AI refers to any computer software that can mimic cognitive functions or human intelligence. However, as the real mechanism of human cognitive functions has not yet been discovered, it is impossible to develop software that follows the mechanism of real intelligence; instead, scientists strategically developed software that appears to have cognitive functions regardless of its underlying mechanism.

AI is a concept with diverse subcategories defined by various principles (Fig. 2), classified according to whether it learns patterns from data or not. An example of AI that does not depend on data is the rule-based system designed by domain experts to handle specific problems well fitted to a hypothesis. In this case, the level of knowledge of the experts is important rather than the amount of data; however, this system is rigid and fails if there is a variation in the problem. The other approach called machine learning is to determine certain patterns from the data. This allows the software to change certain parts such as weights or coefficients in accordance with the inputted data.

Machine learning can be classified into three categories: The first is supervised learning for which two datasets must be prepared namely, input data for pattern training and information labels regarding the category or value of the input data (for example, normal vs cancer or survival vs death). Based on the information labels, the supervised learning algorithm tries to determine a pattern where the output is a category (classification problem) or real value (regression problem). By contrast, unsupervised learning is an algorithm that identifies inherent patterns in the data itself without labeled information. For example, clustering algorithms such as k-means or hierarchical clustering divides each data point into groups (usually the number of groups is predetermined by researcher) based on the distribution. Finally, reinforcement learning aims to set the best policy to solve a certain problem within the computationally designed space by trial and error.

Deep learning, the most advanced form of artificial neural networks, is an architecture that allows diverse methods to fit models to a given data by learning patterns in the data. It can be used for supervised learning, unsupervised learning, as well as reinforcement learning.

Deep learning is a part of machine learning; however, it has distinct characteristics as compared to traditional machine learning algorithms. First, in traditional machine learning, it is crucial to extract important features from the raw data by using domain knowledge; however, deep learning takes raw data as the input, extracts the features within the data, and learns the patterns by itself. Deep learning requires basic preprocessing such as normalization or noise removal; however, it lowers the barrier of entry for specialized domains of data, as it is relatively free from feature engineering based on domain knowledge. For example, suppose we need to develop a model that can distinguish between chest x-ray images of normal lungs and those afflicted by cancer. In earlier machine learning, the researcher had to extract features such as the shape of the nodule or distribution of colors that could be observed in images, and input that to the model. However, deep learning can directly extract the information that is relevant for discriminating between the two. Second, in traditional machine learning, the performance is usually restricted due to the limited predefined set of features and does not improve regardless of the amount of data used for training whereas deep learning requires a larger amount of data as the direct feature extraction process by researchers is omitted. The deep learning model can identify patterns clearly, when more data is added; hence, the diversity, clearance, and usefulness of the extracted features are enhanced leading to an improvement in the performance. Third, traditional machine learning provides the significance of the used features and information that have a major role whereas deep learning is difficult to interpret as it is similar to a black box. This may prove to be a barrier in adopting it for medical applications; however, it is now possible to understand deep learning indirectly through algorithms such as Gradient-weighted Class Activation Mapping [11]. This has led to an increased possibility of its use in the medical field.

Deep learning was particularly useful in visual problem solving in the early stages of its development. In image classification problems like ImageNet datasets, the convolutional neural network (CNN), one of several types of deep learning, showed an accuracy greater than 95% [12]. The CNN extracts features from images through filters that derive the morphological features from a small part (called window size) in an image. The CNN filter uses an element-wise product and weights, and the features from the whole image are analyzed by moving it (Fig. 3A). Learning through a CNN involves finding the appropriate filters and their weights to classify labels. For example, given that the shape of a nodule is a key feature to classify it as malignant or benign, a filter randomly initialized in the first stage cannot react to the specific shape; however, the CNN weights of the filter are updated after training to yield a strong response. The result extracted through the filter is called the feature map. It is possible to derive more abstracted features from the images when the CNN is applied deeply.

The deep learning designed to analyze time series data is the recursive neural network (RNN). Time series data usually have an autocorrelation where the previous value affects the next. To grasp the pattern of such time series data, it is necessary to reflect this autocorrelation in deep learning, for which both the past and current data must be considered. When the RNN analyzes the result in the Nth time step, the result is derived from the (N−1)-th time step as an input value. At this time, the output of the Nth time step is delivered to the next step i.e., (N+1)-th step. The information transferred across the time steps is called a hidden state, which is expected to remember the information of the previous time steps. From the above explanation, it appears that different neural networks are needed for every time step; however, the output from the RNN is transferred back to the same neural network (Fig. 3B). This is why it is referred to as an RNN.

A drawback in the basic RNN model is that the information of past data becomes blurred as the time step length of the analysis data increases. To address this, there are improved models such as long short-term memory [13] and gated recurrent neural network [14], which add a separate neural network to carry the information from past data. In current studies, RNN-based models are also applied with an attention mechanism [15] to provide clues that identify relevant context [12,16,17].

