Fundamental procedures of the anesthesiology decision analysis platform

We constructed an Anesthesiology Decision Analysis Platform (Fig. 2A) equipped with perioperative clinical anesthesia-related data input; some of the fundamental procedures are as follows (Fig. 2B):

i.

raw data acquisition from medical systems, such as electronic hospital records, operations and anesthesia, laboratory information, and radiology information systems (the corresponding icon represents the Second Xiangya Hospital's associated medical system) (Table 1 lists the variables that need to be collected)

ii.

clinical data classification into categories and intermediate outputs. Clinical data including hospitalization information, postsurgical/anesthesia resuscitation records, clinical and biochemical data, and medical image diagnosis report;

iii.

normalization of data storage into major parts, such as basic patient information, principal diagnosis, surgical site and method, and narcotic drug use.

iv.

Creation of a big data platform and dataset that can be identified using registration numbers.

Fig. 2

Design of the Anesthetic Decision Analysis Platform, involving the working loop (A), architecture (B), fundamental procedures (C), and output workflows (D)

Table 1 Variables included in construction of anesthesiology decision analysis platform

The Anesthesiology Decision Analysis Platform aims to create links between artificial intelligence and anesthesia choices (or decisions) for surgical patients. Consequently, the design objectives of the cloud-based big data platform are twofold: first, to improve clinical decision-making at the system level by better analyzing data collected dynamically from diverse medical system sources; and second, to improve the flow of raw data generated from clinical settings for big data research.

The study was approved by the Research Ethics Committee of Second Xiangya Hospital, Central South University, Changsha, China (LYF 20240022).

Architecture of the anesthesiology decision analysis platform

The usage of large amounts of data requires the use of technological tools for data gathering from many sources and systems, as well as data transformation, storage, analysis, and visualization (Fig. 2C, Additional file 2: Table S1). Programming languages, such as C/C +  + , Python, Java, and Perl, are tools used to create purpose-built application programs through which instructions are issued to computers to achieve the desired objectives [18]. In addition, technological infrastructure for data storage is employed, which includes MySQL, HBase of Hadoop, and the Hadoop Distributed File System relational database system built on a Linux server [19]. Big data software, which facilitates the time-constrained processing of continuous information flows to produce actionable intelligence [20], is required for ‘data acquisition & processing’, ‘big data platform running’, and ‘system upgrade and platform security’.

Data acquisition & processing

Because medical records and information are now stored in disparate data formats, they are typically kept in comma-separated values (CSV) or table data format and can simply be stored on platforms utilizing these data formats [21]. The upload of data to the cloud is performed using DataX offline data synchronization software. Configuration environments for Python [22], Anaconda [23], and R language [24] are used for the analysis of intermediate output data. When data are uploaded into the cloud-based platform, the data collecting service begins ETL into a distributed database system [25]. After their upload into cloud storage, data are processed by a cloud-computing engine, such as Hive or Spark, distributed using the Azkaban System, and maintained by the Apache Atlas metadata management system [26].

Data retrieval and presentation

A data exploration portal (or simply a data portal) is employed for viewing, exploring, and downloading data from the platform, which uses the Presto data retrieval engine, a distributed open-source structured query language engine developed by Facebook’s Data Infrastructure Group. Apache Echarts [27] and DataViz [28] technology are used to meet the dynamic needs of visual interfaces and multidimensional visualization.

Upgrades & platform security

As the platform data accumulates, the Q-learning and Feedback algorithm is used to update and optimize parameters [29]. Furthermore, big data are said to be “new oil, but not clean oil,” an expression that seeks to warn that it can be both a critical driver of automation in the field of medicine and a source of information leakage. To maintain data storage security and compliance, the system employs Apache Ranger and Griffin software, which assures data disaster recovery capability and provides audit management and strategy analysis for the platform [30, 31].

Workflow of anesthesiology decision analysis platform

Anesthesia decision-making platforms developed with anesthesia big data input may be knowledge-based, with a sophisticated approach to causality algorithms, or non-knowledge-based (Fig. 2D), essentially designed to provide cognitive aids to the anesthetist [32].

The user logs into the big data platform and preoperatively enters the registration number of the case. The system extracts essential parameters while also detecting anomalous variables, ensuring anesthetic risks and drug compliance, and checking big data. Simultaneously, system identification recognizes and processes the relevant parameters, followed by information conversion and filtering. The filtering rules include the column, equivalent rules (> , = , and), age, sex, height, weight, and diagnosis. The system also detects anomalous findings based on the risk prediction results and whether the patient's clinical parameters exceed the threshold. In the present scenario, the normalization of interactive information triggers system scheduling and matching on a big-data-cloud platform. The platform performs and produces outcomes that match; the anesthetic plan is output directly if the platform matching is successful; otherwise, it is output using predetermined parameters with a confirmation or modification interface, followed by marking the modified cases and uploading the actual data in real time to the platform.

