FedICU: a federated learning model for reducing the medication prescription errors in intensive care units

Abstract

Patients in the Intensive care unit need remarkable observation. This unit consists of people who are critically ill and may tend to lose their lives anytime. Healthcare professionals in critical care tend to commit prescription errors for many reasons. Since patients in intensive care are severely ill and have complicated health issues, mistakes while prescribing medicines can have serious repercussions. In this study, a federated learning model is simulated to reduce the mistakes while prescribing medicines in the intensive care unit, which provides an opportunity for many hospitals to collaborate, keeping their data local to themselves. Local training is performed with Logistic regression, Simple neural network, and Multilayer perceptron in which simple neural network achieves the highest accuracy of 95%. Model weights transferred to a federated server may be vulnerable to data and model poisoning attacks, eavesdropping, and model inversion attacks. So, model weights are encrypted using Paillier homomorphic encryption (PHE), achieving a model accuracy of 93.26% for a key size of 2048. With key size, the effect of encryption and decryption time is observed. The model is also applied with differential privacy, which achieved an accuracy of 94.24% when c = 0.5 and sigma = 0.05. Thus, this privacy-preserving federated learning model can be used to reduce drug prescription errors in critical care.

Keywords:

• Federated learning
• medication prescription error
• intensive care unit (ICU)
• Paillier homomorphic encryption
• differential privacy

REVIEWING EDITORS:

• Chen, Chang Gung University, Taiwan; Marko Robnik-Šikonja, University of Ljubljani, Slovenia

SUBJECTS:

• Artificial Intelligence
• Health & Society; Medicine

1. Introduction

The intensive care unit is one of the critical areas of a hospital where patients struggle between life and death. The patients in these units are highly ill and frequently need complicated treatment plans. Adverse drug events in ICUs can have severe effects (Kane-Gill et al., 2010). Typical reasons for drug mistakes in critical care include incorrect pharmaceutical labeling or storage, a breakdown in communication between patients and healthcare professionals, confusing medicine names or packaging that seems similar, wrong medicine delivery or dosage, and inability to access a patient’s medical history. If there is clear communication, cooperation, and teamwork among the healthcare staff, it is possible to guarantee the most significant outcomes for the patients in the ICU. Any small mistake in treating the patients will lead to severe conditions that can affect a patient’s life. Hence, ICU patients need close observation and assistance from medical experts, including doctors, nurses, and other specialists. The doctors and nurses taking care of them undergo stressful situations. They play an essential role in making the right decisions about the line of treatment concerning the severity of the patient’s condition. Since the patients are critically ill and need close monitoring and prompt attention to any variations in their condition, the ICU is a high-stress and high-risk setting (Kumar et al., 2022).

Many studies were conducted to analyze the reasons causing medication errors in the ICU of neonates (Shawahna et al., 2022) and adults (Mohanna et al., 2022). If integrated with ICU, artificial intelligence can be a potential solution to deal with drug-related issues (Choudhury & Urena, 2022). In literature, artificial intelligence was used to recognize and describe their viewpoints on avoiding medication-related occurrences that cause patients to suffer grave or little injury. In addition to showcasing the usefulness of an automated AI program for analyzing free text datasets, the study attempts to identify the most significant risk management areas associated with these accidents (Härkänen et al., 2021).

Lehmann and Kim (2005) gives many approaches for preventing medication errors, but they must be manually done. If machine learning is used for prediction (Dos Santos et al., 2018), outliers prescriptions can be detected. However, the study does not address other errors like wrong drug selection. Also, It needs patient data to train the model and can benefit only that organization. Medical data is viewed as sensitive information that, if disclosed, could lead to the patient’s genuine and severe harm. This issue can be addressed with federated machine learning. This approach uses a client-server or client-edge cloud (Liu et al., 2020) concept with fair resource allocation (Li et al., 2019) in which the model is trained locally at the client, and only the model weights are sent to the server. By this, the sensitivity of the data is maintained, leading to a generalized model that benefits all the clients taking part in the federated learning.

