specificity (True Negative Rate):

$$ \text{Specificity} = \frac{\text{True Negatives}}{\text{True Negatives} + \text{False Positives}} $$

*

Accuracy:

$$ \text{Accuracy} = \frac{\text{True Positives} + \text{True Negatives}}{\text{Total Predictions}} $$

*

Latency: The time taken for each prediction.

Simulation

Hyperparameters like learning rate (0.01), LSTM units (128), batch size (32), number of local epochs (10), and global iterations of 20 rounds were adjusted based on practical experimentation using centralized validation before starting federated training. These configurations were selected to optimize model performance while maintaining computational efficiency among different clients. To select clients, we implemented a tactic to choose a subset of them from a bigger group, depending on their resource consumption and a certain budget limit. This simulates assigning an arbitrary budget for every client, symbolizing the maximum resources each can manage to use. For all individual clients, it obtains system resource measurements (CPU usage, memory usage, and network latency), then calculates overall “resource cost” utilizing a helper function. When the resource cost of a client fits within both their budget and the overall budget limit, they qualify for selection. The function then picks a final number from the eligible clients’ pool using the minimum value between the qualified clients’ number and another value coming from the total clients count and the budget limit. Randomly selecting the last set of clients happens from this eligible group, promoting variety while adhering to resource limits. This method makes sure of efficient and budget-conscious client involvement in the learning process.

The key components in setting up a simulation of binary classification and continuous regression models involve several steps that must be considered for effective data learning and performance evaluation. First, the dataset is extracted from a CSV file containing various medical features relevant to predicting citric acid overdose and changes in anticoagulation. Different steps in preprocessing include handling missing values by replacing them with the mean value of a respective feature, creating binary labels about citric acid overdose or not based on certain conditions, and creating continuous labels about modifications to anticoagulants based on health measures. The values in these features are scaled using StandardScaler to contribute equally to training.

Then, the data is split into an 80–20 ratio for training and testing groups. This gives the model sufficient data to learn from and a dataset to keep for evaluation. The binary classification model is developed using BiLSTM and GRU layers, which are ideal for sequential or time series information. This results in an output layer powered by a sigmoid function. On the other hand, the continuous regression model has a similar structure, but it includes a linear output layer to project continuous values.

The training of these models is conducted with a federated learning method, which creates a situation where many edge devices do their local training using their data. This practice involves numerous rounds of communication, during which chosen clients update the global model according to results from local training. Every client trains on its data set for several epochs or cycles before sending back updated information about the model to the central server. The server collects these updates to enhance the weight of the global model, maintaining the advantages of decentralized data handling.

After every round of communication, the overall model is tested on metrics such as loss, accuracy, sensitivity, specificity, and mean absolute error (MAE), including latency. This gives a perspective on how suitable this model is for clinical applications. While training happens continuously, these parameters are plotted to see changes in training accuracy, like loss rate or sensitivity over time, which helps us thoroughly gauge the proposed model’s performance.

Results and discussion

Performance comparison of the proposed method in different clients

Results on IID datasets

The model of binary type, given the job to find out overdoses of citric acid for ten clients, gives a well-balanced result with an accuracy level of 96.15%, which means it rightly marks more than 96% cases. What is even more significant is that its sensitivity stands at 98.08%. This suggests that this model effectively identifies the majority of instances involving true overdose. In healthcare, such ability holds high importance since failing to spot a positive case could lead to serious consequences. The model also shows a specificity of 94.49%, proving that it is just as skilled in identifying patients who do not suffer from overdose, thereby reducing the possibility of needless treatments. Moreover, with a latency duration of merely 0.229 seconds, it gives quick predictions, which makes it suitable for use in real-time clinical settings. The continuous model is good at forecasting changes in anticoagulation. For 10 Clients, the continuous model exhibits solid performance, with a Mean absolute error (MAE) of 0.2081. This conveys that, normally, the predictions from this model are quite near to the actual adjustments needed—this being very important, specifically in medical scenarios where exact doses can notably affect patient safety and results. Moreover, the loss value (mean squared error) stands at 0.0788, indicating a high accuracy level of overall predictions by this model. Moreover, having a latency of around 0.229 seconds, the model is very effective and can provide predictions speedily enough to back immediate decision-making in a healthcare setting.

