Data quality for safer and more personalized perioperative care: a scoping review

Study protocol

A study protocol was developed in accordance with the Joanna Briggs Institute (JBI) methodology for scoping reviews and the PRISMA-ScR reporting guidelines. The protocol is available upon reasonable request. The completed PRISMA-ScR checklist is available as Supplementary Materials.

Eligibility criteria

We included peer-reviewed articles, reports, and frameworks addressing perioperative data quality. Eligible studies addressed one or more data quality dimensions such as accuracy, completeness, standardization, harmonization, or interoperability. We did not restrict the search to any specific perioperative setting, data type, or phase of clinical care.

Information sources

The literature search was conducted on PubMed, using relevant keywords (e.g., “perioperative”, “data quality”, “missing data”, “EHR”, “standardization”, “interoperability”).

Search strategy

The search covered literature published between January 1, 2000, and December 31, 2024. Additional filters restricted the search to English-language articles and excluded preprints.

The following query was used to search for relevant articles:
(“big data”[Title] OR dataset[Title] OR “remote monitoring”[Title] OR wearable[Title] OR “data standardization”[Title/Abstract] OR “missing data”[Title/Abstract] OR “data imputation”[Title] OR “data harmonization”[Title/Abstract] OR FHIR[Title/abstract] OR EHR[Title/Abstract] OR Common Data Models[Title/Abstract]) AND (anesthesiology[Title] OR preoperative[Title] OR perioperative[Title] OR postoperative[Title]) AND (excludepreprints[Filter]) AND (2000/1/1:2024/12/24[pdat]) AND (english[Filter]).

In addition to the primary search conducted, the references from the initial studies were examined to include relevant articles. This cross-referencing approach helped capture studies that were not retrieved through the database search. Articles identified by the authors as relevant were included.

Selection and charting process

Two reviewers independently screened all retrieved references. A third reviewer resolved conflicts. We charted relevant information, including data sources, identified quality issues, and proposed interventions. The selection of articles was based on the following criteria:

Inclusion criteria:
• English-language publications from January 2000 to December 2024
• Methodological papers, observational studies, registry reports, and narrative reviews
• Studies conducted in any perioperative phase (preoperative, intraoperative, postoperative)
• Focus on perioperative data quality, missing data, standardization, or harmonization
• Articles addressing data integration, interoperability, or common data models (CDMs) in perioperative care

Exclusion criteria:
• Non-English language publications
• Studies not focused on perioperative settings
• Articles that did not address any aspect of data quality
• Preprints and non-peer-reviewed publications
• Studies in which data quality was neither a primary nor a secondary outcome

The study flow chart is reported in Figure 1.

Figure 1. Study flow chart.

Synthesis of results

We organized findings into thematic clusters: (1) perioperative data sources and inherent challenges, including missing data and imputation; (2) standardization and interoperability frameworks; and (3) impact of data quality on patient safety and outcomes.

Sources of data in perioperative medicine

The operating room is among the most intensively monitored environments in the hospital. Each operating room typically includes ECG monitoring, pulse oximetry, cutaneous or core temperature measurement, invasive or non-invasive blood pressure monitoring, respiratory rate, and capnography (EtCO₂).

Monitoring the depth of anesthesia helps to avoid excessive use of hypnotics by relying on continuous electroencephalogram (EEG) analysis. Predictive models have also been used to avoid hypotension, alerting physicians several minutes before it becomes apparent. These monitoring tools enable anesthesiologists to detect and manage complications at an early stage. Figure 2 reports distinct types of monitoring systems, grouped by clinical setting. Continuous postoperative vital parameter monitoring can also be applied in standard wards. This is useful for detecting and preventing clinical deterioration in patients who are at risk for complications, even if they do not require intensive unit-level monitoring^[4-7]. In contrast, when surgical patients experience complications or need the highest level of assistance, they are admitted to the ICU. In the ICU, advanced monitoring techniques are used besides continuous monitoring, including calibrated pulse-contour monitoring devices and pulmonary artery catheters, which enable therapeutic interventions targeted at cardiac output, preload, and afterload. Both hemodynamic monitors and ventilators provide time-series data with high temporal resolution, ranging from 125 to 500 Hz. These data allow the development of computational models to guide optimal treatment decisions, including the use of vasoactive drugs and fluid therapy^[2].