In this study, a federated learning model is simulated and trained on the prescription file, which can identify potential medication prescribing and administration errors, such as drug interactions or incorrect dosages. To make it privacy preserving (Śmietanka et al., 2020) protect the gradients transferred between the client and the server, Paillier encryption (Fazio et al., 2017) is used, which involves encryption and decryption, giving an accuracy of 93.26%. The model is also tested by applying differential privacy, which achieves an accuracy of 94.24% if the noise parameters are properly tuned. There has to be a trade-off (Paragliola, 2022) between accuracy and privacy. This model can assist healthcare providers in prescribing decisions and can be used to improve and adapt over time. The model proposed in this paper will function as a support system for the physicians and nurses working in the intensive care unit to prescribe the drug with the proper dosage and frequency.

1.1. The contributions of this study are as follows

1. A Federated learning model FedICU is simulated to reduce medication prescription errors in critical care units.

2. Training is performed with different models like logistic regression, simple neural network, multilayer perception, and performance is observed.

3. Homomorphic encryption and differential privacy techniques are applied to protect the model from various attacks, and performance is observed.

In the second half of the paper, earlier work is discussed. The third section is a discussion about an overview of the model. The fourth section discusses the steps of the intended work. In the fifth section, the findings of numerous experiments are discussed. The final portion presents the drawbacks of the study, conclusion, and references.

2. Related work

In literature, recent technologies like Blockchain (Lakhan et al., 2023) and Deep learning (Al-Fahdawi et al., 2024) have shown tremendous effects in medical diagnosis (Ku et al., 2022). Machine learning can be used to assist patients in taking medicines. A study suggests a medication reminder system that helps patients, taking into account their unique physical and/or mental limitations, with their at-home treatment (Naeem et al., 2019). The system communicates with patients using a range of messages (text, visual, and audio) and adjusts to meet their needs. The patient can be guided to the appropriate medication box by choosing the appropriate message mode. A method was proposed for teaching a software agent to manage risks in medical software systems based on reinforcement learning (RL) (Paragliola et al., 2018). The aim of the RL agent is to keep patients out of harmful and undesired states while ensuring they get to a safe state or leave as soon as possible. A study proposed a risk management strategy for remote agile software projects using multi-agent reinforcement learning (Adel et al., 2021). Also, federated learning is used in the field of healthcare (Paragliola, 2023) and has been beneficial in predicting diseases like cancer (Beguier et al., 2021). A federated learning combined CNN-LSTM framework for identifying Spectrum Disorder (ASD) in children was presented by Abdullah Lakhan et al. (2023). It discusses how federated learning, IoT, cloud, and fog computing are used in healthcare to identify ASD in kids. Federated learning was also used to predict COVID-19 (Feki et al., 2021; Zhang et al., 2021).

Table 1 compares various existing federated learning models with the proposed work.

Table 1. Comparison with state of art.

Reference	Motivation	Contribution	Novelty	Dataset Used
Liu et al. (2018)	To improve healthcare outcomes and precision medicine	FADL can be used to predict patient mortality without compromising data privacy and security	Federated-Autonomous Deep Learning (FADL) for predicting patient mortality	eICU
Hwang et al. (2023)	To provide a practical guide for adopting Federated Learning in the medical domain	To evaluate the performance of FL on structured, image, and signal datasets	Empirical benchmarks, evaluation of FL algorithms	eICU
Rajendran et al. (2023)	Compare FL models’ performance with that of local and pooled models, and to comprehend how data heterogeneity impacts FL models’ performance	Methodology for predicting the risk of sepsis and acute kidney damage (AKI) in the intensive care unit	”unique importance score” (UIS) to examine the relationship between heterogeneous features	EHR from different hospitals
Manni et al. (2022)	To solve the issues with security, privacy, and data access in the medical sector	Predicting hemodynamic instability in ICU patients; comparing performance of decision tree models embedded in FL and centralized setups	Addressing security and privacy issues in the medical sector using FL	Philips eICU database
Mehrjou et al. (2022)	To investigate federated learning’s (FL) performance in multi-center critical care	Predicting the survival of each ICU stay	determine appropriate FL conditions on clinical data, which serves as a foundation for trustworthy classification algorithms	eICU dataset
FedICU (proposed model)	Increasing death in ICU due to medication errors	FL model to reduce drug errors in ICU along with HE and DP	Federated learning for reducing prescription errors with homomorphic encryption	MIMIC-IV

The existing work related to this study has been discussed in two sections.

