Harnessing of Artificial Intelligence for the Diagnosis and Prevention of Hospital-Acquired Infections: A Systematic Review

Harnessing of Artificial Intelligence for the Diagnosis and Prevention of Hospital-Acquired Infections: A Systematic Review

Introduction

Healthcare-associated infections (HAIs) are a leading global public health concern, representing the most common adverse events in healthcare. Effective surveillance is crucial for HAI prevention and control, but traditional methods are often costly and labor-intensive. Artificial intelligence (AI) and machine learning (ML) present opportunities to enhance HAI surveillance, improve patient risk stratification, and predict and prevent infections.

AI applications have been explored in clinical laboratory testing, imaging, and antimicrobial resistance profiling. The COVID-19 pandemic highlighted the need for automation and ML in infectious disease management due to increased patient volumes and diagnostic workload. AI can help interpret laboratory results, especially amidst shortages of trained personnel, thereby improving workflow in clinical microbiology.

With over 2.6 million new HAI cases occurring annually in Europe, leveraging AI for forecasting infections in hospitals and intensive care units (ICUs) is essential. While AI has become a valuable tool for organizing and implementing effective infection control programs, challenges remain.

This review aims to summarize the current literature on AI applications in HAI diagnosis and prevention, discussing advancements, potential applications, and future directions in this emerging field. Following PRISMA guidelines, 162 relevant articles were screened from databases including PubMed and Scopus, providing insights into AI’s role in improving infection practice.

Materials and Methods

Eligible studies assessed AI applications in infectious disease surveillance, diagnosis, and prevention until November 2023. Exclusion criteria included commentaries, non-English publications, and research with significant bias (e.g., theses). Reference lists of selected articles were screened for additional relevant studies.

The titles and abstracts of all obtained studies were reviewed by two independent reviewers, who resolved discrepancies through consensus.

Data extracted included authors, publication year, sample size, study design, techniques used, and performance measures (e.g., AUROC, accuracy). The risk of bias in the studies was assessed by the same two reviewers, with any disagreements discussed with a third researcher.

Results

A total of 549 articles were found in the aforementioned databases after the searches, of which 194 were excluded as they were duplicates or were marked as ineligible. The initial screening of the remaining 283 articles by title and abstract resulted in the exclusion of 24 that did not meet the inclusion criteria of the current review. Among the remaining 259 articles, 8 of them were excluded by two reviewers. A total of 251 articles were analyzed and 89 were excluded for different reasons such as being unrelated and off-topic, not using AI algorithms, being not in English language and full-text not being available. Finally, 162 articles met the inclusion criteria and were included in the current systematic review. Figure 1 illustrated the PRISMA flow diagram for the search process.

PRISMA-flow-diagram

Figure 1. PRISMA flow diagram for the search process

AI Applications in Microbiology

The gold standard for COVID-19 diagnosis is the detection of SARS-CoV-2 using reverse transcriptase polymerase chain reaction (RT-PCR). The pandemic led to a surge in PCR testing, prompting the development of AI models to enhance diagnostic efficiency.

Several AI-driven solutions have emerged:

  • qPCRdeepNet: A deep learning model that analyzes RT-PCR fluorescent readings to improve specificity and reduce false positives.
  • LSTM Models: These models utilize fluorescence data from RT-PCR cycles, integrating patient clinical data to expedite diagnosis.
  • ML Algorithms: Various machine learning (ML) models have been developed to analyze RT-PCR curves and predict SARS-CoV-2 positivity based on blood test results, achieving accuracies between 81% to 86%.

AI has also been employed in imaging diagnostics, particularly with chest CT and X-ray:

  • AI systems have shown high diagnostic performance in categorizing COVID-19 pneumonia and distinguishing it from other lung diseases.
  • A study reported an area under the curve (AUC) of 0.92 for AI integration of CT findings and clinical data.

Comprehensive systematic reviews highlight the effectiveness of AI in interpreting chest scans, with some models achieving sensitivities and specificities exceeding 90%. These AI applications have been crucial in optimizing testing processes and enhancing the rapid identification of COVID-19, ultimately aiding in infection control efforts during the pandemic.

