Conducting Retrospective Studies, Audits and Chart Reviews: A Practical Guide for Clinicians

Abstract

Background/Objectives: Retrospective projects including audits and observational research advance the practice of emergency medicine but face methodological challenges affecting data quality. This guideline presents an 11-step framework to guide the conduct of high-quality retrospective projects, minimizing bias and enhancing reproducibility for clinicians. Methods: The stepped approach mirrors the standard sections of a study protocol but reframes them as guiding questions to make each section’s content and purpose more practical, intuitive, and clear for users. Conclusions: This framework equips clinicians with a practical entry point to retrospective study design, distilling methodological nuances and strategies to bridge theory and application. Systematic adherence promotes rigor, reduces bias, and elevates retrospective chart review from a convenient tool to a robust method for evaluating practice patterns, interventions, and quality improvement in emergency care. Implementation also fosters a culture of evidence-based inquiry essential to advancing emergency medicine.

1. Introduction

Rigorous systematic analysis of existing medical records has advanced clinical research, enhanced healthcare knowledge through methodological rigor, and improved patient outcomes through evidence-based practice [1,2,3]. The use of historical data is well established in epidemiology, quality improvement and health services evaluation, professional education, and emergency care research [1,3,4,5,6,7,8,9,10,11]. The main advantage of this design is its easy access to existing data, eliminating the time and cost of collecting new data. Chart reviews also allow clinicians to address clinical questions that are uncommon, difficult or unethical to study in prospective or randomized trials, such as those involving rare or harmful exposures [12,13,14].

Retrospective studies including audits, chart reviews and observational research use existing clinical data in different ways and for different reasons. Clinical audit formalizes the longstanding tradition of healthcare professionals reviewing case records to improve practice through structured, data-driven evaluation. By measuring care against specific standards with the goal of quality improvement through an iterative cycle, clinical audit is an indispensable tool for advancing patient safety and service quality in emergency care [15]. Retrospective observational research, on the other hand, aims to generate new knowledge or test hypotheses and often requires ethical approval and the use of inferential statistics [2,12,16]. Such designs therefore suit estimation of prevalence, identification of associations, and hypothesis generation. Due to the retrospective nature, establishing causality or temporality between exposures and outcomes is often impossible. Chart reviews are a data collection method used in both clinical audits and retrospective observational research, depending on the study’s intent [5,12].

The purpose of this article is to present a practical guide for conducting high-quality clinical audits and retrospective observational research in the emergency care setting.

2. Key Steps for Performing a Retrospective Project

An 11-step process (

Table 1. Key steps of a retrospective project and underpinning questions.

Step	Prompting Question	Study Protocol Heading/Phase Descriptor
1	What is the question?	Aims, Objectives, Outcomes
2	How long will this take?	Timeline
3	Who is helping?	Team composition
4	What is already known?	Literature Review
5	Who should I include?	Inclusion and Exclusion Criteria
6	How many samples should I use?	Sample Size, Power Considerations, and Sampling Strategy
7	Where do I find the data?	Data Sources and Data Collection Tools
8	Do I have permission to do this?	Ethical and governance considerations
9	How do I make the data usable?	Data Collection and Cleaning
10	What does the data mean?	Data Analysis, Figures, and Charts
11	What does the data really mean?	Significance and Limitations (see step 11 for a checklist)

) is described as a practical and accessible starting point (or reminder) for clinicians undertaking retrospective studies. Achieving high-quality results requires a systematic and well-organized approach, best represented as a study protocol. A study protocol (sometimes referred to as a study plan) may seem unnecessary, but it is essential for maintaining a systematic and transparent approach to any project or audit. It provides prescriptive details that serve as a reference for the team throughout the study and when replicating the work in the future. The stepped approach mirrors the standard sections of a study protocol but reframes them as guiding questions to make each section’s content and purpose more practical, intuitive, and clear for users. If one seeks to publish, the Standards for Quality Improvement Reporting Excellence (SQuIRE) guidelines or STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines should also be used [17,18]. Appendix A provides a worked example of these steps.