2.1. Medication errors

Many studies were conducted in the literature to highlight the causes and prevention of drug-related errors. There was a review conducted by O’Shea (1999) which discusses the causes of medication errors in nursing practice. It analyses numerous studies that have linked pharmaceutical errors to interruptions, inadequate arithmetic abilities, incomplete or ambiguous prescriptions, and a lack of communication among healthcare professionals. The research highlights the value of a multidisciplinary approach in avoiding medication errors and offers suggestions for preventative measures in clinical settings. Similarly, there was an analysis conducted on the frequency and prevalence of intravenous drug mistakes in the UK (Sutherland et al., 2020). Incorrect doses and incorrect drugs were shown to be the most frequent types of intravenous medication errors. The analysis also discovered that hospitals experience more errors than other settings. The authors suggest several tactics for lowering the prevalence of intravenous medication errors, including enhancing training and communication, utilizing technology to assist with medication administration, and implementing set norms and recommendations. To decrease prescription medication errors, a clinical decision assistance system based on evidence-based treatment is discussed in this study by Lohani and Mukhopadhyay (2017) based on the Hive system. The necessary medicine is recommended based on the evaluation of the patient’s past medical history and recent evidence. Both patients and doctors can use this technique to learn more about a specific medicine. There was a study conducted by Ayhan et al. (2022) to prevent medicine-related issues. The study categorized various reasons for drug-related problems in the intensive care unit. Dionisi et al. (2022) proposed a study to pinpoint the key tactics and measures for avoiding drug errors in critical care units. The databases used for the search were as follows: Scopus, PsycInfo, CINAHL, PsychoInfo, and Embase. The findings highlight the need for a coordinated set of interventions to protect the system from harm and lessen the consequences of errors. A clinical decision support system designed by Hashemi et al. (2022) successfully reduced drug-related errors in the ICU. This study sought to determine how this CDSS affected the types and frequency of protocol deviations, which measure prescription mistakes, in a medical ICU. Dehghan Nayeri et al. (2021) proposed an approach to assess how a risk management program affects the frequency of adverse drug events among nurses working in intensive care units. The outcomes demonstrated that risk management program deployment successfully lowered nurses’ medication mistakes. Nurses were advised to implement a risk management program to promote safe drug prescriptions and get safe and effective nursing care. In all these studies, various methods are mentioned to reduce medication errors.

2.2. Federated learning for ICU

Regarding federated learning for ICU, for forecasting clinical results in patients with COVID-19, Ittai Dayan et al. (2021) provided a model named EXAM. The EXAM model was trained retrospectively using data from previous COVID-19 patients, and researchers already had actual measures of how much oxygen should be administered to a patient. Trung Kien Dang et al. (2020) have proposed a federated learning prediction model for ICU-in-hospital mortality. The findings suggest that federated learning can produce similar outcomes to centralized learning, with the added benefit of not having to share the dataset. It compares the performance of FedAvg and FedProx, two extensively used aggregation algorithms in federated learning. Deng et al. (2023) explored the difficulties in implementing machine learning in healthcare organizations and suggested a cutting-edge approach called Personalised Federated Learning (PFL) to encourage ML in numerous geographically dispersed healthcare organizations. The research provides a thorough design of the proposed PFL scheme and assesses how well it performs in activities that predict in-hospital mortality. According to the study, PFL beats other approaches regarding prediction accuracy and overall convergence rate, which compares PFL’s performance to that of other FL methods.

Li Huang et al. (2019) have suggested a distributed machine learning system that uses electronic medical records to forecast death rate and ICU stay time (EMR). The K-means method is used to cluster patients, improving federated learning efficiency. The suggested approach uses community-based federated learning (CBFL), assessed using three metrics: ROC AUC, PR AUC, and ROC AUC. The system was designed to analyze non-IID and IID datasets of Intensive care patients to predict death and how long patients will stay in the hospital. A method for federated learning to predict death in intensive care units was proposed by Mondrejevski et al. (2022). Lab results and vital signs were extracted as multivariate time-series data from the MIMIC-III database to compare the forecast accuracy of four different classification techniques (FRNN, LSTM, GRU, and 1DCNN). Reports also consist of window durations (8h, 16h, 24h, and 48h) and the total number of clients (2, 4, and 8). The results demonstrate that Central Machine Learning and Federated training are alike regarding AUPRC and F1-score. Also, Korbinian Randl et al. (2023) proposed a study on early prediction of ICU death with Deep Federated Learning. It explores the capacity of Federated Learning to detect the risk of ICU death rate at an early stage while respecting privacy regulations. The literature shows much work has been done using federated learning in intensive care. Still, to our knowledge, no federated learning model has been developed to reduce medication prescription errors, which becomes the motivation.