Image Analysis—Bacterial, Viral, Fungal, Parasitic

Image analysis plays a crucial role in clinical microbiology diagnostics, enabling the examination of samples like Gram stains, blood smears, and cultures. Experienced microbiologists interpret the phenotypic characteristics of microorganisms, providing insights into infections. However, the increasing demand for diagnostic services and a shortage of laboratory personnel have necessitated the automation of these processes, leading to significant advancements in artificial intelligence (AI) within this field.

AI, particularly through the use of Convolutional Neural Networks (CNNs), has proven effective in image classification and the identification of pathogens in various clinical samples. Numerous studies have demonstrated the ability of AI to automatically detect infectious agents such as Plasmodium, the causative agent of malaria, and Mycobacterium tuberculosis, responsible for tuberculosis. For instance, CNNs have been applied to analyze blood smears for Plasmodium, achieving diagnostic accuracies of up to 99.7%. Techniques like pixel classification and quantitative phase spectroscopy further enhance the ability to identify these parasites at different stages of infection.

In tuberculosis diagnostics, automated systems utilizing CNNs have significantly improved the detection of acid-fast bacilli in sputum samples, with reported accuracies reaching as high as 99%. Similarly, deep learning models have been employed to analyze Gram stain images and chromogenic agar for identifying bacterial species, consistently achieving accuracy rates over 97%. These automated systems streamline the identification process, allowing for quicker and more reliable results in clinical settings.

The application of AI extends beyond bacterial infections to include the detection of fungi and viruses, such as SARS-CoV-2. Various imaging techniques and analyses have been utilized to enhance diagnostic capabilities for these pathogens, showcasing the versatility of AI in clinical microbiology.

Overall, the integration of AI in image analysis is transforming laboratory diagnostics, making them more efficient and accurate. By reducing the reliance on human expertise and labor-intensive processes, AI-driven technologies enhance diagnostic accuracy and response times, ultimately improving patient care in clinical microbiology laboratories.

Automated Factor Analysis

Machine learning (ML) approaches have been effectively applied in clinical sample analysis to enhance detection rates for various infections. For example, Wang et al. developed an automated urine analysis system to improve the detection of Trichomonas vaginalis, the parasite causing trichomoniasis. By using classification models such as random forest, linear regression, and support vector machines, they assessed key variables in urine analysis—including nitrite levels, protein, and white blood cell counts—suggesting that their ML-based method could significantly increase detection rates in a cost-effective manner.

Similarly, researchers utilized an XGBoost algorithm to create an ML model for COVID-19 diagnosis, relying on routine blood parameters like eosinophil count and prothrombin activity. This model serves as a potential diagnostic tool in clinical settings. Another study applied multiple logistic regression, random forest, and naïve Bayes algorithms for the differential diagnosis of viral and bacterial meningitis, using cerebrospinal fluid and blood parameters as predictors. Their findings indicated that an accuracy above 95% for viral meningitis and 78% for bacterial meningitis is necessary for optimal prediction.

Antimicrobial Resistance Analysis

Antimicrobial resistance (AMR) poses a significant challenge in modern medicine, prompting increased attention to machine learning (ML) algorithms over the past seven years. This surge is attributed to the rapid growth of experimental and clinical data, enhanced algorithm performance, and the urgent need for innovative solutions against multidrug-resistant (MDR) microorganisms. Deep learning has been applied to predict antibiotic resistance genes from metagenomic and genome sequence data and identify mutations related to AMR. ML techniques, combined with high-throughput digital PCR and amplification methods, have shown promise in accurately detecting carbapenem-resistance genes in clinical isolates.

Additionally, surface-enhanced Raman spectroscopy (SERS) has emerged as a valuable tool for detecting antibiotic-resistant bacteria due to its high sensitivity and low cost. SERS, when paired with deep learning, has effectively discriminated between resistant strains such as methicillin-resistant Staphylococcus aureus (MRSA). Recent studies, including one by Jeon et al., have utilized MALDI-TOF spectral data and ML for identifying MRSA, achieving sensitivities and specificities of 91.8% and 83.3%, respectively. Various reviews have highlighted the potential of AI in addressing antibiotic resistance challenges.