Moving to 20 Clients, the binary model experiences a slight decline in performance. The loss increases to 0.4460, and accuracy drops to 0.8195, indicating that the model struggles slightly more with classification as the client size grows. Sensitivity decreases to 0.8466, suggesting a reduced ability to correctly identify positive cases, though specificity remains steady at 0.7961, implying that the model’s capacity to detect negative cases remains unchanged. Notably, the latency decreases to 0.5279 seconds, which, while not as fast as for 10 Clients, still reflects efficient processing times, especially given the increase in client size. This suggests that while classification accuracy and sensitivity suffer slightly with more clients, the model becomes more efficient in processing data. For 20 Clients, the continuous model maintains nearly the same level of performance, with a loss of 0.0663 and an MAE of 0.1931, reflecting a decrease in prediction error. The latency increases to 0.6901 seconds, indicating a slight drop in processing efficiency as the client grows, but the overall performance remains stable. This suggests that the continuous model can handle moderate increases in data size without significant loss in accuracy, though some processing overhead becomes evident.

For 40 Clients, the binary model’s performance declines even further. The loss increases to 0.4957, while the accuracy drops to 0.8047, showing that the model has fallen a bit more in its performance of correctly classifying data. Sensitivity stays relatively stable at 0.8498, while specificity goes down to 0.7658, which means a greater number of false positives. Hence, the model struggles more to tell apart negative cases. However, this comes with a drop in accuracy. It improves the latency at 0.2390 seconds, meaning this can significantly speed up when some optimization kicks in with larger data sets. Overall, even though the binary model faces certain difficulties in maintaining the high values of accuracy and specificity while increasing the number of clients, it does turn around with its compensation on the processing time, which is considered vital for real-time medicine applications. This means that finally, for 40 Clients, the continuous model realizes its finest performance. It can be seen that there is a decrease in the loss to 0.0613, with the MAE decreasing to 0.1806, which means this model improves in terms of prediction accuracy as the number of clients increases. Furthermore, coupled with this is the considerable reduction in latency, coming down to 0.1573 seconds, hence making this model of higher accuracy and even quite efficient in handling larger volumes of data. This trend suggests that the continuous model scales well in practice; as the size of a client increases, it has both better performance and lower latency.

Results on non-IID datasets

The model for binary classification, aimed at identifying early signs of citric acid overdose, displays changing performance with the rise in client numbers. When it comes to 10 clients, the model reaches a high accuracy of 95.71%, having a loss of only about 0.2469. The balance between sensitivity (observed as 0.9968) and specificity (observed as 0.9229) suggests strong capability in both detecting positive overdose cases and accurately pointing out non-overdose situations, as shown in Table 3. The latency, a very important aspect in edge computing for real-time detection, stays low at 0.233 seconds. Figure 4 illustrates that the training accuracy is rising constantly, hitting about 92% by round 20. On the other hand, even though the training loss varies at times, it generally falls over time. This shows a good learning process despite some early unpredictability. Even if there are some small surges now and then, the overall decrease in loss hints that the global model is doing a great job reducing error as rounds progress; this happens sometimes when client updates don’t perfectly match up, though. After the initial few rounds pass by, latency comes down drastically and steadies around 0.2 seconds, which tells us communication becomes quite effective post the first phase of learning. In Table 3, the continuous regression model, aimed at adjusting citrate anticoagulation, also experiences variations in performance across different client numbers. For 10 clients, the model achieves a low loss of 0.0812 and an MAE of 0.1731, with a latency of 0.123 seconds, indicating effective and timely adjustments in anticoagulation.

[IMAGE OMITTED: SEE PDF]

[IMAGE OMITTED: SEE PDF]

When the client number grows to 20, an accuracy of 92.75% is seen in a binary model with a loss rate of 0.2487, as shown in Table 3. Sensitivity reaches 0.9361, and specificity 0.9201, hinting at an overall better result for adequately classifying both positive and negative cases. Latency stays around 0.2008 seconds, revealing that despite having more clients, the model can work in a real-time edge environment without much delay. The model achieves its MAE of 0.2329 in a continuous model, with minor deviations during prediction, with a low loss of 0.0946. The latency is about 0.1486 seconds, displaying that the model’s efficiency has been increased for managing a large number of edge devices, but still keeping good performance levels intact. Figure 5 demonstrates a significant enhancement in accuracy, hitting around 96% by the 20th round. This indicates superior learning convergence, probably because of marginally higher data diversity that can lead to more thorough model updates. Compared with Fig. 4, the training loss is generally steadier and has a lower average loss, implying that the model adjusts better, even though there is an increase in non-IID. Like the situation with 10 clients in the non-IID dataset, the latency diminishes rapidly and retains a low, consistent value. This indicates good communication even when data distribution is more varied.

