Setting

This retrospective study was conducted at Queen Mary Hospital, a 1,700‑bed university‑affiliated hospital governed by the Hospital Authority in Hong Kong, providing specialized services, including solid organ and bone marrow transplantation. An ASP targeting broad-spectrum intravenous antimicrobials began in January 2005, ^{Reference Cheng, To and Li12} building on a vancomycin usage audit initiated in 1997. ^{Reference Kumana, Ching and Kong13} The Smart ASP initially focused on piperacillin–tazobactam and meropenem, later expanding to include cefoperazone–sulbactam, cefepime, imipenem–cilastatin, and vancomycin, in accordance with the Interhospital Multidisciplinary Programme on Antimicrobial ChemoTherapy (IMPACT) guideline first published in Hong Kong in 1999. ^{Reference Ma, Tsui, Wu and Ho14}

ASP before implementation of the rule-based decision support system

Since 1997, a multidisciplinary team manually conducted the ASP. ^{Reference Cheng, To and Li12,Reference Kumana, Ching and Kong13} Pharmacists generated daily lists of inpatients receiving targeted antimicrobials for ICN review. Each morning, ICNs reviewed patients’ medical records at the bedside to evaluate antimicrobial indications. Every working day at midday, ICNs and clinical microbiologists met to determine the appropriateness of antimicrobial prescriptions. Prescription appropriateness was determined based on predefined clinical justifications, requiring specific criteria for either targeted or empirical therapy. For targeted therapy, appropriateness required matching the antimicrobial to available microbiological culture and susceptibility results. For empirical therapy, appropriateness required adherence to guideline-approved indications for specific disease states or clinical scenarios, including: documented allergy history precluding narrower-spectrum first-line agents, neutropenic fever, documented clinical failure of first-line antibiotics, immunocompromised status, nosocomial infection, unreliable or impossible oral intake/absorption, severe clinical infection, or other specified valid clinical indications. For noncompliant prescriptions, ICNs provided immediate feedback to prescribing physicians through memos in the afternoon. In complex cases without consensus, clinical microbiologists conducted bedside evaluations and offered recommendations. Beginning in January 2010, the ASP was incorporated into the inpatient consultation service provided by clinical microbiologists ^{Reference Yuen, Seto and Chau15} to assess compliance with the IMPACT guideline and appropriateness as defined above.

ASP after implementation of the rule-based decision support system

In early 2024, with support from the Hospital Authority’s information technology (IT) unit, a protocol‑driven IT tool called “Smart ASP” was developed to detect potentially inappropriate antimicrobial prescriptions through advanced data analytics. Smart ASP consists of four main elements. First, the Artificial Intelligence and Data Analytics (AIDA) platform functions as a rule-based expert system applying predefined rules to clinical and laboratory data (eg, inflammatory markers, cultures) to flag potentially inappropriate prescriptions. Second, Medication Genie is a decision‑support module that is directly triggered by AIDA’s output, issuing automated reminders to clinicians when a prescription requires review. Third, Smart ASP patient list and form provide workflow and documentation support such as standardized electronic templates for audit notes and predefined menus for selecting specific appropriateness or inappropriateness reasons during prescription audits. Lastly, the performance reporting function produces summary reports for monitoring program activity and trends over time. These components streamline case identification, deliver timely clinician feedback, facilitate auditing, and enable continuous stewardship monitoring. Automated case identification expanded monitoring from historically assessed units (Medicine, Surgery, and Orthopedics) to previously unassessed units, including Oncology and the Intensive care unit (ICU). The Smart ASP service model and analytic cycle are illustrated in Table 1 and Supplementary Table 1, respectively.

Table 1.

Service model of Smart ASP for selected broad-spectrum antimicrobials in the hospital ^a

Table 1. Long description

A table outlining the steps and performance of the Smart ASP service model for antimicrobial prescriptions. The table has two columns: Step and Performance. It contains six rows detailing the process. Row 1: Step 1, Clinicians initiate antimicrobial therapy based on each patient’s clinical presentation (Day 1). Wherever feasible, bacterial cultures and other microbiological tests are obtained before starting antimicrobials to inform the choice and duration of therapy; additional investigations may be performed afterward if clinically indicated. Row 2: Step 2, For selected broad-spectrum antimicrobials, the IPMOE system prompts a list of approved indications to guide prescribing. Row 3: Step 3, The AIDA platform runs predefined analyses daily to detect potential inappropriate antimicrobial use. When AIDA flags a prescription, the Medication Genie generates a reminder within the IPMOE system prompting the clinician to review the order as appropriate. Row 4: Step 4, The AIDA analytical cycle continues to monitor the prescription (up to Day 15) unless the order is modified or discontinued. An illustration of the analytic cycle is shown in Supplementary Table 1. Row 5: Step 5, Potentially inappropriate prescriptions flagged by AIDA on Day 4 and Day 15 are added to the Smart ASP patient list. The hospital ASP team can then review these cases and send feedback to the prescriber using the Smart ASP form. Row 6: Step 6, An automatically generated monthly management report detailing key metrics such as total antimicrobial consumption (in D D D s), number of cases reviewed, and the percentage of appropriate prescriptions is produced for the Hospital Chief Executive, department Chief of Service, and the hospital ASP team for ongoing monitoring and follow-up.

