Status of AI-enabled Clinical Decision Support Systems Clinical Implementations in China

AI-enabled Clinical Decision Support Systems (AI+CDSSs) were heralded to contribute greatly to the advancement of healthcare services. At present, there is an increased availability of monetary funds and technical expertise invested in projects and proposals targeting the building and implementation of such systems. Therefore, in this context of large funds and technical devotion, understanding the actual system implementation status in clinical practice is imperative. The objective of this research was to understand: 1) the current clinical implementations of AI+CDSSs in Chinese hospitals and 2) concerns regarding AI+CDSSs current and future implementations. A survey supported by the China Digital Medicine journal was performed. We employed stratified cluster sampling and investigated tertiary hospitals from 6 provinces and province-level cities. Descriptive analysis, two-sided Fisher exact test, and Mann-Whitney U -test were utilized for analysis.

While AI+CDSSs were not yet wide-spread in Chinese clinical settings, clinical professionals recognize the potential benefits and challenges regarding in-hospital AI+CDSSs.

Contributions To The Literature
Breakthrough emergence of AI and launch of AI-enabled applications draw attention, but often the status of their clinical implementation is less studied. Studies investigating the implementation status of AI+CDSSs in Chinese hospitals are scarce, and our work helps address this knowledge gap.
AI+CDSSs are yet to become widespread in Chinese hospitals and implemented are often not evaluated as most satisfactory. We explored the three most-common concerns and challenges which accompany the implementation of AI+CDSSs in China.
Findings of this work's China-specific analysis can help alleviate the concerns and challenges of developing and implementing a well-received AI+CDSSs.

Background
Since its birth in 1950s, artificial intelligence (AI) has been used to solve medical problems in the fields of medical management, clinical decision support, patient monitoring, and health intervention (1). In the 1970s, there was a wave of development of initial AI-enabled applications in the medical field, such as the MYCIN (an expert system for advising physicians regarding selection of appropriate antimicrobial therapy (2)) and the Quick Medical Reference (QMR) (a diagnostic decision-support system for internists (3)). After the "cold winter" during which interest in AI waned in the 1980s, in recent years AI has experienced an era of explosive growth and in the 21st century has advanced to a new stage of accelerated development fueled by an enormous increase in computing power and an even larger increase in data, in combination with improved AI methodologies. Many countries including U.S., China, Japan, India, France, UK, Germany, and others (4,5) have raised the development of AI to a level of government strategy, and recently, especially since 2016, a number of relevant policies have been issued to advance country-specific AI priorities. AI is widely seen as having a great potential to improve efficiency across many sectors, and healthcare is seen as one of the major and most important domains to be revolutionized by AI (6,7). From health-care industry perspective, AI-related initiatives have received increased funding and investor interest.
For example, by 2017, there were already 131 AI-focused healthcare companies in China, which attracted over CNY 18 billion in investments (8).
There is no single, universally agreed definition of AI. Broadly speaking, AI can be regarded as the emulation of human intelligence, which enables machines to learn from existing knowledge, adjust outputs, perform human-like tasks, and discover new knowledge (9). AI is generally classified into two types: 1) narrow (weak) AI which typically focuses on a specific task, or works within a narrow set of parameters; and 2) general (strong) AI which can learn to perform a multitude of tasks, or even refer to a sentient machine with elements of mind and consciousness. Narrow AI is widely applied in the healthcare field and uses computer algorithms to uncover relevant information from source data and to assist clinical decision-making (10) in order to have a positive impact on individuals, families, communities, and society as a whole (9).
Clinical Decision Support Systems (CDSSs) are software programs designed to utilize massive data, medical knowledge, and analysis engines with the goal to generate patientspecific assessments and recommendations to health professionals and to assist clinical decision-making through human-computer interaction (11,12). Such systems provide services ranging from reminders to complex risk-prediction algorithms (13). AI-enabled Clinical Decision Support Systems (AI+CDSSs), represent a narrow AI application (14), and refer to CDSSs that include an intelligent (AI) component. The basic design approach to AI+CDSSs is to combine the knowledge reasoning techniques of AI and the basic function models of CDSS (15). These system apply numerical, logical, and intelligent techniques in order to find out suitable solutions to medical problems as well as assist clinical decisions, and provide intelligent behavioral patterns with the ability to learn new clinical knowledge (16). Based on this, AI+CDSSs can outperform the traditional CDSSs and act as systems that support and improve the decision-making process by converting the raw medicallyrelated data, documents, and expert practice into a set of sophisticated algorithm inputs.
AI+CDSSs approaches are believed to: 1) aid physicians in their diagnostic efficiency; 2) improve clinical decisions accuracy and effectiveness, 3) reduce medical errors, 4) improve patient safety, and 5) decrease costs (15). In recognition of the advances, integration, and benefits of information technology in medicine, the FDA released its "Digital Health Innovation Action Plan" (17) in July 2017 to foster innovation in medicallyfocused digital devices and systems. Subsequently, FDA-approved AI+CDSSs products that are used for monitoring, early warning, or diagnostic assistance have emerged, and have been deployed to various medical facilities worldwide.
In this work we examine the implementation of AI+CDSS systems from the perspective of real-world practice in China. AI+CDSSs for triage or screening, diagnostic assistance, images interpretation, etc. have already been deployed and tested in some Chinese hospitals, however, many unanswered questions regarding the implementation status of AI+CDSSs in clinical practice still remain. With the goal to understand: 1) the current clinical implementations of AI+CDSSs, and 2) concerns regarding AI+CDSSs current and future implementations, we performed a web-based survey of Chinese hospitals' information technology (IT) department personnel combined with statistical analysis.
Specifically, we focus to uncover the percentage of hospitals that have deployed and tested such systems; and within hospitals which have tested and deployed such systemsthe current perspectives about the system and key issues and concerns that should be solved in system upgrade iterations or new deployments. To the best of our knowledge, this is the first such report, which provides a unique China-focused perspective, and helps address the knowledge gap regarding the AI+CDSSs status in China.

