Starter Kit for Data Science in Perioperative and Intensive Care

Electronic medical records and artificial intelligence have facilitated big data research, but challenges remain in developing meaningful machine learning models. This review provides key information on open-access perioperative and intensive care datatablesets and data management tools.

With the adoption of Electronic Medical Records (EMR) and artificial intelligence, big data research has become more common. Nonetheless, there are several challenges when obtaining enough data to develop meaningful machine learning models: ensuring subjects' privacy protection, obtaining ethical approval by an institutional research committee, affording and deploying the human and technical resources for obtaining and storing data, and having enough local surgical or intensive care admissions. To overcome the difficulties of obtaining primary data, researchers may opt for utilising open-access datasets, especially for quick testing hypotheses. Researchers who have developed algorithms using their own population data may benefit from using open-access datasets to pursue external validation. This review aims to provide essential information about open-access perioperative and intensive care datasets and data management tools.

Dramatic Increase in Data Use and Collection

In perioperative and intensive care, accurate and rapid decision-making is essential to minimise adverse outcomes and achieve high care quality. With the constant advances in the EMR and patient monitoring and documentation, large amounts of clinical data have become available. These datasets are often referred to as “Big Data”. The key characteristics of Big Data are traditionally defined as the “4V”. These constitute Volume (vast quantity), Variety (multiple sources and methods), Velocity (high speed of generation and processing), and Veracity (high quality and value) (Shu 2016). Given the complexity of these datasets, equally complex analysis methods have emerged to interpret and draw findings.

One of the most powerful emerging ways to analyse massive volumes of data is the use of artificial intelligence techniques, such as machine learning (ML), which trains algorithms to find correlations and patterns in a relatively short amount of time. Nonetheless, obtaining datasets that are large enough to produce meaningful results is a challenge in itself.

Barriers to creating a dataset from zero are numerous. To ensure the protection of the subject's privacy, data must be de-identified and adhere to regulations such as HIPAA, and ethical approval from an institutional review board is typically required. Extensive computational, technological, and human resources are usually needed to obtain, store, and manage the data. Finally, a sufficient number of local surgical cases is mandatory to collect enough data in a reasonable timeframe.   

Some institutions and groups have collaborated to create combined databases, which allow members of the consortium to access a final dataset once they have provided a quote of patients of their own. Nonetheless, complying with the patient quotient requirements still represents a challenge for many researchers for the reasons described above. As an alternative to promote access to research, some groups share their de-identified databases in open-access repositories, which can be used by researchers to either develop or validate ML algorithms. Given the fact that databases are mostly not indexed by traditional search engines like Google, finding clinical repositories could be challenging.

This short review summarises and describes open-access datasets, dataset libraries, and data mining tools, to aid readers in starting their own journey in perioperative and intensive care big data.

Open-Access Perioperative Medicine Datasets

Table 1 contains datasets that are specific for perioperative/critical care. Table 2 provides other online libraries that contain several datasets on multiple medical specialties, including perioperative/critical care. Table 3 includes data mining tools with minimal coding requirements.

Clinical Applications of Big-Data Research

As proposed by Zhu et al. (2024), some of the clinical applications of big data analysis in perioperative care include the following:

Perioperative risk prediction tools: These establish the likelihood of complications or adverse outcomes during surgery and recovery. This could facilitate the identification of patients that demand adjustments to the anaesthetic, surgical, and/or recovery plan. These tools have the potential to inform decisions at a provider level (surgeons, anaesthesiologists, intensivists) as well as at a system level (for example, patient allocation in the ICU versus the general ward after the procedure).

Perioperative anaesthesia and intensive care decision-making: Complications in the intraoperative and postoperative periods, as well as in the ICU, are typically sudden and demand prompt and accurate decision-making. In addition, the fluctuating nature of a patient's health state in the perioperative period creates a challenging environment for adjustments in management; decision models tailored to inform decisions in real-time would be greatly beneficial. A promising concept in artificial intelligence applied to anaesthesia and intensive care is Closed-loop electronic medication management systems. This involves the automatic adjustment of medication dosages (anaesthetic, vasoactive, or fluid therapy) based on real-time analysis of multiple patient-specific variables (Austin et al. 2018).

In conclusion, this short review is intended as a resource to help researchers source data that may be useful to them as they develop and test analytics approaches for perioperative and intensive care use cases.

Conflict of Interest

None.