Antimicrobial Discovery

Artificial intelligence (AI) is emerging as a powerful tool in discovering novel antimicrobial peptides (AMPs) to combat antimicrobial resistance (AMR). Traditional antibiotic discovery relies on extensive synthetic chemical libraries, which are costly and limited in chemical diversity. Recent advancements in machine learning (ML), particularly deep learning, enable more effective molecular property prediction, allowing researchers to identify new classes of antibiotics.

For instance, a study by Wang et al. utilized various ML models to predict new antimicrobial molecules targeting Staphylococcus aureus. Similarly, Stokes et al. employed a message-passing neural network (MPNN) to screen approximately 107 million structurally diverse chemicals, leading to the discovery of eight novel antibacterial compounds, including halicin, which demonstrated efficacy against broad-spectrum bacterial infections in vivo. Liu et al. also identified a new antibacterial molecule, abaucin, against Acinetobacter baumannii. Additionally, deep learning techniques, such as long short-term memory (LSTM) models, have been used to design novel AMP sequences. These innovative approaches are gaining traction and are the subject of various systematic reviews in the literature.

Microbiome Analysis

With the critical role of the microbiome in various human diseases and the growing importance of microbiome research, characterization of the microbiome and host–microbiome associations remains critical for our understanding of various complex diseases. The omics-based methods, such as metagenomics, metatranscriptomics, and metabolomics, are widely used in the study of gut microbiome due to their ability to provide high-throughput and high-resolution data. The vast amount of data generated via these methods has led to the development of computational methods for data processing and analysis, which is a field where ML can be used as a powerful tool [108]. In this regard, ML provides new insights into the development of models that can be used to predict outputs, such as classification and prediction in microbiology, extrapolation of host phenotypes to predict diseases, and the use of microbial communities to stratify patients by the characterization of state-specific microbial signatures [109]. Advances in this field have been highlighted by several recent studies [110,111,112,113].
The applicability of AI in the field of microbiology and infectious diseases is summarized in Figure 2.

AI-applications-in-the-field-of-microbiology

Figure 2. Artificial intelligence applications in the field of microbiology and infectious diseases with a focus on antimicrobial resistance and hospital-acquired infections.

AI and Hospital-Acquired Infections

Healthcare-associated infections (HAIs), also known as hospital-acquired infections, are infections not present in a patient prior to hospitalization. They frequently occur in intensive care units (ICUs), where patients face a 5 to 10 times higher risk due to factors like immunodeficiency and the use of invasive medical devices. Common types of HAIs include central line-associated bloodstream infections (CLABSIs), catheter-associated urinary tract infections (CAUTIs), ventilator-associated pneumonia (VAP), and surgical site infections (SSIs). Clostridium difficile is also a prevalent cause of nosocomial diarrhea.

HAIs pose significant clinical and financial burdens, leading to longer hospital stays, increased mortality, and rising healthcare costs, with an estimated annual burden of $10 billion in the U.S. Effective surveillance of HAIs is crucial for implementing infection prevention and control programs. This data helps monitor HAI rates, detect outbreaks, and identify risk factors. Machine learning (ML) is being increasingly applied to enhance understanding of HAI risk factors, improve patient risk stratification, identify transmission routes, and facilitate early detection and control measures. ML techniques can be classified into supervised and unsupervised learning, both of which optimize data-driven models for better predictive accuracy.

Intensive Care Units

A significant number of AI and machine learning (ML) models have been developed for event prediction, commonly referred to as “forecasting.”(Figure 3). Given the high incidence of healthcare-associated infections (HAIs) globally, the early detection and prevention of HAIs using AI has gained considerable attention. AI and ML show promise in developing surveillance algorithms that help understand risk factors, improve patient risk stratification, identify transmission pathways, and enable timely detection of infections.

Electronic health data serve as a crucial information source, facilitating the implementation of real-time decision support systems for HAI surveillance. ML-enabled clinical decision support systems in ICUs focus on monitoring, early identification of clinical events, and outcome prediction, aiding clinicians and policymakers. Various ML models have been created to forecast ventilator-associated pneumonia (VAP), central line-associated bloodstream infections (CLABSIs), surgical site infections (SSIs), and the risk of infections with multidrug-resistant pathogens. However, the majority of research has concentrated on forecasting sepsis and septic shock.