2.1. Step 1: What Is the Question? Aims, Objectives, Outcomes

A common approach to prioritizing audit focus is to select problems that are high volume, high cost, or high risk, ensuring that resources are allocated to areas with the greatest potential to improve safety, effectiveness, and patient experience [19]. Once an area of focus has been decided, aims, objectives and outcomes must be clearly articulated. Aims provide a broad statement of the overall purpose. For example, an audit might seek to determine whether clinical practice aligns with current standards for fractured neck of femur care [20]. Objectives are specific, measurable steps in achieving the aim, and in the previous example, might serve to assess time to analgesia administration within 30 min of diagnosis. Outcomes represent the expected results, with primary outcomes being the main measure of success (for example, proportion of patients receiving analgesia within 30 min) and secondary outcomes measuring additional metrics (such as pain score reduction or time to orthopedic review).

Clinicians often risk deviating from their original aims and objectives by attempting to address multiple questions. It is therefore crucial to focus on a single primary outcome, which serves as a clear point of reference to maintain direction and coherence throughout the process. As you go through the steps, your primary outcome may need to be adjusted based on the availability and feasibility of collecting relevant data.

Once the aims, objectives, and outcomes have been drafted, a feasibility check is necessary.

Table 2. Common questions to consider if audit and/or chart review is a feasible method.

Does the topic lend itself to the audit process? Are reliable sources of data readily available for data collection purposes? Can data be collected within a reasonable time period? Will someone on the project team be able to interpret the results and identify limitations?

poses some questions to assess if moving to the next step is worthwhile, or if refining the aims, objectives and outcomes are necessary.

2.2. Step 2: How Long Will This Take? Timeline

It is important to align project expectations with practical constraints with limitations often related to resourced time and data availability (see Step 7). Gantt charts provide a structured means to establish realistic timelines, particularly when prior experience is limited [21]. They enable accurate estimation of required durations, align team expectations, and offer a comprehensive visual roadmap of project demands.

2.3. Step 3: Who Is Helping? Team Composition

Even small-scale clinical audits benefit from a dedicated project team. This may simply include you and a supervisor or mentor or can be quite comprehensive depending on your experience and study methodology. The composition should reflect the multidisciplinary nature of healthcare delivery, incorporating staff with complementary expertise and practical knowledge of the system under review. In addition to identifying members of the study team, an outline of individual responsibilities is also essential.

When quality improvement is central to the study, a multidisciplinary group directly involved in relevant care processes should participate, either as study team members or stakeholders. This inclusive approach provides operational insights critical for implementing practical service enhancements in addition to establishing validity with the data collected [16].

2.4. Step 4: What Is Already Known? Literature Review

A comprehensive review of the existing literature is an essential foundation for any study, including retrospective studies. The purpose of the literature review is to ensure that the planned study is contextually grounded, scientifically justified, and methodologically informed by prior research. An additional advantage of the review process is that it may provide reproducible methods (including defining and measuring key variables) and statistical approaches to reduce the burden of having to create this de novo. For quality improvement studies, reviews may also offer standards of care for which to benchmark against [1,6,19,22,23,24,25,26].

Investigators unfamiliar with search techniques may collaborate with medical librarians who can provide expert guidance on database selection, phrasing of search terms, reference management, and useful tools such as large language models (artificial intelligence) [27].

2.5. Step 5: Who Should I Include? Inclusion and Exclusion Criteria

Retrospective studies utilize data originally generated for medical record-keeping, rendering data collection inherently challenging due to inconsistencies in documentation and completeness [6,11]. Explicit inclusion and exclusion criteria are therefore essential, with justifications underscoring their selection to ensure that captured data accurately reflects the intended study population and purpose.

Inclusion criteria define the key characteristics of the target population that eligible medical records must possess to address the study question (

Table 3. Common inclusion and exclusion criteria themes.

Inclusion Criteria	Exclusion Criteria
Hospital(s) Department(s) Age range Gender Time range Defined population	Duplicate records Missing documentation Conditions that may confound results

). These criteria ensure homogeneity among selected charts, enhancing the study’s internal validity and generalizability to the intended patient population.

Exclusion criteria, on the other hand, are often used incorrectly and expressed as the opposite of the inclusion criteria. Exclusion criteria should instead be defined as criteria that should be excluded from the population identified from the inclusion criteria. For example, a study investigating techniques to remove foreign bodies from the eye may have included a patient who had a complication with their contact lens, which was not considered within the scope of the study aims and objectives and therefore should be excluded from the final population.