The autoencoder is another type of artificial neural network architecture consisting of an encoder that compresses the original raw data to a smaller size and decoder that restores the compressed information into original data (Fig. 4). It is classified as unsupervised learning as only original data is needed, without label information. It is also called representation learning in terms of identifying the intrinsic structure of data and extracting features from them. Identifying key information and compressing it into smaller dimensions is similar to the traditional methods of reduction of dimensions such as principal component analysis; however, the autoencoder has the advantage of representing more complex data spaces through non-linear functions.

The autoencoder tries to learn to minimize the reconstruction error that represents the difference between the original data X and reconstructed data X ', which is the output from the decoder. The information compressed by the encoder is the latent variable (h). If the original data is restored with the latent variable of a size smaller than the original data, it indicates that the encoder identified the pattern or internal structure of the data and extracted the main features. In such a case, the latent variable extracted by the encoder can be used as input data for other machine learning models like clustering or classification (Figs. 4B and 5).

h= f_encoder (X)

X'=f_decoder (h)

Reconstruction Error(Mean Squared Error)=‖X-X'‖²

As aforedescribed, deep learning is capable of finding patterns in complex data that are difficult for humans to recognize, and abstracting them into smaller data. To solve problems with less complexity, the existing statistical models or traditional machine learning also provides satisfactory results. However, for analysis of complex data, where the hidden features are abstract and difficult to extract, deep learning which can extract and utilize more diverse hidden information shows more accurate .

CDSSs are computer systems designed to support the decision of a clinician regarding individual patients. The primary purpose is to reduce errors by clinicians, such as prescribing drugs that may react adversely with those being ingested by patients or failing to conduct crucial tests. These traditional functions follow rule-based algorithms designed by medical experts. However, if AI is adopted in a CDSS, it can provide additional valuable information to increase the efficiency and quality of regular practice by predicting diverse clinical events such as mortality with reliable accuracy. Moreover, it can prove to be a novel digital biomarker that is valuable to make next-step decisions that the clinician might have missed.

The EWS is a similar system to aid the decision of the physician, though the focus is on the prompt notification of impending risk to ensure timely management to reduce the consequences of an adverse event. The algorithm used in current EWSs is based on simple rules; therefore, the accuracy is low and applicable area is limited. However, the ability of AI to extract valuable information not easily known to humans provides timely notification with greater accuracy and wider range of applicable areas.

These systems for supporting clinicians provide an opportunity to ensure high quality practice at low cost in advanced hospital care. However, they are based on existing medical records that are limited in detailed information compared to biosignal data. Applying AI to these systems based on biosignals can solve these problems as the number of hospitals that have systems for automatically recording continuous biosignals is increasing. The real-time biosignal recording systems will enable to fill the gap in information in regular medical records required for CDSSs or EWSs.

SaMD is software that can be used for medical purposes independent of a hardware medical device. Traditionally, software is a part of hardware; however, owing to advanced AI technology it is now a type of medical device supporting the clinician. AI models, such as diagnostic software for detecting diseases through radiologic images or ECGs, that can detect malignant nodules or cardiovascular disease are considered to be SaMDs. The US Food and Drug Administration (FDA) and Korea FDA have prepared the management and regulation of this software as a medical device [30,31].

MIMIC-III is a database collected from more than 40,000 patients in the ICU at Beth Israel Deaconess Medical Center in the United States from 2001 to 2012 [32]. It consists of demographics, biosignal measurements (including waveform data for some subsets), laboratory test results, procedures, drugs, caregiver notes, imaging reports, and mortality (in and out of the hospital). This large database is freely available worldwide and is the most popular database in biosignal research.

eICU database is one that includes data from 200,859 patients in 335 units of 208 hospitals in the United States from 2014 to 2015 [33]. It consists of biosignal measurements (vital signs but not waveforms), laboratory test results, drugs, acute physiology and chronic health evaluation (APACHE) components, care plan information, diagnostics, patient history, and treatment. It is publicly available and managed by the same research group as the MIMIC-III database.

The Ajou University Hospital Biosignal database contains data of patients who were in the ICUs or emergency rooms of the Ajou University Hospital in Korea since August 2016 [34]. Until November 2019, this large database consisted of more than 19,000 patients with all the biosignal waveform data and vital signs during the period of the patients' admission to discharge from the hospital. All the biosignal data were linked with the electronic medical record data. An automated system collected data from approximately 140 bedside patient monitors continuously. The database is not publicly available yet; however, data access is possible by conducting collaborative work with the research team.

VitalDB database is a biosignal database consisting of data from 6388 patients who underwent surgery in 10 operating rooms at Seoul National University Hospital from 2016 to 2019 [35]. It includes biosignal waveform data, vital signs during surgery, various clinical parameters related to surgery, and laboratory test results.