[IMAGE OMITTED: SEE PDF]

On further increasing clients to 40, accuracy decreases to 86.24%, and loss increases to 0.5658, as shown in Table 3, which shows the overall decrease in performance compared to clients 10 and 20. Sensitivity increases to 0.9712, while specificity falls to a lesser value, such as 0.7686, showing many false positives compared to clients 10 and 20. The latency remains around 0.2061 seconds, but with this trade-off between sensitivity and specificity, it gives the impression that the model faces problems in generalizing well if client numbers increase, especially in a non-IID setup. A similar case arises in the continuous model as efficiency reduces compared to 10 and 20 clients. The MAE increases to 0.3155, and the loss increases to 0.1126. Even so, latency stays low at around 0.1899 seconds, showing that the model can still provide fast outcomes in real-time, although with slightly less prediction precision. Figure 6 shows that the model gets very close to full accuracy, almost reaching 100% by round 20. This tells us that our model improves a lot from having more varied data, maybe recognizing many different patterns. After an early jump, the loss settles at a lower baseline compared to previous situations, which means there is fast learning in the beginning and then steady progress over time. But still, this starting high point indicates some initial difficulties with aligning lots of diverse local models. The delay curve resembles the previous figures. It has a steep drop at first, then it levels out, showing steady communication times. This shows that our communication performance remains reliable even when data heterogeneity increases.

[IMAGE OMITTED: SEE PDF]

The outcomes show that the use of federated learning on edge devices to identify citric acid overdose and regulate anticoagulation works well with up to 20 clients. Both binary and continuous models achieve high precision and low delay. Yet, as client numbers rise to 40, performance begins declining, especially in the binary model, probably because data and non-IID characteristics of our dataset become more varied. This implies healthcare applications using federated learning frameworks within edge computing might be better off working with a perfect number of clients to maintain both model correctness and ensure communication efficiency.

Comparison with state-of-the-art

Results on non-IID datasets

The binary model performance shown by the proposed RAFLFed framework outperforms competitive state-of-the-art FLRL and FLSL. RAFL-Fed meets a loss of 0.2469, respectively, which is comparatively reduced compared to the loss concerning FLRL and FLSL. Compared with both FLRL (0.2664) and FLSL (0.3372), RAFL maintains more stable and optimum performance. This yields an accuracy of 0.9571, higher than FLRL (0.9482) and FLSL (0.8580). In addition, the sensitivity of RAFL-Fed, represented by 0.9968, and specificity by 0.9229, also outperformed other frameworks, which is very important to reduce false positives and false negatives in any classification task. The latency of 0.233 seconds is remarkably lower compared to FLRL (0.634s) and FLSL (0.365s); hence, RAFL-Fed is faster and more efficient for real-time applications on edge devices. This combination of high precision, low loss, good sensitivity, specificity, and low latency showcases the proposed model’s effectiveness for binary classification tasks.

For continuous model performance in Table 4, RAFL-Fed continues to show robust results. The RAFL-Fed, with a loss of 0.0812, outperforms FLRL with 0.1015 and FLSL with 0.0813 in terms of loss; therefore, it is more accurate in regression predictions. The MAE of 0.1731 further ascertains the model’s aptitude for making precise estimates. RAFLs MAE of 0.1731 is competitive compared to FLRL at 0.2158 and FLSL at 0.2304, making the model viable for continuous prediction tasks. In terms of latency, RAFL-Fed performs with a response time of 0.123s, better than FLRL at 0.687s and FLSL at 0.2304 seconds. Although latency can vary across different deployment environments, RAFL-Fed generally strikes a good balance between prediction accuracy and timely responses. This makes RAFL-Fed a viable alternative for applications that require real-time performance and robust continuous predictions in edge computing. RAFL-Fed has demonstrated very strong performance both on the binary and continuous models by outperforming the state-of-the-art methods currently on key metrics such as accuracy, loss, sensitivity, and latency.

