Keywords

1 Medical Problem and Background

Antimicrobials are medicines used for the treatment and prevention of infections; this includes antibiotics, antivirals, antifungals and antiparasitics. Antimicrobial resistance occurs when bacteria, viruses or others change over time and stop responding to said medicines, making the treatment difficult and increasing the risk of spread, severity of diseases and death [14]. Specifically, multidrug-resistant (MDR) bacteria are one of the critical threats to public health at the moment, as their spread is associated with an increase in the morbidity, mortality and healthcare costs [7].

MDR bacteria are primarily associated with hospital-acquired infections: in a hospital, a patient’s exposure or infection with an MDR bacteria can occur through contact with a carrier, exposure to contaminated environments, use of contaminated medical equipment or following the use of antimicrobial agents [13]. The control and prevention of infections caused by MDR bacteria is, therefore, a priority problem today. This control is divided in a series of relevant tasks, which are: early detection of infection outbreaks, communication to healthcare staff and monitoring of cases within a hospital to help prevent the spread [4]. However, the detection of MDR-bacteria outbreaks inside a hospital is a complex process that requires interweaving reports with temporal and spatial information. Computer-aided visualization techniques might play an important role to explain and understand data-driven decision making in the following years.

In the most recent literature, there is growing interest on the computer-based detection and notification of infections in hospital settings. We highlight the following:

  • Myal et al. [11] presented a network-based analysis of movement of patients carrying drug-resistant bacteria across several hospitals so as to detect disease transmissions.

  • Baumgartl et al. [3] presented a visual analytic approach to support the analysis of contacts between patients, transmission pathways, progression of the outbreak and patient timelines during hospitalization.

  • Arantes et al. [2] developed a system for monitoring occurrence trends of hospital-acquired infections using statistical process control charts.

In most cases this research focuses on the monitoring of specific infection scores, the detection of abnormal events and the development of tailored statistical models from a classical approach [1]. Some of them also focus on bacteria’s features and contact network models, but neglect the spatial characteristics of hospital buildings.

We have also identified that there is a lack of general reproducibility of the experiments due to (1) the absence of open-access data of this nature; (2) the clinical datasets from hospitals used are confidential, under the personal data regulations. From the scientific point of view, this fact could lead into two problems: (a) the difficulty to establish a fair analysis between previous and current computational models; and (b) the limited use of Machine Learning and Deep Learning methods for building, evaluating their performance and robustness, and their clinical validation.

Therefore, there is a need for new tools able to generate, analyze and visualize spatial-temporal data for MDR-bacteria infections to support decision-making by health personnel and to ease their work.

2 Hypothesis and Planned Approach

Considering all the factors mentioned, the research hypothesis of this PhD thesis is that spatial-temporal modeling and visualization techniques allow clinicians to increase their confidence and comprehension of AI-based epidemiological analysis and prediction models.

This PhD thesis aims to:

  • Identify current and most suitable visualization methods for explaining results in epidemiological analysis considering spatial and temporal dimensions.

  • Get high quality spatial-temporal dataset of hospital MDR-bacterial infection at our disposal. We will consider both approaches: real datasets from hospital settings and realistic simulated data.

  • Propose and design new visualization methods for explaining epidemiological indicators (e.g. incidence, prevalence, mortality) considering spatial and temporal dimensions. We plan to implement an interactive visual tool for the detection of outbreaks and endemics using traditional clinical approaches.

  • Develop new ML-based prediction models for hospital outbreak analysis and an extension of the interactive visual tool based on explainable and trustworthy principles.

3 Preliminary Results

During the first phase of this PhD project, the following results have been obtained:

3.1 Visualization Models for Epidemiology Analysis

We have carried out a detailed analysis of what has been done in the last years on the application of spatial-temporal visualization techniques on epidemiological data. To do this, we have conducted a systematic review following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [10] and we have searched for papers in various scientific databases. Following this, we have reviewed 1049 articles, of which 71 were deemed suitable for inclusion in the study. With this systematic review, we have put the efforts made during the last two decades into perspective, paying special attention to how epidemic measures, temporal and spatial information are displayed. We identified that there is an increasing development of visualization tools, since the need to display the data for a better and more efficient interpretation is more valued. We have inferred the current trends in the development of visualization programs and identified some gaps concerning the lack of a standard evaluation methodology, the study of individual epidemiological data and, more specifically, of diseases acquired by patients in hospitals as the result of an epidemic. This research is currently under evaluation at [8].

3.2 Agent-based Simulation of Clostridium Difficile Infection

A second problem addressed is the need of high quality spatial-temporal data of hospital infection spread. To this end, we have proposed and implemented a simulation model and its simulator. The objective of the simulation model is to generate clinically realistic spatial-temporal data of infection outbreaks within hospitals. We adopted an agent-oriented approach combining spatial-temporal logic of the agents together with clinical semantic obtained from external epidemiological models. The main objectives of the simulation are the movements of patients through a hospital, and the infection and transmission of an MDR-bacteria disease. For this reason, our model consists of one type of agent: patients. To study the evolution of patient’s infection, each one is associated with a health state. This state is an abstraction of their situation at any given time so as to adapt it to the SEIRD epidemic model (Fig. 1).

Agents can interact with each other if they share a room, and they can also interact with the environment. A contagion can occur in these interactions if one of the patients is infected or if the environment is contaminated.

For our hospital model, we have considered a two-story hospital, where we take into account the most likely areas in which a hospitalized patient can become infected: emergency room, radiology rooms, operating rooms, ward rooms and the Intensive Care Unit.

The simulations follow a discrete-time step-based approach. In each step of the simulation, patients can move from their current location to an available place in the hospital. Our proposal also simulates admissions and discharges. In addition to this, the health status of patients can change and the different places within the hospital can be contaminated or disinfected.

Our preliminary experiments have been conducted considering Clostridium Difficile infection (CDI) since it is the main cause of infectious diarrhea in hospitalized patients. Both its incidence and the severity of clinical manifestations have increased notoriously in recent years [9]. In order to obtain a fair comparative with state-of-the-art proposals, we have based part of our methodology on [5, 6, 12]. The results from these experiments are outlined in Fig. 2.

Fig. 1.
figure 1

SEIRD epidemic model.

Full size image
Fig. 2.
figure 2

Simulator results.

Full size image

4 Current State

This thesis project started in October 2020 and we are reaching the halfway point of work. The current pandemic scenario has slowed down certain activities, particularly advice from the medical team.

We are waiting to publish the systematic review and we expect to have results from the simulator in the coming weeks. We are starting to run some tests to visualize the data output that we are getting from the simulator.