Application-of-machine-learning-models

Figure 3. Application of machine learning models to the predictive surveillance and forecasting of hospital-acquired infections (HAIs).

Ventilator-Associated Pneumonia

Ventilator-associated pneumonia (VAP) occurs more than 48 hours after a patient has been intubated and is the most common nosocomial pneumonia in critically ill patients. Early recognition and prevention of VAP are crucial in critical care units. Risk factors include patient-specific elements like age and pre-existing conditions (e.g., COPD) as well as care-related factors such as head-of-the-bed angle, aspiration, and previous antibiotic treatment.

Several studies have applied machine learning (ML) algorithms for early VAP prediction. For example, Liang et al. used a random forest algorithm to analyze 38,515 ventilation sessions, identifying key predictive features such as the PaO2/FiO2 ratio and APACHE III score. Their model achieved an AUC of 84% in validation. Another study by Giang et al. compared various ML models using electronic health records, finding that XGBoost provided the highest AUROC of 0.854. Key features for this model included the duration of mechanical ventilation and recent GCS assessments.

These ML models could lead to the development of warning systems that improve diagnosis and optimize antibiotic treatment. Similar studies have also explored individualized risk assessment for VAP and mortality prediction in patients with severe pneumonia.

Central Line-Associated Bloodstream Infections (CLABSIs)

Central line-associated bloodstream infections (CLABSIs) are defined by the CDC as bloodstream infections that develop 48 hours after central line placement and cannot be attributed to other sources. Identifying high-risk patients can facilitate earlier treatment and better monitoring, such as timely catheter replacements.

Several studies have focused on developing machine learning (ML) algorithms to predict CLABSIs. For example, Rahmani et al. analyzed electronic health records of 27,619 patients and used classifiers like XGBoost, logistic regression, and decision trees. XGBoost outperformed the others with an AUROC of 0.762, identifying key predictors such as age, race, temperature, and comorbidities.

Another study involving 70,218 patients found that a random forest model outperformed logistic regression, with important features including age and the number of days with a central venous line. Additional research from Boston Children’s Hospital developed an ML model predicting CLABSIs in 7,468 cardiac patients, with predictors like previous infections and elevated vital signs.

These models underscore the potential for early identification of CLABSI risk, which can lead to improved patient outcomes and reduced healthcare costs.

Surgical Site Infections (SSIs)

Surgical site infections (SSIs) are common postoperative complications leading to significant morbidity, mortality, prolonged hospital stays, and financial strain on healthcare systems. To address this, various groups have developed machine learning (ML) predictive models for SSIs.

In 2021, Petrosyan et al. created a three-stage algorithm using health administrative datasets, employing a random forest algorithm for preliminary screening followed by logistic regression to select the top predictors for SSIs. Their model demonstrated high accuracy in predicting SSIs.

More recently, Wu et al. developed nine XGBoost models to detect SSIs after total hip and knee arthroplasties, achieving an ROC AUC of 0.906. Similarly, Chen et al. used perioperative data from 4,019 lumbar spinal surgery patients to identify predictors of SSIs, reporting an impressive ROC AUC of 0.986.

Some healthcare systems, like the University of Iowa Hospitals & Clinics, have successfully implemented ML to reduce SSIs by 74% over three years through their DASH analytics system, which integrates with electronic health records to predict risks and prevent infections.

Sepsis

Timely detection of sepsis is crucial for patient outcomes, and electronic health records (EHRs) serve as effective tools for building machine learning (ML) models to enhance early detection in intensive care units (ICUs). Several studies have utilized ML models trained on individual patient EHR data for this purpose.

For instance, Wang et al. analyzed 55 features from the EHRs of 4,449 infected patients, employing a random forest algorithm that achieved an ROC AUC of 0.91, demonstrating strong predictive capability for sepsis in Chinese hospitals. Lauritsen et al. compared different models for early sepsis detection using retrospective data from multiple Danish hospitals, showing AUROC values ranging from 0.856 three hours before onset to 0.756 24 hours prior.

