Hospital-acquired infections (HAIs) are a major concern nowadays, since they entail a big threat to society and an increase in healthcare costs. AI techniques show great performance in the development of efective systems to help in their control and prevention. However, many recent studies highlight the lack of available datasets for reproducing their experiments, claiming for more trustworthy medical AI models. Realistic data simulation is a valid approach for testing these models when data is publicly unavailable or when clinical data gathering is cumbersome or impossible. Main simulators often focus on implementing compartmental epidemiological models and contact networks for validating epidemiological hypotheses. However, very little attention is paid to hospital infrastructure (e.g. hospital building, policy, shifts, etc.) which plays a key role in the infection and outbreak processes. This paper proposes a novel approach for a simulation model of HAI spread, combining agent-based patient description, spatial-temporal constraints of the hospital settings, and microorganism behavior driven by epidemiological models.
Multidrug-resistant bacteria (MDR-bacteria) are bacteria that evolved and acquired resistance
to antimicrobial drugs. This resistance makes the treatment more complex, increasing the risk
of infection, its spread and mortality [
In this context, machine learning (ML) and deep learning (DL) techniques present an
opportunity to develop efective systems that can help in the clinical decision and planning process.
The development of these systems requires access to big volumes of high-quality data, both
for training and validating. However, several concerns might compromise the use of health
data in AI research, such as the quality of accessible data and the risk of bias, the protection of
individual privacy, or the individual loss of autonomy, among others [
Some approaches to preserving privacy are the anonymization or pseudo-anonymization of
data. This could encourage the data usage, though not without challenges (e.g. data triangulation)
[
This paper presents a simulation model to study the propagation of infections within a hospital. The core of the simulation is to study the movement of a population of patients inside a hospital, and to allow the analysis of the spread and outbreak of an infective disease in this population. Therefore, it consists in the combination of an agent-based model (micro-scale model), the dynamics of an epidemiological compartmental model (macro-scale model), and the policies and the physical structure of a healthcare environment.
Our proposal is based on a discrete event time model in order to simulate a bacterial infection spread among patients in a hospital setting. In each step of time, hospitalized patients are going to move around the hospital, and they are going to be involved in the transmission of a bacterial infection as well as in its potential outbreaks. To achieve this, the simulation model consists in 3 main components: the input parameters, the core of the simulator (i.e. micro-scale modeling via an agent-based model, macro-scale modeling with the use of a compartmental model, and spatial-temporal constraints of the hospital settings), and the simulation outputs. The framework of the simulation model is presented in Figure 1 and the components are explained below.
Compartmental models are modeling techniques often applied in the simulation of infectious
diseases. The population is assigned to labeled compartments and they can progress from one
to another. In these models, the population dynamics are well known [
Population are admitted at the hospital in state S, I, or NS. They can get infected while in state S. If this happens, they go to state E, which means that they are incubating the disease, but are not contagious. They are going to remain in state E a period of time that depends on the incubation duration of the infection. Once this period is over, they go to state I, thus they can infect others and contaminate the environment. If they survive the disease, they go to state R and, if not, they go to state D.
Agent-based approaches ease micro-scale simulations, describing individual and their
interactions. Unlike other approaches, like [
With the patient simulation proposed, we track their stays from the arrival to the hospital to their discharge. Each patient has a unique ID and a set of attributes that include the localization where they are, age, gender, length of stay (LOS) in the hospital, health state, incubation period, infection duration, and applied treatment. We have considered only adult patients and we track their movements around the hospital in each time step through the diferent areas and services.
Patients can interact with other patients and with the environment. Interactions between patients take place when they share a room. During these interactions, a sick patient can infect another with probability . Interactions with the environment happen when a patient has spent an amount of time in the same place, and they can infect it with probability . This probability is going to be higher if the infected patient has not started the treatment yet, and lower if they have been on it for more than 3 days. Finally, a contaminated place can infect a susceptible patient with probability .
Regarding the recovery, infected patients may have a quick recovery without treatment or a longer one with the need of treatment. Both have a probability of success and a duration. In case of non-recovery, a patient may die with probability . All of these probabilities and time periods depend on the modeled pathogen and are part of the parameters explained in Section 3.
The spatial distribution of the hospital plays a key role in the spread of infections. We have considered a two-story hospital, where we took into account the most likely places in which a hospitalized patient can become infected: emergency room (ER), operating rooms, radiology rooms, wards which contain several patient rooms each, and an intensive care unit (ICU). The ER, the ICU and each room can have a user-defined number of beds. Each bed and place have a unique ID and a state indicating whether they are contaminated by the infection or not.