3. System overview

Given a set of training ICU dataset having n samples ($x_i,$ $y_i$), $1\le i\le n.$ Let $P_1,$ $P_2,$ $P_3$……$P_k$ denote the different ICU’s involved in federated learning and n is the total number of instances considered. The main aim of federated learning is to minimize the loss function f(w) given whole data ${\sum }_{k=1}^k{n}_k$ trainable federated learning weights with d parameters $wϵ$ $R^d.$ (1) $$f(w)=\sum _{k=1}^k\frac{n_k}{n}{F}_k(w)$$(1) where (2) $$F_k(w)=\frac{1}{n_k}\sum _{iϵP_k}{f}_i(w)$$(2)

Initially, local training of the data takes place at the individual clients using an algorithm that generates model weights.

Let $\Delta w^1,$ $\Delta w^2,$ $\Delta w^3,$ …. $\Delta w^k$ represent the weights generated by each of the individual ICU’s after training with an algorithm. These weights are encrypted with additive homomorphic encryption or differential privacy and sent to the federated server. After receiving the model weights from each ICU, the server averages them to form a global weight given by Equation (2)(2) $$F_k(w)=\frac{1}{n_k}\sum _{iϵP_k}{f}_i(w)$$(2) . (3) $$w_{t+1}\leftarrow \sum _{k=1}^K\frac{n_k}{n}{\mathrm{w}}_{t+1}^k$$(3)

The global weight $w_{t+1}$ is sent back to the hospitals and is decrypted at each hospital, preserving privacy between the hospital and the server. This is one communication round of federated learning. This is repeated in many rounds until the desired accuracy is achieved. At each round, learning occurs, improving the model’s accuracy and minimizing the loss. The overview of FedICU is shown in Figure 1.

Figure 1. Overview of FedICU.

Algorithm 1.

The FedICU Algorithm

Result: Generalized model for ICU

1: initialize $w_o$;

2: for each communication round do

3:  for each client 1,2,3…k do

4:   download the initial model w from server $w^k\leftarrow w$;

5:    Perform local training at each client;

6:    $\mathit{\text{client}}\to \Delta w^k;$

7:    Secure the model weights sent to server;

8:    Server performs aggregation $w_{t+1}\leftarrow {\sum }_{k=1}^K\frac{n_k}{n}{\mathrm{w}}_{t+1}^k$;

9:   Decrypt and update $w_{t+1}$ at clients

10:  end for

11: end for

3.1. Additive homomorphic encryption

The model gradients transmitted from each of the ICUs to the server may be prone to various kinds of attacks (Lyu et al., 2020). The attacker may be able to interpret the gradients to recreate an original model out of it. Hence, to provide security to the model gradients, homomorphic encryption or Paillier encryption (Park & Lim, 2022; Yang et al., 2020; Zhang et al., 2020) is applied. It is a public key encryption technique that has the advantage of performing operations on the encrypted data. With additive homomorphic encryption, the gradients can be sent to the server for averaging without decryption, and the data is protected. The data is decrypted only by the clients after receiving the global model. At the same time, the model is being learned from it. This ensures that only the finalized model parameters are exposed to the server and that the raw data is kept private and secure. Additive Homomorphic Encryption can perform operations on the encrypted data. This is possible because of the properties of homomorphic encryption that allow the addition of two ciphertexts given in Equation (4)(4) $$D_{\mathit{priv}}\left(E_{\mathit{pub}}\right(m_1\left)E_{\mathit{pub}}\right(m_2\left)\mathit{\text{mod}}{n}^2\right)=m_1+m_2\mathit{\text{modn}}$$(4) , which says the result of the multiplication of two ciphertexts will lead to the sum of their plaintexts when decrypted. (4) $$D_{\mathit{priv}}\left(E_{\mathit{pub}}\right(m_1\left)E_{\mathit{pub}}\right(m_2\left)\mathit{\text{mod}}{n}^2\right)=m_1+m_2\mathit{\text{modn}}$$(4)

The flowchart for the steps performed in Paillier encryption is shown in Figure 2.