[IMAGE OMITTED: SEE PDF]

Results on IID datasets

The binary model performance of the proposed RAFL-Fed framework is superior to state-of-the-art methods. RAFL-Fed achieves a loss of 0.2625, which is significantly lower than 0.4630 and 0.4938 in FLRL and FLSL, respectively. That proves that RAFL-Fed holds much more stability and keeps performance optimized, producing an accuracy of 0.9615, which is considerably higher than 0.5370 and 0.8314 as achieved from FLRL and FLSL, respectively. Moreover, sensitivity and specificity were also high compared to other frameworks, hence being crucial in keeping false positives and negatives to a minimum during classification. The latency amounts to 0.269 seconds, which is visibly lower than FLRL 0.904 seconds and FLSL (0.3548); hence, RAFL-Fed has promising potential for being used against real-time edge applications because of its faster and more efficient processing state. High accuracy, low loss, decent sensitivity and specificity, and latency prove that our proposed RAFL-Fed is a very efficient technique for binary classification.

On the continuous models, the expected best performance was by RAFL-Fed. The performance of RAFL-Fed outperformed with an MAE of 0.2081, which indicates that it yields more accurate regression predictions, whereas FLRL and FLSL achieved 0.2252 and 0.2452, respectively. Regarding loss, RAFL-Fed also performs better than the state-of-the-art. Similarly, in latency, RAFL-Fed outperforms other state-of-the-art. RAFL-Fed has generally presented strong performance for both binary and continuous models, showing the best results compared to state-of-the-art methods in major metrics such as accuracy, loss, sensitivity, and latency.

Performance comparison of different methods under various noise levels

The Table 5 compares the performance of three methods (Proposed, SCN-Deep BiLSTM, and FLRL) at different noise levels (0.02, 0.04, 0.06, 0.08, and 0.10). The experiment is carried out for 20 clients with a learning rate between 0.01 to 0.001. Each method is evaluated on both binary and continuous models with metrics including loss, accuracy/MAE, sensitivity, specificity, and latency. The proposed method performs well across noise levels, with the lowest loss values and high accuracy. For binary models, accuracy is high (above 90%) even at higher noise levels. Sensitivity and specificity are also strong, particularly at lower noise levels, maintaining effective classification with minimal false positives and negatives. Latency remains low for both model types, indicating good efficiency. SCN-Deep BiLSTM shows significantly lower accuracy and higher loss, especially at higher noise levels. Sensitivity and specificity are lower compared to the proposed method, which means it struggles more with false positives and false negatives. The latency is higher, particularly in continuous models. FLRLS shows the poorest performance, especially as noise increases. Loss values rise significantly, and accuracy drops drastically at higher noise levels. Both binary and continuous models exhibit reduced sensitivity and specificity, and the latency is similar to SCN-Deep BiLSTM in continuous models but is lower for binary models. The Fig. 7 visually compares the accuracy of the proposed SCN-Deep BiLSTM and FLRLS methods under different noise levels. The proposed method consistently outperforms both SCN-Deep BiLSTM and FLRLS across all noise levels, showing its robustness to noise. FLRLS and SCN-Deep BiLSTM both show a significant drop in accuracy as noise increases, with FLRLS particularly struggling at higher noise levels. This reinforces the proposed method’s superior performance in noisy environments.

[IMAGE OMITTED: SEE PDF]

[IMAGE OMITTED: SEE PDF]

Detailed trial results

The Table 6 gives a summary of how the suggested model performed in five separate trials, each with different random seeds. These experiments were done to calculate the 95% confidence intervals and check if the model is stable when starting conditions are randomly changed.

[IMAGE OMITTED: SEE PDF]

Note: The results we are discussing were achieved using specific random seeds to ensure they can be repeated. The trial seeds used were 2700, 3663, 2965, 5068, and 683.