Figure 2. Monitoring system used in the hospitals to collect data.

A promising advancement is the use of wearable devices (WD) to monitor vital parameters and detect clinical deterioration. Typically, vital signs are monitored intermittently in hospital wards (e.g., every 8 h), which may result in missed episodes of clinical deterioration in high-risk patients. In contrast, WDs allow continuous data collection, potentially reducing the necessity for ICU admissions and facilitating high-level monitoring in standard wards or even at home. WD could also be employed to replace parts of the traditional preoperative assessment, such as cardiovascular fitness assessment, relying on data collected by patients at home without direct clinician supervision. Another advantage of WD is that they enable patient mobility without the discomfort of being tethered to stationary monitors. In general, further studies are needed to assess and validate the role of WD in the perioperative setting, also considering how their low cost is facilitating their adoption^[7-11].

The electronic health record (EHR) system typically manually records monitoring data, clinical evaluations, radiology, and laboratory data. Originally designed as a tool for clinical documentation and billing, the EHR has evolved into a valuable tool for clinical research and quality improvement^[12].

Prior to the widespread adoption of modern EHR systems, most epidemiological data were derived from multicenter registries or healthcare claims databases, while other sources of data came from administrative databases for reimbursement purposes. This was particularly relevant when considering rare events. Singapore’s SingHealth Perioperative and Anesthesia Subject Area Registry (PASAR), one of these registries, had collected data on approximately 153,312 patient admissions, spanning from preoperative evaluation to postoperative treatment^[13].

More recently, the Multicenter Perioperative Outcomes Group (MPOG) has collected over 489 million medical records and 10.7 million anesthesia records from dozens of hospitals. The achieved numbers indicate the importance of multicentric collaborations in research and quality improvement projects. These datasets allowed studying postoperative outcomes such as opioid use, the efficiency of multimodal pain management strategies, and the efficacy of regional anesthesia^[14].

While these registries were created to collect data for scientific research and quality purposes, EHRs were designed primarily to document clinical information for internal use and thus may lack the quality, standardization, and harmonization needed for research use. Accordingly, we need to improve the quality of EHR data to ensure reproducibility, external validity, and data harmonization among different centers.

Data accuracy and consistency: automatic labeling of artifacts and false readings

A model can only be as accurate as the data on which it is based. Consequently, the quality of the data is crucial for allowing validation, reliability, and the concrete application of any artificial intelligence (AI) system in clinical practice. Different manufacturers, device specifications, and software versions within each class of devices result in differences in quality, resolution, and dictionary of collected data, as well as compatibility issues^[12]. Furthermore, the large volume of data, considering thousands of physiological observations per anesthetic record, can further complicate the analytical procedure^[1].

Physiological monitoring devices, such as arterial lines or pulse oximetry, are prone to mechanical distortions or interferences, leading to inaccurate readings. For instance, catheter damping can lead to an underestimation of blood pressure, while resonance effects can cause an overestimation^[12]. Very low or very high values are often caused by incorrect readings, detachment, and distortion, and while readily recognized by the perioperative team, these alterations can be difficult to recognize later in time when considering massive amounts of data. In other words, while clinical staff can often identify artifacts during real-time monitoring, retrospective detection of inaccurate reading - known as labeling - is labor-intensive and resource-consuming. Advanced algorithms can help mitigate these inaccuracies by filtering out noise and detecting the artifacts generated during data collection. This can be a significant contribution to the advancement of the analysis of high-frequency data^[3].

Another issue is related to the prevalence of missing or incomplete data, which constitutes a significant challenge^[15,16]. They may result in the introduction of uncontrolled biases, particularly when used in clinical research, leading to reduced performance of predictive algorithms. Missing data in EHR may arise primarily due to (1) technical issues, where the data does not appear in the system due to device errors or faults in data transmission; and (2) clinical issues, where the data are not entered into the EHR because they are considered unnecessary by the physicians or due to time constraints. Missing data can be classified into three categories, which influence the choice of analysis strategies:

• Missing completely at random (MCAR): where there is no systemic pattern behind the missing values, which could be related to observed or non-observed data.
• Missing at random (MAR): where the missing values are related to the observed data, not to the unobserved. Here, the probability of missingness is related to observed data. An example is when missing data or inaccurate monitoring data are present in less critical patients due to human intervention focused on more critical cases.
• Missing not at random (MNAR): refers to situations where the missing values arise from a consistent bias associated with unobserved data, which can influence the results.

