Background

Antimicrobial resistance (AMR) is a complex public health issue. A modeling study estimated that 4.95 million deaths globally in 2019 were related to bacterial AMR, with a disproportionate impact on resource-limited areas. ^{Reference Murray1} Inappropriate use of antimicrobials remains a key driver of AMR, particularly in outpatient settings and hospitals. For example, 25–28% of antibiotic prescriptions for respiratory infections in ambulatory care are unnecessary, ^{Reference Hersh, King, Shapiro, Hicks and Fleming-Dutra2} while up to 50% of hospital antibiotic use is considered inappropriate. ^{Reference Hulscher, Grol and van der Meer3}

To preserve the effectiveness of antimicrobial agents, it is imperative to improve the use of antimicrobials globally. ^{Reference Hulscher, Grol and van der Meer3} Antimicrobial stewardship (AMS) programs aim to optimize antimicrobial prescription by selecting the appropriate agent, dose, and duration tailored to individual patients. ^{Reference Dyar, Huttner, Schouten and Pulcini4} AMS strategies include clinical guideline adherence, de-escalation, intravenous-to-oral switches, drug monitoring, and specialist consultation. ^{Reference Davey5} These interventions have been associated with reduced antibiotic use, ^{Reference Zay Ya, Win, Bielicki, Lambiris and Fink6} but are often resource-intensive, cover only a subset of prescriptions, and are typically confined to regular working hours. ^{Reference Divecha, Tullu and Karande7} Moreover, they are usually reactive, implemented after antimicrobials have already been prescribed, potentially missing opportunities to prevent unnecessary exposure. ^{Reference Rawson8,Reference Tamma9}

Computerized clinical decision support systems (CDSS) are software designed to support healthcare professionals in clinical decision-making. For AMS interventions, CDSS can overcome limitations of “human” interventions by overseeing antibiotic prescriptions in electronic health records (EHR) continuously. CDSS function as proactive support systems, assisting in decision-making before antimicrobial prescription. ^{Reference Rawson8} CDSS integrate with EHR data and incorporate key AMS components like preauthorization, duration suggestions, medication switches, de-escalation strategies, allergy alerts, reassessment protocols, and restrictions, enabling multimodal interventions.

In this review, we aim to summarize current evidence available on CDSS as tools for AMS interventions, with a focus on evidence of their effectiveness and challenges to implementation.

Computerized clinical decision support systems in healthcare

CDSS offer evidence-based recommendations and information relevant to patient care. CDSS combine patient data, medical knowledge, and clinical guidelines, enabling healthcare providers to effectively diagnose diseases, choose appropriate treatments, and ultimately improve patient outcomes. ^{Reference Sutton10} Originally developed for academic research in the 1970s, CDSS have evolved to include stand-alone applications as well as tools integrated into EHR and computerized provider order entry (CPOE) systems. ^{Reference Sutton10}

CDSS can be classified into two types: expert systems and machine-learning (ML) systems. ^{Reference Peiffer-Smadja11} The former, which constitutes the majority, so far operates based on rules established according to expert opinion and guidelines. In these systems, the quality of the experts’ knowledge and recommendations is crucial. In contrast, ML-based systems use algorithms—including supervised and unsupervised learning methods, such as neural networks and deep learning—to identify patterns and improve predictions based on available data. The performance of these systems depends on the quality, completeness and relevance of the input data, which directly influences the accuracy and validity of the outputs. Unlike expert-based systems, ML algorithms continuously learn from new data and adapt over time, allowing for dynamic, real-time insights into complex prescribing scenarios. ^{Reference Peiffer-Smadja11} The integration of ML into AMS advances the development of precise, predictive decision support systems. These models are able to synthesize a wide range of extensive datasets—including clinical history, microbiology results, pharmacy data, and genomics—to predict AMR profiles and optimize empiric therapy. ML-based CDSS remain mostly experimental. Systems used in real settings like the Veterans Health Administration (VA), UK National Health Service (NHS), or Epic modules still rely on expert-rule-based algorithms rather than adaptive machine learning. ¹² The development of ML techniques and the interoperability of ML algorithms with EHR data may be used to optimize the use of antimicrobials among patients. ML algorithms can predict individual infection resistance profiles, improving the appropriateness of empiric therapy even before culture and susceptibility results—traditionally several days—are available. ^{Reference Liu13} Table 1 highlights three recent examples of ML-driven CDSS designed to improve antimicrobial prescribing through personalized antibiotic resistance prediction and empirical treatment selection.