Observations from multi-trial results

To settle doubts about the stability of our outcomes, we carried out 5 separate training experiments. Each one was started with a different random seed to understand the stochastic diversity in training results. The models were trained and assessed on MIMIC-IV dataset through our federated learning arrangement, and we present the average as well as 95% confidence intervals (CIs) for each important metric. The confidence intervals were computed using the t-distribution, which accounts for the small sample size and allows a statistically grounded estimation of variability across runs. This method reflects both central performance and dispersion, thus enabling a better understanding of the model’s stability. As seen in Table 7, the classification efficiency stays very stable throughout trials. The binary accuracy reached an average of 0.9672 with a small confidence range of \(\pm\)0.0068, indicating constant good generalization. Both binary sensitivity (0.9770 \(\pm\) 0.0081) and binary specificity (0.9587 \(\pm\) 0.0080) also show close grouping, showing that the model consistently recognizes both positive and negative instances in various attempts. These measures confirm how effective the model is when it comes to binary classification tasks within a federated arrangement. The binary loss shows more variability (0.1801 \(\pm\) 0.1025), which could indicate responsiveness to early-round convergence actions or exceptional batches in some federated cycles. In the same way, binary latency demonstrated a moderate difference (0.2700 \(\pm\) 0.1324 seconds), probably because of system-related elements like communication burden during various experiments. For the regression part, the regular MAE had an average of 0.6153 with a broader CI of \(\pm\)0.3715, showing somewhat higher variation in predicting continuous values. This matches well with seen changes in the ongoing loss (0.5045 \(\pm\) 0.3371), which could be affected by diverse data from edge devices. However, despite this variability, the constant latency was very steady (0.2225 \(\pm\) 0.0281 seconds), indicating predictable performance during runtime.

[IMAGE OMITTED: SEE PDF]

Conclusion

This work introduces a new solution for protecting privacy that functions in real-time to detect citric acid overdose. In addition, it successfully modified RCA protocols by utilizing federated learning along with edge computing techniques. RAFL-Fed is the novel algorithm developed for selecting clients efficiently based on available resources to maximize the use of edge devices. The model employs a mixed structure of BiLSTM-GRU, significantly boosting its capability to forecast results by comprehending complex time-based connections and emphasizing key features in the data. The experimental findings reveal that the proposed models are extremely accurate and responsive, which is essential for immediate medical assistance. The binary classification model demonstrates high responsiveness and a robust ability to detect possible overdose scenarios, whereas the continuous regression model provides precise adjustments for RCA protocols that are critically important for patient safety.

This way enhances decision-making in clinical environments while ensuring patient information security by processing data on-site. Edge computing enables instant analysis at the location, making it suitable for healthcare settings that are constantly evolving. However, this approach has not been tested in real clinical scenarios where unstable networks, diverse devices, and inconsistent quality of data could affect its performance. It also assumes a constant connection, which may not always be possible, and the data set might lack uncommon overdose situations, limiting generalization. Also, we saw that as the number of clients increased, performance got worse. This was more noticeable in non-IID environments. What this shows is that there are problems with scalability in different settings. To address these issues, future work will look into grouping clients based on how similar their data distribution is, creating federated learning models tailored to specific needs, and developing flexible aggregation methods intended to improve stability and performance when dealing with large deployments or non-IID situations.

Though the suggested RAFL-Fed framework shows good outcomes for observing and modifying regional citrate anticoagulation with federated learning, there are various paths left to explore in future research and development. One important step is validating these findings in real-world clinical situations by partnering with healthcare institutions. This will help us evaluate its performance under actual ICU conditions and make necessary adjustments based on doctors’ feedback. Merging this system with electronic health records (EHRs) can improve usability as it allows automatic data entry and instant dosing suggestions within current clinical setups. Further, creating customized anticoagulation models considering unique patient details like age, co-existing ailments, and liver performance can enhance the accuracy of treatment. The structure could be expanded to cover other essential care procedures such as insulin administration, fluid regulation, and antibiotic treatment, where adaptive decision support with confidentiality is equally crucial. Lastly, upcoming studies may concentrate on refining federated learning approaches by using stronger collection methods, adjusting well to client changes, and improving privacy-protective techniques like differential privacy and secure multiparty computation for safe, widespread deployment in health care settings.

Data availability

The raw data presented in this study are openly available in the MIMIC-IV database at PhysioNet (https://doi.org/10.13026/s6n6-xd98), reference number 499. The processed data is available in the GitHub repository https://github.com/malisaroj/crrt.

Materials availability

Not applicable.

Code availability

The code implementations, including instructions for reproducing the results are made publicly available on GitHub under the following repository: https://github.com/malisaroj/crrt.

References

Pace P, Aloi G, Gravina R, Caliciuri G, Fortino G, Liotta A. An edge-based architecture to support efficient applications for healthcare industry 4.0. IEEE Trans Ind Inf. 2018;15(1):481–89. https://doi.org/10.1109/TII.2018.2843169.

Rieke N, Hancox J, Li W, Milletari F, Roth HR, Albarqouni S, Bakas S, Galtier MN, Landman BA, Maier-Hein K, et al. The future of digital health with federated learning. npj Digit Med. 2020;3(1):119. https://doi.org/10.1038/s41746-020-00323-1.