The places that a patient can go to are divided into two types: temporary and indefinite. Temporary places are those in which a patient is going to spend a short period of time (e.g. radiology or surgery). Indefinite places are those where a patient can stay for a long period of time (e.g. a bed, the ICU). The patients movements are spatially and temporally constrained. In order to implement these constraints as realistic as possible, we have designed a series of rules following the suggestions of medical doctors: • Temporal constraint: in each step of the simulation, only a limited number of patients can move to each ward. • Spatial-temporal (ST) constraint: patients must have spent a minimum number of steps without having gone to a temporary place to go back. For example, if a patient has just undergone surgery, they will not go back into the operating room right away. • ST constraint: patients that have been in the ER or the ICU for a certain period of time can be transferred to a ward. • Spatial constraint: when a patient goes to a temporary place, in the next step they return to the same bed where they were before. • ST constraint: patients can change of bed in the same ward during 1 simulation step. • ST constraint: patients in a ward can change to another ward during 1 simulation step. • ST constraint: patients in a ward or the ER might be transferred to the ICU.
The simulation model receives input parameters with user-assigned values. These parameters can be classified into three types: population parameters, epidemic-model parameters and configuration parameters.
Population parameters are those that represent the population and the characteristics of the hospital (i.e. the occupancy rate, admission rate through the ER, number of beds, rooms and wards, etc.). Among these parameters, we also find the patient’s age and LOS distributions.
Epidemic-model parameters are those that configure the infection behavior and allow the
change from one state of health to another (i.e. probability of contamination, probability of
recovery, treatment duration, etc.). We obtained or calculated both these and the population
parameters from public access data and information from the literature. Those for which there
was not enough information to infer their distribution follow a triangular distribution defined
by a mean value that represents its mode, a minimum and a maximum value, based on [
Configuration parameters comprise the settings for each run of the simulator (i.e. the cleaning frequency, the time between the movements of a patient, etc.). We defined these parameters based on the hospital size, data published by hospitals, and information obtained from an expert.
When working with micro and macro-scale modeling, one advantage is the derivation of results with low and high level of abstraction. Two straightforward outputs are daily statistics of the processes under study, and a database of the patients during the time of the simulation. This database includes information on each patient and the places from the hospital where they have been to. Another output is the extraction of aggregated information of the infection: in this case, the number of patients in each health state per day is stored. By combining this with the localization of the patients in each step, it is possible to compute any epidemiological indicator that can be calculated with these data (e.g. prevalence, incidence density, etc.) for the diferent areas from the hospital.
To experiment with this simulator, we have configured a hospital with 212 beds and the input
parameters according to the Santa Lucia General University Hospital from Murcia, Spain. Each
step is 8 hours, since it is the duration of a standard work shift. For the infection, we have
chosen the Clostridium Dificile (CD) pathogen, which is the main cause of infectious diarrhea
in hospitalized patients [
Codella et al. [
Lee et al. [
Haber et al. [
This paper proposes an agent-based model coupled with the infection dynamics extracted from an epidemiological model. This model can be used as a generator of synthetic clinical data on MDR-bacteria infections within hospitals. The use of an agent-based model together with the role of the hospital topology in an infection spread can enhance the detection of spatial and temporal patterns to help in the monitoring and the decision-making process. This is thanks to a more precise study of the infection process and the consequences in hospitalized patients. The capacity of tracing patients at a low level and to also obtain aggregate results from them can play a key role and be a step forward in the creation of a more explainable AI and in the generation of higher-quality synthetic data. In future we plan to carry out a thorough evaluation of the model to ensure its correct implementation, clinical meaning and utility. Once we perform this evaluation, the simulator is going to be available in an open repository. This work was partially funded by the CONFAINCE project (Ref: PID2021-122194OB-I00) by MCIN/AEI/10.13039/501100011033 and by "ERDF A way of making Europe", by the "European Union" or by the "European Union NextGenerationEU/PRTR", and by the GRALENIA project (Ref: 2021/C005/00150055) supported by the Spanish Ministry of Economic Afairs and Digital Transformation, the Spanish Secretariat of State for Digitization and Articial Intelligence, Red.es and by the NextGenerationEU funding. This research is also partially funded by the FPI program grant (Ref: PRE2019-089806).