Figure 2. Flowchart for Paillier cryptosystem.

The encryption and decryption processes can be computationally demanding; homomorphic encryption could also result in a computational overhead. Data is shared among several parties in federated learning, and encryption can be employed to protect the data and stop illegal access or breaches. Encryption can also be utilized to safeguard interparty communication during the training procedure.

3.2. Differential privacy

Differential privacy is applied in the federated learning system to secure the confidentiality of model weights being exchanged from client to server (Choudhury et al., 2019; El Ouadrhiri & Abdelhadi, 2022). To prevent the attacker from interpreting the original data, noise is attached to the model weights before transmitting them to the federated server. The $ϵ$ parameter, which regulates the degree of privacy protection, provides the base for the definition of differential privacy. A randomized algorithm A is said to be $ϵ$-differentially private if for any two datasets D and D’ that differ in at most one record, and for any subset S of the output space of the algorithm A, the following inequality holds: (5) $$\mathit{\text{Pr}}[A(D)\in S]\le \mathrm{\text{exp}}(ϵ)*\mathit{\text{Pr}}[A(D^{\prime })\in S]$$(5)

According to this inequality, the likelihood that algorithm A will generate output in S on dataset D is only marginally higher than the likelihood that it would generate the same output on a nearby dataset $D^{\prime },$ with a difference of no more than one record. Smaller parameter values, which define how much greater the likelihood can be, indicate more robust privacy protection. The probability density function (PDF) of the Gaussian distribution is given by Equation (6)(6) $$p(x)=1/\sigma \sqrt{2\pi }{e}^{\left(\frac{{(x-\mu )}^2}{2{\sigma }^2}\right)}$$(6) . (6) $$p(x)=1/\sigma \sqrt{2\pi }{e}^{\left(\frac{{(x-\mu )}^2}{2{\sigma }^2}\right)}$$(6) where x is the random variable, $\mu$ is the mean, and $\sigma$ is the standard deviation. Assuming there is a function f that accepts dataset D as input and returns a real-valued result to apply the Gaussian process. By incorporating Gaussian distribution noise, this function is made differentially private. The differentially private function $f^{\prime }(D)$ produces the following result: $f^{\prime }(D)=f(D)+\eta$ Where $\eta$ is a random variable drawn from a Gaussian distribution with mean 0 and standard deviation $\sigma .$ The standard deviation $\sigma$ is determined by the sensitivity of the function f and the desired level of privacy $ϵ,$ according to the following formula: $\sigma$= sensitivity(f) * sqrt(2 * ln(1.25 / $\delta$)) / $ϵ$ Where $\delta$ is a parameter that controls the probability that the output of the mechanism deviates from the actual value by more than a given threshold.

4. Methodology

The positivist paradigm is adopted by the proposed study of developing a federated machine learning model to reduce drug-related errors in intensive care units. The research method is quantitative since it involves numerical data that can be measured. Data collection is secondary since it is publicly available and is downloaded. The working of a simulated federated learning model is as follows.

1. The server sends the initial model to each ICU.

2. The data at each ICU is trained using a suitable algorithm. This study is tested with three algorithms: Logistic regression, simple neural network, and Multilayer perceptron, out of which a simple neural network is chosen for further experimentation.

3. Step 2 will generate model weights, which are encrypted and sent to the federated server.

4. After receiving weights from each ICU, the federated server performs averaging and generates global weight, which is sent back to each ICU. Then, the global weights are decrypted and applied to each of the ICUs. This is one round of federated learning shown in Figure 3. Learning is performed for several rounds until the desired accuracy is achieved.

Figure 3. One of federated learning.

The parameters used for the study are learning rate, which is a tuning parameter that establishes the step size at each iteration as the algorithm moves towards a loss function minimum, which is taken as 0.05. The number of local epochs used for local training is taken as 3. Number of clients considered are 15.