2.6. Step 6: How Many Samples Should I Use? Sample Size, Power Considerations and Sampling Strategy

A representative sample is often required to allow inferences about the broader population. Two fundamental questions inform how well represented the data is: what approach will ensure that it accurately represents the study population (sampling technique) and how large should the sample be (sample size)?

2.6.1. Step 6.1 Sampling Techniques

Broadly, three sampling techniques exist and include convenience, quota, and systematic sampling (

Table 4. Description of sampling techniques and their use cases.

	Description	Use Cases
Representative Sampling
Simple	The population eligible for inclusion are selected in such a way that each has an equal chance of being included in the sample.	When the population is the same or highly similar for the characteristics that are key to the objective
Stratified	The population eligible for inclusion are divided into groups (such as age), with a random sample selected from each group.	When the population is not the same or highly similar for the characteristics that are key to the objective
Systematic (interval)	The population is arranged in a sensible order, with the first person eligible for inclusion selected at random, and every person that falls at a fixed interval thereafter is selected.	When the population is the same or highly similar for the characteristics that are key to the objective
Non-representative Sampling
Purposive	The population eligible for inclusion are selected for specific purposes.	When the population is known and a small sample will suffice
Convenience	The population eligible for inclusion are selected because you can get them relatively easily.	When you do not want to generalize the findings to a population and you want a manageable sampling method
Quota	Subgroups of a population are identified and a desired number of the population eligible are sought until the quota for each subgroup is achieved.	When you do not want to generalize the findings to a population and you want a manageable sampling method

). Choice of technique is often based on feasibility of probability-based sampling and available resources but is also dependent on the study’s objectives and epidemiologic frequency. For emergency medicine, convenience sampling is the most frequently used approach and involves selecting all eligible cases within a defined time period or data availability window. Be careful to account for seasonal and temporal factors (such as large workforce changes at the start of each year). Quota sampling entails identifying a predetermined number of cases from specific subgroups to ensure that critical strata are adequately represented. In systematic sampling, investigators select every nth record from a complete list of eligible cases, yielding a quasi-random distribution of observations.

More novel sampling techniques include interval sampling (beginning with a small sample and expanding only if results are inconclusive) and rapid-cycle sampling (using small sample sizes at increased frequency).

2.6.2. Step 6.2 Sample Size

The required sample size depends primarily on the desired confidence in the findings and the available resources for data collection. In audit studies, sample size calculations frequently involve comparing proportions of patients who meet predefined care criteria before and after implementation of an intervention. A general rule of thumb for audits is to sample at least ten cases or events per variable to produce results that are both reliable and clinically interpretable [28]. This “ten-per-variable” rule is commonly applied in regression modeling and other multivariable analyses [29,30].

2.7. Step 7: Where Do I Find the Data? Data Sources and Data Collection Tools

Clinical data is frequently dispersed across multiple media and stored in different departments or even separate institutions. The required information is often unavailable in an analyzable manner and requires collecting, linking, and/or cleaning. At other times, the data may be missing or no longer exists.

2.7.1. Step 7.1 Data Sources

Many existing health information systems already contain data suitable for audit and quality-improvement purposes. Hospital and departmental management systems commonly capture performance and process measures and should be the first port of call for understanding what routinely collected data is available for extraction [3,10]. Understanding what data is available will also assist in understanding feasibility of the study.

2.7.2. Step 7.2 Data Collection Tools

Once an understanding of what data is routinely available and extractable, customized data collection tools should be used. This can be in paper form or electronic form. A popular electronic form often used in research is RedCap and should fulfill most ethical and research governance requirements [31]. When designing a data collection tool, it is important to define the type of data (numerical or categorical) you will be collecting as this will inform the statistical analysis (see Step 10). Outlining a table or graph by creating a ‘dummy table’ can often help contextualize what critical information is required and save you the effort of having to review records for a second or third time to collect additional data. Standardizing the collection options and minimizing free text will also help when it comes time for data analysis.

It is important that data collection tools are informed by a detailed abstraction manual or coding guide. This guide should provide explicit coding rules and decision criteria for each data element to ensure consistency across abstractors. When reporting study methods, clinicians should describe these coding procedures clearly and, when feasible, include the coding guide and data dictionary as appendices to enhance transparency and reproducibility.