Uchino S, Bellomo R, Morimatsu H, Morgera S, Schetz M, Tan I, Bouman C, Macedo E, Gibney N, Tolwani A, et al. Continuous renal replacement therapy: A worldwide practice survey. The beginning and ending supportive therapy for the kidney (b.e.s.t. kidney) investigators. Intensive Care Med. 2007;33(9):1563–70.

Joannidis M, Oudemans-van Straaten HM. Clinical review: Patency of the circuit in continuous renal replacement therapy. Crit Care. 2007;11(4):218.

Gulla K, Sachdev A, Anand K. Continuous renal replacement therapy (crrt) in critically ill children—experience of a tertiary care centre in a developing country. Pediatr Crit Care Med. 2014;15(4_suppl):40.

Zhang W, Bai M, Yu Y, et al. Safety and efficacy of regional citrate anticoagulation for continuous renal replacement therapy in liver failure patients: A systematic review and meta-analysis. Crit Care. 2019;23(1):22. https://doi.org/10.1186/s13054-019-2317-9.

Zhang Z, Hongying N. Efficacy and safety of regional citrate anticoagulation in critically ill patients undergoing continuous renal replacement therapy. Intensive Care Med. 2012;38(1):20–28. https://doi.org/10.1007/s00134-011-2438-3.

Mehta RL, McDonald BR, Aguilar MM, Ward DM. Regional citrate anticoagulation for continuous arteriovenous hemodialysis in critically ill patients. Kidney Int. 1990;38(5):976–81.

Zarbock A, Küllmar M, Kindgen-Milles D, Wempe C, Gerss J, Brandenburger T, Dimski T, Tyczynski B, Jahn M, Mülling N. Effect of regional citrate anticoagulation vs systemic heparin anticoagulation during continuous kidney replacement therapy on dialysis filter life span and mortality among critically ill patients with acute kidney injury: A randomized clinical trial. JAMA. 2020;324(16):1629–39.

Oudemans-van Straaten HM, Kellum JA, Bellomo R. Clinical review: Anticoagulation for continuous renal replacement therapy-heparin or citrate? Crit Care. 2011;15(1):202.

Khwaja A. Kdigo clinical practice guidelines for acute kidney injury. Nephron Clin Pract. 2012;120(4):c179–84.

Liu S-Y, Sheng-Yong X, Yin L, Yang T, Jin K, Zhang Q-B, Sun F, Tan D-Y, Xin T-Y, Chen Y-G, et al. Management of regional citrate anticoagulation for continuous renal replacement therapy: Guideline recommendations from Chinese emergency medical doctor consensus. Mil Med Res. 2023;10(1):23.

Pistolesi V, Morabito S, Pota V, Valente F, Di Mario F, Fiaccadori E, Grasselli G, Brienza N, Cantaluppi V, De Rosa S, et al. Regional citrate anticoagulation (rca) in critically ill patients undergoing renal replacement therapy (rrt), expert opinion from the siaarti-sin joint commission. J Appl Psychol Anesthesia, Analgesia and Critical Care. 2023;3(1):7.

Fisher R, Moore GW, Mitchell MJ, Dai L, Crichton S, Lumlertgul N, Ostermann M. Effects of regional citrate anticoagulation on thrombin generation, fibrinolysis and platelet function in critically ill patients receiving continuous renal replacement therapy for acute kidney injury: A prospective study. Ann Intensive Care. 2022;12(1):29.

Chen H, Yingying M, Hong N, Wang H, Longxiang S, Liu C, Jie H, Jiang H, Long Y, Zhu W. Early warning of citric acid overdose and timely adjustment of regional citrate anticoagulation based on machine learning methods. BMC Med Inform Decis Mak. 2021;21:1–11.

Harmony HT, Kwan AM, Leung NY, Chan WM, Ngai CW, Wong AS, Tsai PN, Tammy SM, Yam I, Pauline Yeung N, et al. A prospective study on serum citrate levels and clinical correlations in patients receiving regional citrate anticoagulation. Clin Kidney J. 2023;16(2):285–92.

Mishra RC, Sodhi K, Prakash KC, Tyagi N, Chanchalani G, Annigeri RA, Govil D, Savio RD, Subbarayan B, Arora N, et al. Isccm guidelines on acute kidney injury and renal replacement therapy. Indian J Crit Care Med: Peer-Reviewed, off Publ of Indian Soc of Crit Care Med. 2022;26(Suppl 2):S13.