4.1. Dataset and preprocessing

In this paper, all the experiments are conducted using the publicly available dataset MIMIC-iv (Johnson et al., 2020), downloaded from physionet.org. Since it deals with drug-related problems, only the prescription file is considered. The prescription file contains information about the medicines for critically ill patients in intensive care. The entire file is checked for duplicates, and redundant rows are filtered. The file had 20 attributes; after discussing with doctors and nurses working in the ICU, seven prominent attributes were considered for experiments. The attributes include the name of the drug, the dosage of the drug, a unit of dosage, the amount of the drug to be given is taken as count, the unit of count, and the path through which the drug is given is taken as a route. The last attribute is the target, which is taken as a 0 for the right medications and 1 for the wrong ones. So, in total, there are 7 attributes and one label. The prescriptions dataset contained non-numeric values for attributes like drug name, dose unit, count unit, and route. These attributes are converted to numbers using the label encoding package of the sklearn library. Converting the attributes into a numerical representation so that machines can read them is known as label encoding. The snapshot of the dataset is shown in Figure 4.

Figure 4. Prescription dataset with attributes and label.

There are around 28,000 instances, of which around 25,000 are without errors and 2,000 with errors. It can be seen in Figure 5. 0 indicates those instances without errors, and 1 indicates those instances with errors.

Figure 5. Count of instances.

The correlation matrix shown in Figure 6 assesses the relation between two variables. Every element in the table is a coefficient of correlation where 1 is considered a strong relationship between variables, and 0 is considered a neutral relationship. It can be seen from Figure 6 that attributes are not much correlated. The highest correlation is between count and error, which is 0.65.

Figure 6. Correlation between features.

5. Results

A Federated learning model was simulated to reduce medication prescription errors considering 5 clients or hospitals. All the experiments were conducted on a jupyter notebook on a Windows 10 system with an i7 processor using Pytorch. Initially, local training is performed on each of the clients. The entire dataset is split as 80% for training and 20% for testing. The model performs local training with logistic regression, which is a base for neural networks. The base model of the logistic regression classifier is like a single-layer neural network. Then, the model is trained with other models, like a simple neural network with one hidden layer and a multilayer perceptron with two hidden layers for different epochs, and performance is observed. The Aggregation algorithm used is Federated Averaging. Figures 7 and 8 show the performance of the federated learning model with communication rounds against testing accuracy and training loss. It can be observed that the testing accuracy of the model increases with the number of rounds, and training loss decreases with the number of rounds. The simple neural network performs better than the other two models. So, a simple neural network is chosen for local training of the model.

Figure 7. Communication rounds vs testing accuracy.

Figure 8. Communication rounds vs training loss.

Table 2 shows the other performance metrics, like F1 score, Precision, and Recall, after training the model for 20 rounds using a simple neural network. Precision tells about how accurate the positively predicted instances are. High precision means that the model is likely to be accurate when it predicts the positive class. The recall is the proportion of correctly predicted positive outcomes to all positive events. Precision and recall are frequently inversely connected; the F1 score is especially helpful when trying to strike a balance between both.

Table 2. Performance of other metrics.

Type of data	F1 Score	Precision	Recall
Training	0.73	0.63	0.85
Testing	0.75	0.62	0.95

Table 3 shows the performance with varying numbers of clients. When more clients are added, it is observed that the testing accuracy decreases slightly.

Table 3. Performance varying number of cleints.

No of clients	Training accuracy (%)	Testing accuracy (%)	Training loss	Training time (s)
5	93.42	94.07	0.379	0.228
10	93.31	93.62	0.378	0.216
15	93.51	93.27	0.379	0.222
20	93.53	93.18	0.379	0.23

To secure the model weights sent to the federated server, weights are encrypted with additive homomorphic Encryption. The model is experimented with different key sizes, and performance is observed. Figure 9 is a graph that shows the varying Encryption and decryption times for various key lengths. As the key size increases, there is an increase in the encryption time. It is also observed that it takes more time to encrypt than decrypt.

Figure 9. Effect of Paillier encryption.

Table 4 shows the effect of applying Paillier homomorphic encryption, which takes training time, average training time, and accuracy into consideration with varying key sizes. Adding longer keys would increase more protection, but that adds computation overhead. There is a massive increase in the training time when the key size increases. However, the accuracy of the model is not affected by Encryption. The model gives accuracy similar to what it showed without applying encryption.

Table 4. Performance after applying Paillier encryption.