Note: AIDA, Artificial Intelligence and Data Analytics; IPMOE, Inpatient Medication Order Entry.

^a The selected broad‑spectrum antimicrobials included piperacillin–tazobactam, cefoperazone–sulbactam, cefepime, imipenem–cilastatin, meropenem, and vancomycin. Smart ASP was rolled out in phases: meropenem was added on 1 February 2024, piperacillin–tazobactam on 5 March 2024, and cefoperazone–sulbactam, cefepime, imipenem–cilastatin, and vancomycin on 7 May 2024.

^b IPMOE (Inpatient Medication Order Entry) is the locally standardized closed-loop electronic medication administration subsystem utilized across Hospital Authority institutions in Hong Kong.

^c Day 4 was selected to allow sufficient time for final culture and susceptibility results to be reported and acted upon by the primary team, while Day 15 was chosen to flag unusually prolonged courses of therapy that require clinical microbiologist or infectious disease specialist review.

Consumption of antimicrobials before and after implementation of Smart ASP

Antimicrobial prescriptions from 2016 to 2025 were retrieved from the Clinical Data Analysis and Reporting System (CDARS) as previously described. ^{Reference Wong, Chau and So16} This included overall consumption of all antimicrobial classes and individual agents across various specialties, such as Medicine, Surgery, Orthopedics, Oncology, and the ICU. Monthly antimicrobial consumption was calculated as defined daily dose (DDD) per 1,000 patient days using patient days data from CDARS.

Incidence of MDROs before and after implementation of Smart ASP

The corresponding incidence of new MDRO cases, including methicillin-resistant Staphylococcus aureus (MRSA), vancomycin-resistant Enterococci (VRE), extended-spectrum beta-lactamase (ESBL)-producing organisms and carbapenem-resistant Acinetobacter (CRA) species, was retrieved from the CDARS. Data on carbapenemase-producing Enterobacterales (CPE) were obtained from the laboratory information system. A new case was defined as a patient without a laboratory culture of the respective MDROs in the preceding 12 months. Incidence is expressed as the number of new cases per 1,000 patient days. ^{Reference Wong, Chau and So17}

Statistical analysis

Descriptive statistics summarized antimicrobial consumption from 2016 to 2025 as DDD per 1,000 patient days, reported as medians with interquartile ranges (IQRs). Continuous longitudinal trends in baseline antimicrobial consumption over the 10-year period were assessed using ordinary least squares (OLS) linear regression. Nonlinear longitudinal trends in MDRO incidence rates (per 1,000 patient days) were evaluated using Locally Weighted Scatterplot Smoothing (LOWESS) with a smoothing fraction of 0.3.

To assess the rule-based Smart ASP’s impact, an interrupted time series (ITS) analysis was conducted using segmented OLS regression, aggregating data into monthly intervals. Specifically, total monthly antimicrobial DDDs were divided by total monthly patient days to calculate the continuous variable of DDD per 1,000 patient days, which was utilized in the regression models. The ITS models estimated two parameters: immediate level change (intercept step-change) at implementation and trend change (slope) during the postintervention period relative to the preintervention baseline. Residual autocorrelation was checked using the Durbin–Watson statistic, and subgroup ITS analyses evaluated the intervention’s impact in specific clinical specialties (Medicine, Surgery, Orthopedics, and ICU). Sensitivity analyses with 1-, 2-, and 3-month lag periods assessed potential delays in intervention effects.

To investigate the bidirectional temporal relationships between broad-spectrum antimicrobial consumption and MDRO incidence, a cross-correlation function (CCF) analysis with double-prewhitening was performed to prevent spurious correlations driven by underlying temporal trends or autoregressive artifacts. Optimal autoregressive integrated moving average (ARIMA) models were fitted using the lowest Akaike information criterion (AIC) to independently reduce both the antimicrobial and MDRO time series to white noise. To confirm the appropriate use of this statistical approach, Ljung–Box testing was performed in advance to verify that the ARIMA-derived residuals for all evaluated series were successfully reduced to white noise. Subsequently, cross-correlations of the prewhitened residuals were calculated for time lags ranging from −3 to +3 months, with statistical significance of lagged correlations evaluated using Z-scores. Changes in prescribing quality were assessed by comparing the proportion of appropriate antimicrobial prescriptions between the preintervention (2019 Q1–2023 Q4) and postintervention (2024 Q2–2025 Q4) periods using the Mann–Whitney test. Data from the transition period of Smart ASP in Q1 2024 were excluded from the analysis. Specialties without preintervention data were excluded from this analysis. All statistical analyses were performed using Python (version 3.14), with a two-sided P value < .05 considered statistically significant.