Once the tool has been created, testing the tool with either dummy data or a few cases (10 cases is ample) will help determine usability. Conducting a pilot review before full data collection can help identify areas prone to missingness and inform feasible strategies to mitigate its impact. If pilot testing was not performed, this omission should be acknowledged as a study limitation.

2.8. Step 8: Do I Have Permission to Do This? Ethical and Governance Considerations

Prior to collecting data, it is important to understand ethical, legal and organizational requirements. For internally focused audits and chart reviews, organizations may have established processes addressing both ethical and governance requirements. Approval is often delegated to departmental heads and may range from something as simple as verbal confirmation to a formal application, including study plan.

Should you wish to publish, most academic journals will require an ethical review process that may include a waiver or procession through low or high-risk pathways. Whilst an ethical review seeks to protect the patient, research governance adds an additional layer of protection for the organization, including compliance with local legislation. Most organizations will require research governance approval once ethical approval has been granted. More progressive organizations will have both ethical and research governance reviews in a single application. In some countries (such as Australia), ethical approval from only one ethics committee is enough to cover multiple sites and organizations [32].

2.9. Step 9: How Do I Make the Data Usable? Data Collection and Cleaning

2.9.1. Step 9.1 Data Collection

Among the most significant threats to validity in chart review research are biases introduced during the data collection phase. Standardized instructions, abstractor training, and regular meetings between investigators and abstractors to resolve ambiguities and reinforce coding rules with periodic monitoring should be utilized to minimize this bias.

The most common method to quantify the degree of potential bias is to perform a test for interrater reliability, which is the degree of agreement between two or more independent abstractors examining the same record [6]. There is no formal standard that dictates what proportion of records should be extracted by two independent abstractors, although 5% is often considered adequate [33].

Depending on the type and scale of the variables, reliability may be expressed using Cohen’s kappa for categorical measures or an intraclass correlation coefficient (ICC) for continuous data. The acceptable degree of interrater reliability is set by the study investigators and, as a rule of thumb, an 85% agreement could be considered the minimum while 90–95% is preferable [6,34,35]. Revising the methodological approach or even the study aims should be considered if interrater reliability falls below 70% [34].

2.9.2. Step 9.2 Data Cleaning

Following collection, all data should undergo systematic cleaning to ensure completeness, accuracy, and internal consistency. Prior to cleaning, it is important to preserve the original, unaltered dataset in its entirety, with any cleaning or manipulation performed on duplicate copies. Data cleaning involves detecting and correcting errors (e.g., age 999), inconsistencies (e.g., merging Y and Yes responses), duplicates, or missing values in abstracted data to ensure analytical reliability.

Missing data can introduce bias, particularly if cases with incomplete records differ systematically from those with complete data [13,14,36]. Two common strategies used to address missing data during cleaning include deletion and imputation. Deletion is more commonly used and excludes the case and/or variable if data is missing and often defined as an exclusion criteria. Although straightforward, deletion can substantially reduce sample size and exacerbate selection bias [37]. Imputation can handle missing data in a variety of ways such as replacing missing data with the mean of the data available (mean substitution) or fitting the missing data to a statistical model such as a maximum likelihood ratio or Bayesian Monte Carlo simulation [38]. Imputation methods are more commonly used in large electronic datasets and rely on the assumption that data are missing at random [14].

2.10. Step 10: What Does the Data Mean? Data Analysis, Figures, and Charts

Data are often analyzed through two broad approaches: descriptive and inferential statistics [39]. Descriptive statistics summarize and describe the characteristics of a dataset, such as measures of central tendency and range, without making any predictions (such as associations). Descriptive statistics are often enough for audits and chart reviews. They are also a necessary step when conducting inferential analysis and, in this context, are often referred to as summary statistics. Inferential statistics use sample data to draw conclusions, test hypotheses, measure association, or make predictions about a larger population (see step 11) [40]. Inferential calculations are often categorized as parametric or non-parametric. Parametric calculations require that the data have a normal distribution and are often more accurate than non-parametric.

2.10.1. Step 10.1 Proportions

Proportions are another measure that can be either descriptive or inferential, depending on context. For example, a finding where 25% of 100 ED patients were diagnosed with sepsis is descriptive, summarizing the observed data. For the same cohort, a population proportion estimate can be calculated such that a finding of 25% (95% CI: 17–33%) of ED patients were diagnosed with sepsis.