The output of the anesthesia strategy is generated by combining various physiological and preoperative data to generate advice, such as the precise identification of the appropriate use of narcotic and auxiliary drugs with the most recent recommendations and dosage schedules, smart alerts, and efficient gas administration. An anesthesiologist may accept or improve the output data. After completing an anesthetic case, the data are returned to the platform for data optimization and updating using Q-learning and Feedback learning algorithms [33]. To ensure quality control, our team involves three or more researchers in charge of data filtering to ensure that patient data are consistent, complete, and accurate. Data processing necessitates consistency and comparability when analyzing the same patient data from many sources. Time series data must have timestamps. Additionally, we sample and evaluate data regularly, anesthesia professionals review any incorrect results, and renew them to the platform following changes. Further, we provide an example of default parameter development, performance of machine-learning, and results’ interpretation in patients with oral cancer to make the platform understandable, because oral lesions significantly influence anesthetic decision-making in clinical practice (Fig. 3, Additional file 1: Fig S1, Additional file 3: Table S2).

Fig. 3

Example of default parameter development and results interpretation in oral cancer patients. A, B Variable shrinkage and selection using LASSO regression. C Nomogram of the model for predicting the wake tracheal intubation. D DCA curves were used to evaluate the model. LASSO Least absolute shrinkage and selection operator; DCA decision curve analysis; BMI body mass index; BSA body surface area; CCI Charlson comorbidity index. E A potential user interface presentation. ① User name; ② Password; ③ Log in; ④ Input of registration number; ⑤ Search in Platform; ⑥ Parameter identification; ⑦ Output of anesthesia strategy

Featured functions of the anesthesiology decision analysis platformCase retrieval

The scientific record retrieval system of the platform modified the previous process of obtaining medical records from various business systems [34]. The platform's system is based on clinical data centers and it can search for medical records that match the indicators from illness data marts using a hospital identification number, a single condition, or several combined diseases. The search results include information such as the anesthetic method and dosage, and expected surgery and resuscitation times.

Special case: anesthesia view

The anesthesia views’ special case shows all the medical data generated during the entire process of a specialized case, from the initial diagnosis and surgery to postoperative complications and hospital release. The information created by the case during diagnosis and treatment is displayed fully and accurately using the time axis and presentation dimensions of the type of medical information. The anesthesia view reflects more refined clinical reasoning throughout the perioperative period. Focusing on specialist scientific instances can demonstrate the scientific research value of medical information.

Self-service data analysis

Machine learning analysis of big data offers significant advantages for absorbing and evaluating enormous amounts of complex medical data [35]. The learning module of the Anesthesiology Decision Analysis Platform for parameter and model selection is a refinement and summary of traditional learning models that use various algorithms. The training data are used to screen key variables and construct the model, while the validation data are used to verify the results the training cases. Cross-validation and model evaluation indicators (including the F1 score, accuracy, precision, recall, area under the curve, and so on) are used to evaluate the performance of the models [36]. The best-performing models will be selected for further analyses, including feature importance evaluation, identification of key risk factors, and establishment of a comprehensive prediction model. Decision curve analysis and reasonability analysis are used to determine the net clinical benefit and interpretability of model [37].

To ensure that the platform is ready for clinical use as soon as possible, the developers (both anesthesiologists and software engineers) will validate it during the development phase, and then two clinical anesthesiologists provide real-time feedback during the trial process. When a given disease accumulates to a certain level, the platform will analyze the data again, determining the best model, and then testing and validating it once more. The platform applies classical machine-learning algorithms (logistic regression, random forest, extreme gradient boosting [XGBoost], support vector machines [SVM], k-nearest neighbors [KNN], and light gradient boosting machine [LightGBM], among others) to clinical anesthesia data and encapsulates them into a series of machine-learning service options, freeing users from the problems of complex manual operations, simplifying the research and analysis process, and improving scientific research efficiency [38].

Role of the Anesthesiologists in Big Data Platform Development

The widespread use of big data platforms will likely to improve computer-assisted human performance. Clinician–data interactions have previously been shown to improve decision-making [39]. Using big data, anesthesiologists may actively drive cloud-based platform advancements in anesthesia-related medical care, rather than passively waiting for the technology to become useful. First, because a lack of data can limit big-data platform predictions, anesthesiologists should attempt to broaden their involvement in perioperative data registries to ensure that all factors and patients are included. These can contain registers at numerous institutions, regions, and even at national and international levels. As data cleaning and processing techniques improve, registries may boost their utility and the availability of genomic, proteomic, and pathology data. As key stakeholders in adopting cloud-based big data technologies for perioperative decision-making and postoperative resuscitation, anesthesiologists should seek opportunities to collaborate with data scientists to explain how the big data platform can help decision-making with interpretable risk predictions [40, 41].

Furthermore, anesthesiologists can add value to data scientists by sharing their understanding of the relationship between seemingly simple topics, such as physiology, and more complex phenomena, like the dosage of narcotics or postoperative problems. These types of interactions are vital for accurately modeling and predicting clinical events, as well as enhancing the interpretability of cloud-computing platforms. In addition, anesthesiologists are ultimately responsible for making anesthesia decisions for patients and can establish a patient communication framework to relay the data made available by big data platforms and convey the results of complex analyses, such as risk predictions, prognostications, and treatment algorithms, to patients within the anesthesia decision.