TL;DR
This paper demonstrates that electronic health records (EHRs) can be effectively leveraged to automate the monitoring of antimicrobial stewardship (AMS) practices in hospitals. By developing novel data modeling techniques, the researchers created a feasible system for generating actionable insights into prescribing quality, which can support quality improvement initiatives and clinician engagement.
What Problem Does This Paper Address?
Antimicrobial stewardship (AMS) programs are crucial for optimizing antibiotic use, preventing resistance, and ensuring effective patient treatment. However, traditional methods for monitoring AMS, such as manual audits and point prevalence surveys, are time-consuming, labor-intensive, and require significant specialist expertise. This creates a bottleneck, hindering hospitals' ability to monitor prescribing practices on a large scale and provide timely feedback to clinicians. The UK's Antimicrobial Resistance National Action Plan specifically calls for the comprehensive use of EHRs to drive AMS by 2025, highlighting the need for efficient, data-driven approaches. This paper addresses the gap by exploring the feasibility of using routinely collected EHR data to automate the generation of meaningful AMS metrics.
Methods
The core of this study lies in developing an informatics methodology to analyze hospital AMS practices using routine electronic prescribing and laboratory records. The approach involved three key steps:
- Translating AMS Standards into Measurable Concepts: The researchers identified concepts relevant to AMS performance from clinical guidelines and infection surveillance protocols and translated them into operational definitions that could be extracted from EHRs.
- Data Modeling for Therapy Episodes: A novel data modeling approach was developed using graph theory principles to reconstruct "therapy episodes." This involves linking related prescription records to define periods of uninterrupted antibiotic therapy, allowing for the analysis of therapy changes, duration, and discontinuation. A dynamic table with hourly resolution was constructed for each episode to track changes in therapy in relation to clinical parameters.
- Pilot Application and Metric Computation: The developed methods were piloted on a large EHR dataset. This involved computing various AMS metrics, including:
- Antimicrobial Consumption: Measured by Defined Daily Doses (DDD), Days of Therapy (DOT), and Length of Therapy (LOT).
- Changes of Therapy: Tracking transitions between different antimicrobial choices, including de-escalation and IV to oral switches.
- Congruence with Prescribing Guidelines: Comparing first-line empirical therapy choices against local guidelines.
- Microbial Culture Taking: Assessing adherence to guidelines for collecting samples prior to treatment initiation.
The data processing was performed using SQL, R, and tidyverse, with the development of a prototype for the "Ramses" package.
Datasets
The study utilized pseudonymized electronic health records (EHRs) from adult patients admitted to six specialities (general medicine, respiratory medicine, geriatric medicine, cardiology, general surgery, urology) at Queen Elizabeth Hospital Birmingham (QEHB), UK, between September 2017 and August 2018. This dataset comprised:
- Patient demographics
- Clinical diagnosis codes (ICD-10)
- Clinical procedure codes (OPCS-4)
- Episodes of care (consultant code, speciality)
- Ward movements
- Key investigation results (blood counts, vital signs, etc.)
- Antibacterial drug prescription and administration records from the locally developed Prescribing, Information and Communication System (PICS).
- Microbial culture results.
A training dataset of 4200 prescriptions from 2712 patients, collected during a previous audit, was used for supervised classification to infer drug indications.
Key Insights and Ideas
- Automating AMS is Feasible: The most significant insight is that routine EHR data, when appropriately modeled, can be used to generate meaningful AMS metrics, moving beyond manual audits.
- "Therapy Episodes" are Key: The concept of reconstructing "therapy episodes" by linking prescriptions is crucial for understanding the full lifecycle of antibiotic treatment and enabling detailed analysis of changes and duration.
- Beyond Consumption Metrics: The study demonstrates the ability to measure aspects of AMS beyond just antibiotic consumption, such as adherence to guidelines and timely IV-to-oral switches, which are critical for optimizing care.
- Granular Monitoring is Possible: The developed methods allow for fine-grained monitoring of AMS practices down to the speciality, ward, and individual clinical team level, facilitating targeted interventions and peer comparison.
- Challenges in Guideline Congruence: While feasible, assessing adherence to prescribing guidelines can be complex due to data limitations (e.g., inferring indications) and the inherent ambiguity in clinical decision-making.
Key Results
- Therapy Episodes Characterized: The study analyzed 24,511 antibacterial therapy episodes, with a mean length of therapy (LOT) of 5.8 days and a mean DOT per admission of 8.7 days.
- IV to Oral Switch Analysis: Of the IV therapy sequences, 36% were converted to oral therapy. The average delay between meeting criteria for switching (ABCD criteria) and actual conversion was 3.6 days (95% CI: 3.4–3.7), indicating potential for improvement.
- Guideline Congruence for CAP: For low-severity community-acquired pneumonia (CAP), only 22% of therapy episodes were initiated with the recommended first-line drug, with a tendency towards broader-spectrum antibiotics.
- Microbial Culture Adherence: Only 22% of antibacterial prescriptions were preceded by a microbial sample taken within 3 days of initiation.
- Team-Level Variation: Significant variation was observed in IV-to-oral switch rates and time to switch across different consultant teams and specialities, suggesting opportunities for tailored interventions.
Limitations and Future Work
The authors acknowledge several limitations:
- Prescribing Indication Inference: The reliance on statistical classification for drug indications introduced potential errors, with an estimated 20% of CAP episodes potentially having a different indication. Future work could leverage more robust electronic prescribing systems that capture indications directly.
- "To Take Away" Medications: The analysis did not include take-away medications issued at discharge, which can extend total LOT.
- Free Text Data: Information recorded in free-text medical notes was not analyzed, potentially missing valuable insights.
- Case Mix Adjustment: While some metrics were computed at a team level, further research is needed to statistically adjust for case mix to enable more robust comparisons.
Future work suggested includes:
- Expanding the approach to include surgical prophylaxis, timely initiation of therapy, and adequate empirical coverage of microbial isolates.
- Developing interoperable software and standardized vocabularies to make these analytics accessible to a wider range of hospitals.
- Conducting user-centered research to tailor these metrics to the needs of individual clinical teams, AMS teams, and hospital managers.
- Developing evidence-based standards of care for antimicrobial stewardship, particularly for de-escalation decisions.
- The development of an open-source software package (Ramses) is underway to facilitate the implementation of these analytics in other hospitals.
Summary updated