Additionally, AI algorithms, such as Fagerström et al.’s LiSep LSTM model, which used data from 59,000 ICU patients, achieved an AUROC of 0.8306 for predicting sepsis. These studies indicate that ML models can significantly improve the early diagnosis of sepsis compared to traditional methods like the SOFA score.

Clostridium difficile Infection (CDI) and Complications

Despite the growing burden of healthcare-associated infections (HAIs) worldwide, the use of machine learning (ML) in this area remains underexplored. Clostridium difficile infections (CDIs) are common in healthcare settings, prompting studies to develop risk prediction models. One study at two academic health centers found that L2 models achieved AUROC values between 0.75 and 0.82, but noted variability in predictive factors across facilities.

Panchavati et al. compared ML approaches for predicting CDI in hospitalized patients, finding that XGBoost outperformed other methods with an AUROC of 0.815 after just six hours of hospitalization. Conversely, Escobar et al. evaluated over 150 predictors but failed to develop models that could predict recurrent CDI effectively.

CDI complications can lead to serious outcomes, including ICU admission and death. Li et al. explored ML for risk stratification of complications in CDI patients. They used various patient features, reporting an AUROC of 0.69 for complications diagnosis, which improved to 0.90 when using data collected two days post-diagnosis. This highlights the potential of ML to accurately stratify CDI cases by their risk of complications.

Multidrug-Resistant (MDR) Pathogens

Colonization or infection by multidrug-resistant (MDR) microorganisms poses a significant threat to vulnerable ICU patients. Since MDR confirmation tests can take up to 48 hours, various AI approaches have been employed to predict risk factors during this critical period.

In a 2021 study by Mora-Jimenez et al., researchers identified significant features for predicting MDR infections, such as SAPS III, Apache II score, age, and department of origin, based on clinical and demographic data along with mechanical ventilation and antibiotic use.

Similarly, Liang et al. developed a model for predicting carbapenem-resistant Gram-negative bacteria (CR-GNB) carriage in 2022. Analyzing data from 2,910 patients, they found that the random forest model outperformed others, achieving an overall accuracy of 85.92% by successfully predicting 74 out of 86 positive CR-GNB cases.

These studies underscore the potential of ML models to predict MDR colonization or infection within the ICU, enabling timely identification of high-risk patients.

Hand Hygiene

Healthcare personnel’s hands are a primary source of healthcare-associated infections (HAIs), and effective hand hygiene (HH) practices are crucial for infection prevention. While HH is a simple strategy, its implementation can be challenging in hospital settings.

Many studies have leveraged the Internet of Things (IoT) to enhance healthcare workers’ self-awareness regarding HH. Innovations include automatic hand hygiene surveillance systems (AHHMS) using cloud-based servers and wearable devices to monitor compliance, often supplemented by reminders.

AI training systems have also been explored as cost-effective methods for monitoring and improving HH quality. For example, Greco et al. utilized convolutional neural networks (CNNs) to analyze handwashing sequences in real time, while Nagar and colleagues developed a CNN-based model to assess an individual’s microorganism levels during handwashing. Their model achieved high accuracy in detecting handwashing steps, microorganism levels, and overall compliance with WHO guidelines.

Recent studies continue to investigate ML models for automatic detection of hand hygiene practices, highlighting the potential for improved compliance and infection control.

Conclusions

AI applications are inevitably becoming a part of modern healthcare with a high potential to aid caretakers and decision-makers in the fields of laboratory and imaging diagnosis, antimicrobial stewardship, discovery of antimicrobials, microbiome-based translational interventions, infectious disease surveillance, prediction and prevention. The mass digitalization of health records making data accessible and advances in computer power has been instrumental and will remain crucial for future research and development in the field. Although AI is commonly regarded as a threat for “common” jobs, its integration into healthcare should instead be seen as an opportunity for improved patient care and infection management, increased survival, better allocation of staff and resources and lowered costs in healthcare systems.

Source: Baddal, B.; Taner, F.; Uzun Ozsahin, D. Harnessing of Artificial Intelligence for the Diagnosis and Prevention of Hospital-Acquired Infections: A Systematic Review. Diagnostics 2024, 14, 484.