Table 1.

Examples of ML CDSS for antimicrobial prescribing

ML: machine learning; UTI: urinary tract infection; AMR: antimicrobial resistance.

Despite these promising advancements, many ML-based CDSS have so far demonstrated performance largely through internal validation on retrospective datasets, with limited external validation and prospective evaluation, both of which are needed for clinical implementation. ML implementation is limited by access to quality annotated data, local microbiological variations, and integration into clinical workflows. ^{Reference Ahmed14,Reference Mittermaier, Raza and Kvedar15} There may also be issues related to biases in training and test data set used to develop ML models, as well as limited interpretability—clarifying the relative contributions of various features to the models—or explainability—showing how the models generate their outputs and reach their conclusions—of ML models which may be a barrier to their adoption in clinical settings. ^{Reference Agarwal and Gao16} The continued development of artificial intelligence and ML techniques is expected to significantly impact AMS interventions in the short to medium term. Hybrid models integrating expert rules with machine learning may offer an optimal solution by enhancing interpretability while maintaining prediction. The use of large datasets, for example through collaborations with national, or linked health data systems could also be helpful for assessments of effectiveness.

Modes of implementation for AMS digital tools :“stand-alone” tools versus CDSS integrated within EHR

Digital tools for AMS can be implemented as standalone systems or as CDSS integrated into EHR, as shown in Table 2. Stand-alone tools, typically accessed via mobile devices, function independently of EHR, while integrated tools are embedded within EHR systems and use patient-specific data for decision support. ^{Reference Van Dort, Penm, Ritchie and Baysari17} Stand-alone smartphone applications provide clinicians quick access to antimicrobial therapy guidance from local to international sources. These applications offer accessibility, frequent updates, and comprehensive pathogen and treatment information at lower costs than integrated systems. However, their effectiveness depends on clinicians’ usage, and data source quality must be verified. ^{Reference Catho18}

Table 2.

Examples of stand-alone tools versus integrated CDSS for AMS

^* : The availability and specifics of features like alert systems, dynamic reports (automated, interactive dashboard that continuously aggregates and analyzes real-time clinical data such as antibiotic use, resistance patterns, patient outcomes), dose checking, IV-to-PO switch, and infection prevention and control modules can vary based on the version and customization of the software.

⁺ : See Appendix B for specific DOIs.

IPC: infection prevention and control; IV: intravenous; PO: per oral; EHR: electronic health record.

CDSS integrated within EHR delivers personalized medical recommendations by analyzing patient data and guidelines to suggest optimal treatments, dosages, and durations. These systems incorporate AMS principles, enabling antimicrobial optimization based on clinical data, alerts for revaluation, and suggestions to avoid unnecessary prescriptions. ^{Reference Sutton10,Reference Catho18} Such systems could incorporate complex recommendations based on ML models. While CDSS offers benefits like workflow integration and customizable features, implementation requires substantial IT infrastructure and maintenance. ^{Reference Tokgöz, Hafner and Dockweiler19} Careful workflow evaluation during development should be conducted to maximize end-user uptake.

CDSS process outcomes

Baysari et al conducted a meta-analysis of 45 studies evaluating CDSS for AMS, showing increased appropriate antimicrobial use in hospitals (pooled Relative Risk [RR]: 1.49, 95% Confidence Interval [CI]: 1.07–2.08). ^{Reference Baysari23} However, study quality was suboptimal, and only four types of computerized AMS interventions were included. Analysis of high-quality studies showed inconclusive effects on prescribing (pooled RR: 1.53, 95% CI: .96–2.44), with little evidence of impact on mortality or hospital stay duration. The diverse study designs and outcome measures complicated comparisons between IT interventions. Rawson et al analyzed 58 studies of 38 CDSS for AMS, finding moderate to high-quality evidence that CDSS impact antimicrobial prescribing behaviors. ^{Reference Rawson8} Outcomes varied from guideline appropriateness to empirical treatment aligned with microbiological sensitivity. The review noted only two randomized controlled trials, one outpatient and one hospital-based. ^{Reference Rawson8}