Multiple studies have shown the impact of missing data on perioperative care and how these significantly impact the accuracy and further reliability of perioperative research, as well as patient recovery. In a study conducted by DeCrane et al., incomplete datasets introduced biases in postoperative cognitive disfunction (POCD), leading to misinterpretation of the true estimation of the associated risk factors related to POCD prevalence^[17]. Similarly, a recent study by Aziz et al. showed how missing data hinder the identification of risk factors in postoperative complications, which in turn affects the clinical decision making and patient outcomes^[18].

Selecting an appropriate imputation method is critical, as simple removal of records with missing values leads to data loss. Several approaches have been developed:

(a) Simple imputation methods:
These methods substitute the missing data points with single statistical estimates.
• Mean/median/mode imputation: These simple imputation methods use the mean, median, or mode to fill in the missing values. While computationally efficient, these methods reduce variance in the dataset and fail to maintain the relations between variables^[19].
• Forward fill (FFILL): This method replaces a missing value with the last non-missing value. Suitable for time-series physiological data and still computationally efficient, it may introduce bias in longer gaps^[20].
• Last observation carried forward (LOCF): As the name implies, this method uses the last known non-missing value to fill in all the gaps in the data. Effective for short-term missing intervals^[20].

(b) Advanced statistical and machine learning (ML) methods:
Advanced statistical methods fill in the data by estimating the uncertainty associated with the missing values, while ML approaches use predictive modeling to approximate the missing values^[21].
• Multiple imputation by chained equations (MICE): creates multiple imputations over a single imputation to account for statistical uncertainty^[22].
• K-nearest neighbours (k-NN) imputation: imputes missing values by identifying the k most similar instances (called “neighbors”) and using their value (mean, median, or most frequent value) for imputation^[23].
• Random forest (RF) imputation: demonstrates reliable performance for mixed-type data with the ability to capture non-linear relationships^[24].

(c) Deep learning approaches:
• Generative adversarial networks (GANs): operate on the principle of generative modeling, learning probabilistic data distributions through adversarial training with deep neural networks^[25]. It shows promise for complex multivariate time-series data^[26].
• Recurrent neural networks (RNNs): a neural network that learns from past data to estimate weights without requiring domain knowledge to predict new data^[27]. This model excels at capturing temporal dependencies in physiological data^[28].
• Variational autoencoders (VAEs): provide uncertainty quantification alongside imputation, which is crucial for clinical decision making^[29].

While imputing missing values with summary statistics is a simple and straightforward approach, the method is often inadequate for the more complex data, including intraoperative data^[30]. No single rule applies, so the choice of imputation depends on the pattern of missingness, the complexity of the dataset, the computational requirements, and the intended use of data^[31]. For real-time applications, summary statistics methods can be used for computational efficiency, while for research applications, other approaches provide better accuracy in quantifying uncertainty^[32]. When the proportion of missing data approaches 80%-90%, it is advisable to exclude the variable, since any imputation technique will introduce bias.

Data standardization and harmonization: challenges in interoperability and external validity

Data standardization is pivotal for consistent data collection across centers^[33]. In perioperative medicine, some efforts have been made to promote data standardization, such as the development of a minimum set of data for comprehensive geriatric assessments^[34].

Standardization allows merging data from multiple sources, including EHRs from different manufacturers, imaging and monitoring systems. It works by applying a common data structure, such as a predefined data dictionary, consistent units of measure, and predefined sampling rates. Harmonization goes a step further by aligning data from many sources for integration and analysis^[35]. To answer these needs, the Health Level Seven (HL7) standards and Digital Imaging And Communications In Medicine (DICOM) were proposed to grant interoperability^[36]. In 2011, HL7 created the Fast Healthcare Interoperability Resources (FHIR), a new standard aimed at improving how healthcare information is shared, using technologies such as Representational State Transfer (REST) architecture, application programming interfaces (APIs), XML format, JSON format, and Open Authorization tools^[37]. Similarly, the DICOM standard is designed to maintain interoperability in medical imaging across hospitals. In addition to creating standards for storage, DICOM also ensures uniformity in metadata and imaging acquisition protocols^[38].

