The AI infrastructure of Chi Mei Hospital includes four major parts: original data sources, large medical databases, AI computing platforms, and an AI-enabled hospital information system.
-
Original data sources
Our main sources of big data come from our in-house hospital information system (HIS) including medical equipment and Internet of Things (IoT) data transfer and external data (such as the National Health Insurance database, the National Health Agency’s death data, health screenings, etc.). This will be expanded to include unstructured data (such as medical images, photos, audio files, text files etc.) in the future. Our HIS data is based on the IBM Informix database system.
-
Medical big data server
Chi Mei big data is divided into a large patient database and a large department-based database to facilitate the application of multivariate analysis. Large databases are first transferred to structured data formats (including medical orders, payments, claims, laboratory results, TPR sheets, nursing records, etc.). This process is adjusted regularly to meet the needs of researchers. After this data warehouse is established, a back-end AI operation management platform is available for modeling and application. The data warehouse is stored on a medium-sized Linux-based host with built-in MariaDB.
-
AI modeling platform
In recent years, due to the growth of efficient, parallel, multi-tasking computing capabilities, big data has brought about breakthroughs in accuracy. The explosiveness of deep learning algorithm research has led to the advancement and development of artificial intelligence in various fields. We first acquired high-end hosts and high-performance storage devices along with two sets of NVIDIA graphics processing units (GPU) to build an integrated AI modeling platform. We also utilize a professional modeling and management platform (InfusePrimeHub), which includes complete machine learning, deep learning and graphical tools, effective model building, version management and service delivery. Currently, the platform has already built many AI development tools to support commonly used algorithms. New algorithm tools will also be introduced, including: Logistic Regression, Random Forest, Support Vector Machines (SVM), K Nearest Neighbor (KNN), XGBoost (eXtreme Gradient Boosting), LightGBM (Light Gradient Boosting Machine) and multilayer perceptron (MLP), convolutional neural networks (CNN), etc.
In addition, we use the K-Fold Cross Validation to test and validate our algorithm results. Modeling quality evaluation indicators include traditional medical research indicators such as accuracy, sensitivity, specificity and area under the curve (AUC). In the end, we select the best model for clinical implementation.
-
AI cloud service platform
The other part of this platform is the AI cloud service platform. All AI models created are converted into programs and installed into our existing emergency HIS (health information system). These various medical AI predictive services can be accessed through our web host. We use a Service-oriented Architecture (SOA) point of view to provide appropriate cloud services. We currently have three types of cloud services to bridge between existing HIS and AI services:
-
HIS prediction interface web service (HWS):
This web service receives calls from the existing medical system (such as the emergency discharge assessment system) and sends the prediction results back to the clinician. HWS includes interactive features that allow physicians to adjust parameters for re-prediction.
-
Medical Feature web service (FWS):
This web service receives calls from the HWS, adds patient data such as the Glasgow Coma Scale, bedridden status, etc., then sends data back to HWS. FWS also includes the use of IoT technology to capture bedside physiological information (heartbeat, blood pressure, breathing rate, etc.).
-
AI disease-prediction web service (AWS):
This web service receives calls from the HWS, integrates data from the FWS and returns the results of AI prediction to HWS. For example, the chest pain AI disease prediction tool delivers two kinds of AWS: acute myocardial infarction risk prediction and death risk prediction.
-