CDSS clinical outcomes

Van Dort et al. conducted a qualitative synthesis of eight systematic reviews, revealing a trend toward reduced antimicrobial use and improved prescribing appropriateness when using CDSS interventions. ^{Reference Van Dort, Carland, Penm, Ritchie and Baysari24} However, significant heterogeneity in interventions and outcome measures prevented meta-analysis. The synthesis reported inconsistent findings regarding the impact of digital interventions on clinical outcomes. ^{Reference Van Dort, Penm, Ritchie and Baysari17} Additional reviews support the potential of CDSS in hospital settings and emphasize the need for higher-quality studies and inclusion of clinical outcomes among patients, including mortality. ^{Reference Curtis, Al Bahar and Marriott25} A systematic review of 45 studies by Rittman et al. found that the efficacy of CDSS and patient outcomes fluctuated based on different settings, underscoring the need for context-specific CDSS that are optimally designed and implemented. ^{Reference Rittmann and Stevens26} In the context of perioperative prophylaxis, Simpaoet al. conducted a systematic review of 25 studies and found strong evidence that CDSS could improve antibiotic prescribing practices. ^{Reference Simpao27} A separate scoping review highlighted the potential of CDSS in pharmacies to decrease antibiotic use in community settings, provided that they are implemented appropriately. ^{Reference Moon, Chladek, Wilson and Chui28}

These findings indicate that CDSS show stronger benefits for process outcomes like prescribing appropriateness and reduced antibiotic use, rather than clinical endpoints. The diverse interventions and varying study quality limit comparisons, with modest effect sizes reported.

Schweitzer et al. conducted a comprehensive review of studies from 1950 to 2017, evaluating AMS interventions in various settings to identify factors limiting the applicability of research findings to clinical practice. ^{Reference Schweitzer29} The review found that most AMS studies were of poor quality, lacking proper control groups and relying heavily on before-after designs (675/825, 82%). While common in quality improvement initiatives, these quasi-experimental studies are prone to bias, leading the Effective Practice and Organisation of Care (EPOC) to discourage their inclusion in systematic reviews in the absence of interrupted time series analyses. In addition, studies investigating AMS intervention in acute care settings report only clinical and microbiological outcomes, frequently omitting key process indicators.

When evaluating AMS interventions, De Kraker et al., highlighted that implementation occurs in complex healthcare environments where biases and temporal effects can obscure causal inferences. ^{Reference de Kraker, Abbas, Huttner and Harbarth30} Key challenges include concurrent infection prevention strategies, regression to the mean, and contamination from cluster-level interventions. ^{Reference de Kraker, Abbas, Huttner and Harbarth30} These complexities underscore the need for comprehensive measurement of CDSS effectiveness in AMS. A rigorous evaluation framework would enable better understanding of CDSS impact across healthcare settings. Schweitzer et al proposed guidelines to enhance AMS research design, addressing methodology, timing, data gathering, analysis, outcomes, and reporting. The group emphasized aligning study design with goals, such as determining superiority or non-inferiority, and defining effect sizes. Timing should match study objectives, while sequential data collection is crucial for interrupted time series analyses. To minimize bias, suggested approaches include the Desirability of Outcome Ranking (DOOR) and Response Adjusted for Days of Antibiotic Risk (RADAR). Outcomes should cover clinical, microbiological, and care-related areas. They emphasized transparent reporting of effects and suitable controls for practical application. ^{Reference Schweitzer31} The prevalence of before-after study designs likely stems from financial constraints and logistical difficulties in conducting controlled trials in healthcare settings. Addressing this requires enhanced support for AMS research and increased methodological capacity to prioritize more rigorous study designs.

Overall, the paucity of high-quality and comparable data remains a major obstacle to assess CDSS effectiveness in AMS, particularly regarding the impact of CDSS on AMR and other microbiological outcomes. Results from quasi-experimental AMS studies should be interpreted cautiously, given the inherent challenges in evaluating CDSS effectiveness in this complex field. These collective findings suggest that while CDSS shows promise in improving antimicrobial prescribing practices, the heterogeneity in study designs, quality, and settings warrants cautious interpretation and highlights the need for further high-quality research to establish their effectiveness. Principal challenges related to assessing effectiveness of CDSS in AMS and are summarized in Table 3.

Table 3.

Overview of principal challenges to CDSS in AMS

ML: machine learning; AI: artificial intelligence.