As reported by Guglielminotti et al. (2015), consistency in reporting and methodology in multivariable analysis for prognostic observational studies is needed for reproducing research findings and increasing the reliability of research^[39]. Data integration and harmonization have also been reported to enhance the prediction of postoperative cardiac events from decision support systems^[40]. Standardization facilitates international data collection and enhances the quality of care by enabling benchmarking across institutions. An example is postoperative pain^[41].

Medical records include specific data types, from clinical observations to insurance claims, which vary widely in their organization, structure, and dictionary. Consequently, CDMs were created to merge data from sources and to store information systematically using predefined syntax^[42,43]. Several CDMs have been developed; among these, the most used are the observational medical outcomes partnership (OMOP), the sentinel CDM (SCDM), the national patient–centered clinical research network CDM (PCORnet CDM), and Informatics for Integrating Biology in the Bedside (i2b2)^[44-48].

In particular, OMOP has gained the largest diffusion. OMOP has been successfully employed in key projects such as the European health data and evidence network (EHDEN), which aims to harmonize patient data across European countries to improve patient care and medical research^[43]. It demonstrated superior performance due to the extensive vocabulary, which allows us to handle complex, heterogeneous data in healthcare^[49]. Table 1 presents the most widely used CDMs proposed to enhance the quality of data collection in perioperative care.

Health data are standardized to a CDM through an extract, transform, and load (ETL) process, which is defined by multidisciplinary teams comprising clinicians, data engineers, and data managers. In 2017, Ong et al. isolated 24 technical problems, categorized into six core topics, encompassing challenges from data sources (e.g., heterogeneity and accessibility) to data management (e.g., code maintenance and sharing)^[50].

Standardization of temporal resolution and image data

High-frequency data collected by monitoring devices have temporal resolutions on the order of milliseconds, reaching up to 500 Hz. This is necessary to visualize waveforms. Filtering techniques, derived from physics and mathematics, include low-pass and high-pass filters. These are often applied to preserve the most critical information and to rule out noise^[3]. Filtering, especially low-pass filters, has traditionally been used to reduce noise and improve data quality. With the advent of AI, filtering is even more necessary to handle data analysis even under high-noise conditions.

Several techniques have been developed to reduce noise in pictures, achieving sharper images, a principle that can also apply to complex medical signals. As for handling of missing data, the choice of filters depends on the type of noise and data addressed. Some techniques are:
• Low-Pass Filters: remove high-frequency noises while preserving low-frequency features such as general shape and contrast.
• High-Pass Filters: the opposite of low-pass filters.
• Median Filters: preserves edges using the median values.
• Mean Filters: a simple method used to remove less severe noise by taking the average pixel values.
• Weiner Filter: an adaptive filter that minimizes the difference between the estimated and original signal.
• Bilateral Filter: preserves edges by replacing the intensity of a pixel with the weighted average of the intensities of its neighboring pixels.

For medical images, specific types of noise and appropriate filters are:
• Gaussian noise: follows a standard bell-shaped curve, best addressed by the Wiener filter.
• Salt and Pepper noise: appears as black and white pixels on an image, and is effectively eliminated by the Median Filter.
• Speckle noise: exhibits a granular pattern, and needs Bilateral filters^[51,52].

The performance of these filters can be assessed using the peak signal-to-noise ratio (PSNR) metric.