Huang C, Gengchen X, Chen S, Zhou W, Eddie YN, Albuquerque VHCD. An improved federated learning approach enhanced internet of health things framework for private decentralized distributed data. Inf Sci. 2022;614:138–52.

Kong L, Tan J, Huang J, Chen G, Wang S, Jin X, Zeng P, Khan M, Das SK. Edge-computing-driven internet of things: a survey. ACM Comput Surv. 2022;55(8):1–41.

Mulimani MS, Rachh RR. Edge computing in healthcare systems. Deep Learn and Edge Comput Sol for High Perform Comput. 2021;63–100.

Banabilah S, Aloqaily M, Alsayed E, Malik N, Jararweh Y. Federated learning review: Fundamentals, enabling technologies, and future applications. Inf Process & Manag. 2022;59(6):103061.

Abreha HG, Hayajneh M, Serhani MA. Federated learning in edge computing: A systematic survey. Sensors. 2022;22(2):450.

Zhang C, Xie Y, Bai H, Bin Y, Weihong L, Gao Y. A survey on federated learning. Knowl-Based Syst. 2021;216:106775.

McMahan B, Moore E, Ramage D, Hampson S, Arcas. BA. Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. PMLR; 2017. p. 1273–82.

Brisimi TS, Chen R, Mela T, Olshevsky A, Paschalidis IC, Shi W. Federated learning of predictive models from federated electronic health records. Int J Med Inf. 2018;112:59–67.

Ahmed S, Grønli T-M, Lakhan A, Chen Y, Liang G. A reinforcement federated learning based strategy for urinary disease dataset processing. Comput Biol Med. 2023;163:107210.

Bukhari SMS, Zafar MH, Houran MA, Moosavi SKR, Mansoor M, Muaaz M, Sanfilippo F. Secure and privacy-preserving intrusion detection in wireless sensor networks: Federated learning with scnn-bi-lstm for enhanced reliability. Ad Hoc Networks. 2024;155:103407. https://doi.org/10.1016/j.adhoc.

Wang W, Li X, Qiu X, Zhang X, Brusic V, Zhao J. A privacy preserving framework for federated learning in smart healthcare systems. Inf Process & Manag. 2023;60(1):103167. https://doi.org/10.1016/j.ipm.2022.103167.

Sahu AK, Tian L, Sanjabi M, Zaheer M, Talwalkar A, Smith V. Federated optimization in heterogeneous networks. arXiv: Learning 2018.

Jaladanki SK, Vaid A, Sawant AS, Xu J, Shah K, Dellepiane S, Paranjpe I, Chan L, Kovatch P, Charney AW, Wang F, Glicksberg BS, Singh K, Nadkarni GN. Development of a federated learning approach to predict acute kidney injury in adult hospitalized patients with covid-19 in New York City. medRxiv [Preprint. 2021 Jul. https://doi.org/10.1101/2021.07.25.21261105. PMID: 34341802, PMCID: PMC8328073.

Rahmani AM, Gia TN, Negash B, Anzanpour A, Azimi I, Jiang M, Liljeberg P. Exploiting smart e-health gateways at the edge of healthcare internet-of-things: a fog computing approach. Future Gener Comput Syst. 2018;78:641–58. https://doi.org/10.1016/j.future.2017.02.014.

Khan M, Wang H, Ngueilbaye A, Elfatyany A. End-to-end multivariate time series classification via hybrid deep learning architectures. Pers Ubiquitous Comput. 2023;27(2):177–91.

Dang TK, Lan X, Weng J, Feng M. Federated learning for electronic health records. ACM Trans Intell Syst Technol (TIST). 2022;13(5):1–17.

Huang C-T, Wang T-J, Kuo L-K, Tsai M-J, Cia C-T, Chiang D-H, Chang P-J, Chong I-W, Tsai Y-S, Chu Y-C, et al. Federated machine learning for predicting acute kidney injury in critically ill patients: A multicenter study in Taiwan. Health Inf Sci Syst. 2023;11(1):48.

Pan W, Zhenxing X, Rajendran S, Wang F. An adaptive federated learning framework for clinical risk prediction with electronic health records from multiple hospitals. Patterns. 2024;5(1).

Johnson AE, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, Pollard TJ, Hao S, Moody B, Gow B, et al. Mimic-iv, a freely accessible electronic health record dataset. Sci Data. 2023;10(1):1.