Key size (bits)	Training time (s)	Testing accuracy (%)	Training average loss
128	7.4	93.98	0.354
256	8.07	93.62	0.363
1024	113	92.56	0.357
2048	125.022	93.36	0.375

The model is also experimented with applying another privacy-preserving technique, differential privacy. The two parameters considered are C and $\sigma .$ C is the model clip parameter, and $\sigma$ is the Gaussian noise parameter. The Gaussian noise parameter controls the amount of random noise supplied to the output to achieve differential privacy. In contrast, the clip parameter restricts the range of the model’s output and serves as a type of data normalization. Figure 10 shows the testing accuracies for varying values of C and $\sigma .$ C = 0.5 and $\sigma$=0.05 can be considered the ideal value, after which the accuracy starts falling as their values increase.

Figure 10. Performance after applying differential privacy.

It is observed that C and $\sigma$ parameters harm accuracy. As the value of C and $\sigma$ are increased, the accuracy falls, and the training average loss also increases. $\sigma$ is the noise parameter, and its value must be appropriately chosen so that there is a fundamental trade-off between privacy and accuracy in differential privacy. Adding noise to the output of a function is necessary to protect individual privacy, but this noise can also reduce the accuracy of the analysis.

Table 5 shows the performance of other metrics like f1, precision, and recall after 20 rounds of training with 20 epochs with C = 0.5 and $\sigma$=0.05

Table 5. Performance of other metrics with DP.

Type of data	F1 score	Precision	Recall
Training	0.31	0.61	0.21
Testing	0.29	0.60	0.19

6. Discussion

The simulated federated learning model shows a good performance considering various parameters. The results are observed after applying Paillier encryption and differential privacy to preserve privacy. The impact of using Encryption on a federated learning model’s accuracy relies on several variables, including the Encryption employed, the volume of the encrypted data, and the computing power of the clients involved. Encryption generally increases processing overhead since it increases training time, encryption time, and decryption time. The other technique, differential privacy, requires carefully tuning the parameters. The noise should be added in such a way that there is no fall in accuracy.

7. Limitations of the proposed study

The implementation of this paper is done with Pytorch, a deep learning framework that is not a specific federated learning framework. There are many Federated learning frameworks like TFF, Leaf, Flower, and OpenFL on which the study can be implemented. The various clients or ICUs considered can be distributed on separate systems instead of simulating on a single system, which could add more value to results. Due to the additional computing work required to encrypt and decrypt the data, using encryption might cause latency to increase. Encryption can raise the computing burden and latency of the system, which may significantly affect the model’s performance.

8. Conclusion

The main goal of the study was to help the healthcare professionals working in ICU by developing a model that would reduce drug-related errors in intensive care units. This paper proposes FedICU, a federated learning model to address medication prescription errors in intensive care units. FedICU allows several hospitals to collaborate and keeps the data private without sharing it with other hospitals and the federated server. In this study, FedICU is simulated as a standalone system. The model is trained locally with different models; only the gradients are sent to the server. This is very beneficial because healthcare organizations may want to keep their data private. Also, to protect the model against various attacks, the model gradients are encrypted using Paillier encryption.

If this model is used in real-time, the participating ICUs need not share the data with one another. All the participating ICUs will have the advantage of getting an improvised global model from it. The Federated learning model, FedICU, takes a collaborative approach that allows all the ICUs to participate without transmitting the actual data to the server or each other. These models can enhance trust among healthcare professionals and patients. Federated learning works with continual learning as the local data updates the model. This is helpful in the ICU because new medications, treatment methods, and medical information evolve. The federated model can generalize better in different healthcare settings because it was trained on various datasets from different ICUs. This enhances the model’s capacity to accommodate changes in patient demographics, healthcare procedures, and other elements.

9. Future directions

This setup can be experimented with a real-time set-up as a future enhancement, considering ICUs from different hospitals, training locally in the hospitals, and sending gradients to the server in desired rounds. Also, the drug dosage calculation according to the patient’s height and weight is not done in this work, which is taken as a future enhancement.

Authors’ contributions

Vineetha Pais (primary author) has been responsible for experiments, writing manuscripts, conducting investigations, and creating figures. Dr. Santhosha Rao (second author) and Balachandra Muniyal (third author) were responsible for reviewing the paper and helping write the manuscript. Sheng Yun (fourth author) has been responsible for helping in conducting experiments.

Disclosure statement

The authors declare no conflict of interest.

Data availability statement

The publicly available dataset MIMIC-iv was used.