Convoluted neural networks (CNNs) have emerged as leaders in the analysis of healthcare imaging data, achieving significant performance even in the presence of significant noise^[53]. An example is ClarifyNet, an end-to-end trainable CNN model developed by Susladkar et al. for dehazing photos^[54]. ClarifyNet was originally developed for computer vision applications and not specifically for perioperative care. Nonetheless, the underlying principles of noise reduction and image enhancement using high-pass (to detect sharp edges and finer details) and low-pass filters (for color and contrast) are highly relevant to noise reduction and image enhancement. CNNs have been used for noise and artifacts removal and have outperformed traditional techniques in published studies. They have been used to eliminate eye blink artifacts from electroencephalography signals^[55], while in the context of ultra-high-resolution photon-counting detector computed tomography (UHR-PCD-CT), they achieved up to 89% performance in noise reduction^[56].

Standardized dictionaries

An important aspect of data standardization mentioned above is the use of a standard dictionary as a common language across different centers. An example is the International Classification of Diseases (ICD) coding system, developed by the World Health Organization (WHO). This system uses alphanumeric codes for disease classification and reimbursement purposes^[57]. The ICD has been adopted for recoding diseases and sharing disease data across regions, thereby supporting benchmarking and evidence-based decision making.

The Standard Nomenclature of Medicine Clinical Terms (SNOMED CT) was deployed to provide a standardized clinical dictionary for different EHRs^[58]. In perioperative care, SNOMED CT standardizes the terminology used for surgical procedures, techniques, and complications. On the other hand, the logical observation identifiers names and codes (LOINC) standardizes terminology of clinical observations, including vital parameters and lab results, and is the preferred standard used by HL7 to attain interoperability^[59].

Other standards used include RxNORM for drug delivery systems, providing standardized drug names for pharmacy systems and drug interaction databases^[60]; current procedural terminology (CPT), which serves as the standard language for the communication of healthcare procedures^[61]; and Unified Code For Units Of Measurement (UCUM) for expressing quantities across domains ranging from medicine to business^[62].

There is ongoing collaboration among various organizations to make standards more consistent, interoperable, and user-friendly for both clinicians and researchers. SNOMED, which encodes patient clinical information, must interoperate with LOINC, which covers laboratory results, to ensure comprehensive EHR integration. This is achieved through the Unified Medical Language System (UMLS), providing semantic mapping and links between terminologies^[63]. There has been ongoing collaboration between the two organizations, SNOMED International and LOINC, for cooperative work in linking the terminologies and reducing redundancy^[64].

Impact of data quality on patient safety and outcomes

Safety and quality are pivotal aspects of perioperative medicine, both of which can be significantly influenced by the quality of collected data^[65]. A study by Fu et al. compared the quality of data extracted from two EHR systems in relation to the incidence of postoperative complications. The findings demonstrated that data quality can negatively impact patient outcomes and hinder clinical decision making. Similarly, inconsistent or incomplete documentation may lead to errors in therapeutic management^[66].

A Delphi study investigating the safety indicators in surgical patients showed that 74% of 73 key indicators considered crucial by the experts were related to the care quality, including effective monitoring and communication^[67]. Clear and effective communication is critical in perioperative care, especially during handovers from the operating room to the wards or between different intraoperative teams. The use of formal checklists has been shown to significantly enhance patient safety^[68]. Conversely, a lack of standardization in postoperative handovers has been associated with compromised postoperative care^[69]. Documentation errors are common and can also lead to complications, longer stay, and lower quality of life. Documentation is not only a clinical necessity but also a legal requirement. In a study on hand trauma cases, only 18.3% of the perioperative data were available for extraction, with some procedures exhibiting up to 38% missing data^[68]. High-quality documentation ensures that all relevant information is available for clinical decision making^[70], whereas omissions, errors, and inconsistencies can impair care delivery and complicate the investigation of adverse events^[71].

Such documentation and quality issues are also prevalent in open datasets. Initiatives like the Perioperative Risk Assessment Dataset (PRAD) integrate multi-view clinical data with pairwise comparison labels to improve the reliability of surgical risk prediction models^[72]. However, the effective use of these datasets for research is often limited by poor data quality, including inconsistencies and non-standardized formats^[73].