For audits that compare performance against a predefined standard, results are usually presented as proportions or percentages. Both the absolute numbers and corresponding percentages should be reported; for example: adherence n = 40/50 (80%). When sample sizes are small, decimals beyond a single significant figure should be avoided to prevent false precision.

2.10.2. Step 10.2 Descriptive Statistics

Selecting the correct measure of central tendency (mean, median, mode) requires categorization of the data type (continuous, discrete, ordinal, nominal) and distribution of the data (normal versus skewed). Based on these values, the correct method can be calculated (

Table 5. Approach to central tendency and variance.

	Mean	Median	Mode
Suitable distribution	Symmetrical (normal)	Skewed	Skewed, Double Peaked
Suitable Data Types	Continuous, Discrete	Continuous, Discrete, Ordinal	Any
Calculation	Add all values and divide by their count	Arrange data in order and choose the middle value	Value most frequently counted
Calculate variability	Standard Deviation	Interquartile Range	Minimum, Maximum

). Calculating the variance is also important and should be based on the calculation used (standard deviation for mean, interquartile range for median and minimum/maximum for mode).

Data can be described as four types within two categories (

Table 6. Categorization of data.

Category	Type	Description of Data	Example
Numerical (Quantitative)	Continuous	Numeric values that can take any value within a range	Length of Stay
	Discrete 1	Numeric values that are integers	Number of presentations
Categorical (Qualitative)	Ordinal	Categorical data with a defined order	Pain score
	Nominal 2	Categorical data without inherent order	Blood Type

¹ Also known as count. ² Includes binominal.

). The type of data will determine the approaches to both descriptive and inferential statistics. Although most types are intuitive, confusion can often arise when transforming ordinal data into numerical form for analysis. For example, pain scores can be mild, moderate, and severe (ordinal) and are converted to numerical values where 1 is mild, 2 is moderate and 3 is severe. The numerical scores are still ordinal (categories with a defined order).

Once the data type has been identified, the next step is to determine the type of distribution of the data. There are a variety of ways to determine distribution, with the simplest being plotting a histogram and determining symmetry through visual opinion, which is often enough for smaller studies and audits. Statistical confirmation of distribution can be performed through a variety of tests, including Shapiro–Wilk and Kolmogorov–Smirnov, but lack precision at sample sizes less than 20 or more than 50 [41,42].

Calculating the central tendency is often the focus when summarizing a study outcome. The type and distribution of data identified in the previous steps determines the most appropriate calculation of central tendency and their equivalent assessment of variability (Table 5). Note that for nominal data, rather than central tendency, a frequency (%) is expressed.

2.10.3. Step 10.3 Inferential Statistics

Determining the degree of association and generalizing to population estimates requires more advanced statistical interrogation. Like descriptive analysis, the statistical test is reliant on the type of data, number of groups, and type of distribution. Figure 1 describes a flow chart for commonly used statistics based on these variables.

2.10.4. Step 10.4 Figures

Figures can be effective in describing the data more than (or in conjunction with) statistical analysis. It is important to use the right graphical representation method for your data, which includes a concise but descriptive title, axis labels, and legend. A summary of when to use each method based on type of data is provided in

Table 7. Common figures used based on data type.

Data Type	Common Figures	Strengths
Continuous	Histogram, Box plot, Scatter plot	Visualizes spread/skewness/outliers; compares groups; detects trends/correlations
Discrete	Bar chart, Histogram	Emphasizes discrete gaps between values; shows modes/clusters; simple for counts over categories
Ordinal	Bar chart, Box plot	Reveals order/ranking; compares medians/spread across ordered groups; identifies outliers
Nominal	Bar chart, Pie chart	Clearly shows category counts/percentages; easy to compare relative sizes; highlights dominant groups

.

2.10.5. Step 10.5 Control Charts

For longitudinal monitoring, and, in particular,, audits and quality improvement studies, control charts are often used to distinguish normal variation from significant process changes (Figure 2) [43].

