Table 3.
Summary of recommendations on the application of intensive care big data.
| Part | No. | Recommendation contents | Consensus degree |
|---|---|---|---|
| 1. The concept, significance and necessity of intensive care data | 1 | The concept of intensive care big data: Intensive care big data refers to the datasets with logical connotations formulated by various indicators which are large-scale, multi-heterogeneous, variably dynamic, high-speed and real-time acquisition, low-value density and difficult to analyze traditionally in the whole process of diagnosis and treatment of patients or potential ones with critical symptoms | 97 |
| 2 | The intensive care big data is multi-modal, massive, dynamic, continuous, and objective, and its correct acquisition can provide auxiliary evidential support for diagnosis of critical illnesses and early warning | 98 | |
| 3 | The establishment of a large database for intensive care in China should follow the principles of multiple center, multiple disease and automatic capture, and provide reliable and accurate data support for the application of big data and the development of artificial intelligence | 92 | |
| 4 | Building a large database of patients with critical symptoms in China for their condition monitoring, the research and development of clinical drug and clinical trials can provide the standardized and individualized treatment for patients with critical symptoms | 97 | |
| 2. Clinical scientific issues concerned by intensive care big data in clinical research | 5 | It is recommended to use machine learning method to build modeling to make early warning of sepsis, acute kidney injury (AKI), and acute respiratory distress syndrome (ARDS) | 94 |
| 6 | The prediction model based on machine learning can effectively predict the risk of patients at high risk of potential organ damage in the ICU | 89 | |
| 7 | It is recommended to use machine learning method to build modeling to conduct early screening of hospitalized patients, so as to provide help for clinicians intervene early and reduce the severity of the disease | 88 | |
| 8 | It is recommended that the image data of patients with critical symptoms be included in the intensive care database to provide more comprehensive, accurate and timely diagnostic information, so as to guide clinical decision-making through relevant algorithms | 92 | |
| 9 | It is recommended to divide patients with sepsis, acute kidney injury, and acute adult respiratory distress syndrome into phenotypes with different clinical outcomes and treatment responses by means of cluster analysis, and identify patients who are most likely to benefit from specific treatment strategies | 91 | |
| 10 | In specific clinical scenarios, such as decision making for tracheal intubation and intensive care drug decision, it is recommended to build a decision-making model that can be used for clinical treatment based on machine learning algorithms | 74 | |
| 11 | It is recommended to use machine learning methods to predict the prognosis of patients with critical symptoms | 85 | |
| 12 | A clinical decision support system (CDSS) can be used to improve compliance with guidelines for diagnosis and treatment of patients with critical symptoms and the implementation of clinical pathways | 86 | |
| 3. Establishment, standards and principles of a large database for intensive care | 13 | It is recommended to build a intensive care medicine database and data analysis platform | 98 |
| 14 | It is recommended to form a standard normative intensive care dataset | 97 | |
| 15 | It is recommended to select automatic collection for objective data first. For data that cannot be automatically collected for the time being, targeted collection should be carried out in combination with research needs, data sources and data types | 92 | |
| 16 | It is recommended to establish a standard system for intensive care big data, standardize multi-center source data, and constrain standard codes, measurement units, field standards, as well as naming dictionaries to ensure the homogeneity and standardization of the use of the large database for intensive care | 95 | |
| 17 | It is recommended to establish a data security system to ensure the security of data storage, processing, sharing and use | 98 | |
| 4. Ways and methods to solve big data problems in intensive care medicine | 18 | It is recommended to use processing methods of digital signals such as filters to preprocess time series data, deep learning to process image data, use Natural Language Processing (NLP) technology to process unstructured text data | 93 |
| 19 | It is recommended to use resampling methods to deal with unbalanced datasets | 78 | |
| 20 | It is recommended to convert original categorical variables and numerical variables into variables that can be directly processed by machine learning algorithms through one-hot encoding, sequential encoding, etc. | 83 | |
| 21 | It is recommended to use dimensionality reduction methods such as principal component analysis to perform variable screening of high-dimensional features in intensive care datasets | 90 | |
| 22 | It is recommended to select supervised learning, unsupervised learning and reinforcement learning models for critical disease prediction and identification according to different scenarios and different data types | 97 | |
| 23 | It is recommended to use a causal inference model to explore and discover causal relationships in the intensive care field | 89 | |
| 24 | It is recommended to add external validation to internal validation of the model | 94 | |
| 25 | It is recommended to use indicators such as sensitivity, specificity, F1 score, and AUC to evaluate the performance of classification models, and indexes such as R2, MSE, RMSE, and MAE to evaluate the performance of regression models | 91 | |
| 26 | It is recommended to explore the interpretability of the model to facilitate the clinical transformation of complex machine learning models. The recommended model interpretation methods include Feature Importance, LIME, and Shapley | 91 | |
| 5. Clinical application of intensive care big data | 27 | It is recommended to transform and promote early warning tools that meet critical needs | 91 |
| 28 | It is recommended to use the information system for intensive care as a carrier to access real-time data and output recommendations for decision making | 91 | |
| 29 | It is suggested that the current practice of intensive care diagnosis and treatment should still be led by clinicians with the use of big data technology to coordinate to improve medical efficiency and ensure medical quality and safety | 98 |