Challenges and future directions

As reported above, further efforts to ensure compatibility between different platforms are needed^[74]. One of the main challenges is related to regulations and privacy issues across countries, which hinder the harmonization of data from different jurisdictions^[75]. Regional differences in healthcare practice and in population can also affect the harmonization process. These can lead to ethical issues and socioeconomic bias, resulting in models being trained on incomplete, biased datasets that underrepresent certain populations. For example, a model trained on data from the US Veteran Health Administration may not generalize well to populations in other countries. To facilitate global intensive care data collection, initiatives such as the massive BlendedICU dataset have been proposed. The BlendedICU dataset merges data from prominent open critical care databases, including Amsterdam UMCdb, eICU, HiRID, and MIMIC-IV, and harmonizes them using the OMOP CDM format^[76]. Similarly, the INDICATE initiative by ESICM is designed to create a standardized ICU dataset across a network of European ICUs.

The creation process of these datasets reveals significant real-world challenges. First, the source databases exhibit high variability in data collection methods, recording formats, and treatment policies, reflecting diverse clinical practices and patient populations worldwide. To minimize bias and errors, merging these datasets requires meticulous expert mapping - a lengthy and resource-intensive task. Another example is the Chinese Anesthesiology Benchmark (CAB), a database developed to evaluate medical LLMs in a non-English context. Findings showed that even the top-performing models struggled to achieve clinician-level accuracy regarding safety. This challenge likely stems from the lack of high-quality, domain-specific datasets in non-English languages, posing an additional obstacle for multicentric datasets^[77].

Among future challenges, we should also consider the educational initiatives needed to bridge the gap between data science and clinical practice. A solid clinical understanding is essential to effectively leverage the vast amount of available data, while a strong grasp of data science is equally important to facilitate the adoption of data-driven tools and technologies^[1]. AI is increasingly integrated across medical fields, including anesthesiology, where it enhances patient outcomes by enabling real-time risk prediction and personalized perioperative care. As AI becomes a routine part of clinical practice, training healthcare staff to manage AI systems and to generate and maintain high-quality data is crucial for improving patient outcomes^[78]. The METRIC framework, proposed by Schwabe et al., is a 15-dimensional checklist designed to evaluate medical training data and could serve this purpose. This framework includes clusters that assess data quality dimensions such as Measurement process, Timeliness, Representativeness, Informativeness, and Consistency, enabling thorough evaluation of data prior to their use in training predictive models^[79].

Considerations for real-world application

This review highlights the need for standardized vocabularies (SNOMED CT, LOINC) and harmonization frameworks to assure data quality and facilitate collaboration across institutions. However, implementation remains challenging due to differences in institutional capabilities, regulatory requirements, and resource availability. The adoption of standard data models, particularly the OMOP CDM, represents a crucial step toward achieving interoperability across diverse data systems.

There is no universal solution to achieving high data quality and harnessing its full potential for efficient, sustainable data collection. Each center must weigh factors such as computational complexity, implementation costs, and optimal workflow integration, recognizing that different hospitals may require tailored approaches to data quality management.

We propose that this process can be enhanced through the formation of interdisciplinary teams comprising data scientists, informaticians, and healthcare professionals such as doctors and nurses. In this context, it is imperative to structure the training and education of healthcare professionals in data science, data analysis, and data quality management. Increasing awareness of international standards and regulatory frameworks within the medical community is equally important. Moreover, fostering and effectively disseminating international collaborations should be a priority.

Limitations of this review

This study did not include a formal quality evaluation of the reviewed studies, in line with the methodology of a scoping review. Although the initial search covered Google Scholar and Scopus, we ultimately relied on PubMed to maintain a focused and manageable scope specifically on perioperative data quality. This approach may have limited the comprehensiveness of literature identification and selection. Additionally, restricting sources to English-language publications may have excluded relevant non-English literature. Nonetheless, we think that the references provided by our scoping reviewer provide a solid foundation for understanding the current landscape of data quality in the perioperative context.

Conclusions

Data quality is a cornerstone for advancing clinical quality, patient safety, and research in perioperative care. Amid the ongoing data revolution, clinicians and researchers have access to an ever-increasing volume of data. This review underscores the importance of robust data cleaning, effective handling of missing data, and the adoption of standardization and harmonization strategies as critical components for both clinical care and research. Future initiatives should prioritize promoting standards, unified data formats and dictionaries, improving validation methods, and ensuring compliance with privacy and security regulations.