2.11. Step 11: What Does the Data Really Mean? Significance and Limitations

2.11.1. Step 11.1 Association, Strength, and Causation

Association describes a statistical linkage between variables where they co-occur more frequently than expected by chance, with common analysis of association provided in Figure 1. The strength of these associations is also important in appreciating any clinical significance that may not have statistically significant association (for example, when the sample size is low). For binominal data (such as those seen in case–control studies) and logistic regression analysis, strength can be quantified through relative risk and odds ratios. For continuous data with a linear relationship, correlation calculations measure the strength of association. None of these measures prove effective. Causation is proof of effect and demands rigorous evidence of temporality (exposure precedes outcome), dose–response gradients, biological plausibility, and consistency across studies. This typically requires prospective or randomized designs rather than the retrospective snapshots common in audits, where concurrent data collection confounds inference [2,6,37].

When reporting results from retrospective data, interpretations must also remain proportional to the strength of the underlying evidence. It is appropriate to describe observed changes in adherence or performance but not to attribute these changes to causal factors unless supported by a study design capable of establishing cause and effect (and statistical analysis confirming it).

2.11.2. Step 11.2 Probability of Results: Confidence Intervals and p-Values

Inferential statistics must include a probability of the results being correct. This is often referred to as the p-value, with a threshold of p < 0.05 conventionally used to denote statistical significance. This value is arbitrary and should not replace sound scientific judgment. Moreover, p-values are strongly influenced by sample size where larger samples can produce small p-values even when observed differences are not clinically meaningful [14]. p-values should always be presented alongside a corresponding confidence interval (CI) to assess the practical importance of findings, emphasizing whether the plausible range of effect sizes includes values that are clinically relevant.

When studies use sample data to estimate population parameters, CIs serve to quantify the precision and reliability of those estimates. The 95% CI is conventionally reported, indicating that if the study were repeated many times, 95% of the calculated intervals would contain the true population value. Narrower CIs denote greater precision and are influenced primarily by sample size and data variability.

2.11.3. Step 11.3 Statistical Versus Clinical Significance

The p-value remains one of the most frequently misinterpreted and overutilized statistical measures [40]. Statistical significance does not necessarily imply clinical relevance as it may reflect an effect too small to influence patient care. Clinical significance should instead be judged by the magnitude of effect and its capacity to change care. In the absence of established benchmarks, clinically important differences can often be context-specific.

2.11.4. Step 11.4 Limitations

Despite their value, retrospective designs are often undervalued owing to concerns over data quality and interpretability. Each step in the pathway from patient interaction to the final dataset introduces potential for error. Patients disclose information, which clinicians must understand correctly and record accurately in medical records. If clinical documentation is inaccurate, no amount of analysis will be able to compensate. Abstractors then interpret these records and extract key variables using standardized tools. Abstractors can be prone to fatigue and misinterpretation, especially when dealing with large volumes of complex medical records. Errors such as omissions, misinterpretations, or miscoding at any stage can accumulate, causing systematic biases that distort outcome measures from true effect estimates [9,12,13,37]. Recognizing these limitations is essential when analyzing and reporting audit or retrospective observational data to ensure conclusions accurately reflect the study design’s capabilities. Implementing strong methodological controls to detect and reduce bias is therefore crucial for producing reliable and high-quality results [9,44]. A core tenant set out in the 11 steps for performing a retrospective study is to create meaningful data that is robust and minimizes the limitations of such a methodology [45].

Table 8. Checklist of potential biases and limitations of retrospective methodologies.

Is there a conflict of interest with investigators? Can it be managed? How? Was the sampling strategy representative? Were data variables clearly defined? Was a data collection tool used? How was missing and conflicting data handled? Were abstractors blinded to the study aims? 1 Was there abstractor training and monitoring? Was interrater reliability measured? For which variables?

¹ This is often difficult as investigators are often the data abstractors.

summarizes these potential biases and limitations that can also act as a checklist.

3. Conclusions

This framework serves as a practical and accessible starting point (or reminder) for clinicians undertaking retrospective audit or research. It provides structured direction and integrates experiential insights drawn from established practice to offer guidance on methodological nuances, common pitfalls, and strategies that bridge conceptual understanding with real-world application. In doing so, it aims to demystify the audit process and equip clinicians with the foundational knowledge and confidence required to initiate high-quality retrospective projects.

When performed according to these principles, retrospective methodologies become both an efficient and feasible approach as well as a scientifically robust method for addressing clinically relevant questions in health care. Far from being a secondary design, it remains an essential tool for characterizing practice patterns, evaluating novel interventions, and driving evidence-based quality improvement that shapes the ongoing evolution of health care.