Methods
A web-based survey was initiated to understand: 1) the current clinical implementations of AI+CDSSs in Chinese hospitals and 2) concerns regarding AI+CDSSs current and future implementations. This study was informed by the Checklist for Reporting Results of Internet E-Surveys (CHERRIES) (18).

Survey questionnaire development and pre-testing
Before fielding the questionnaire, the appropriateness, usability, and technical functionality of the electronic questionnaire were tested through semi-structured interviews and a pilot survey (19). Investigators asked a pilot interviewee group, which included two health management experts, two medical informatics experts, and five prospective respondents, whether they understood the questions and whether they

Survey methodology and sample size calculation
We employed stratified cluster sampling, which is a feasible method considering the research goals and target population. We randomly assigned groups (provinces) from the Eastern area, Midlands, and Western area of China, and then surveyed tertiary hospitals within the groups. The areas were defined as follows: Eastern area (Shanghai, Guangdong), Midlands (Henan), and Western area (Qinghai, Chongqing, Yunnan).
According to the figures released by the statistical information center of the National Health Commission of the People's Republic of China, there are 2619 tertiary hospitals in China (June 2019). Hereby, the sample size was calculated as 184. Response rate of at least 70% is desirable for external validity (20).

AI+CDSSs stratification
To understand the details of actual AI+CDSSs practice in Chinese tertiary hospitals, we divided the AI+CDSSs into three types based on clinical diagnosis procedure stage: prediagnosis AI+CDSSs, diagnosis AI+CDSSs, and post-diagnosis AI+CDSSs. Pre-diagnostic AI+CDSSs are defined as software tools that are generally adopted in the process of previewing and preliminary checking, focused on medical history collection, risk factors screening, and ordering inspections and laboratory tests needed for diagnosis. Diagnostic AI+CDSSs are defined as triage or screening tools, which are used to assist clinicians in performing clinical diagnosis and detection by comprehensive assessment of patient information, including clinical phenotype, molecular genetics, reference knowledge, and other related data. Post-diagnostic AI+CDSSs are defined as systems that present recommended treatments to the clinician to select, alert for abnormal laboratory values, monitor medication safety, predict the possible health-related problems reflecting patient conditions and existing knowledge bases, generate and execute individual discharge plans and other post-diagnostic activities.

Survey quality control measures
We preset the logical relationship between questions and designated required questions in the electronic questionnaire. Therefore, logical consistency and completeness checks were ensured before the questionnaire was submitted by the responders. Before submitting respondents can review or change their answers. Furthermore, we asked the liaisons to send online notifications at least three times at regular intervals in order to improve the response rate. Three duplicate questionnaires were removed by considering the IP address and the name of hospital. Investigators verified the three duplicates: two of the duplicates were due to repeated submission, and one duplicate was a resubmission of a pilot test questionnaire.

Calculations and statistical analysis
Based on the survey results, we calculated summary statistics for the respondents (age, gender, professional titles, educational background). We also calculated summary statistics for the responding hospitals (area, grade, beds, number of medical doctors and nurses, number of patients). We also described the current situation of actual

Geographical and demographics characteristics
The survey's analyzable response rate was 86.96%. Survey responses were collected from a total of 160 respondents from 160 hospitals located in 6 provinces across China.
Geographical distribution of responding hospitals is shown on  (Figure 1).

Overall AI+CDSSs implementations
Less than a quarter of the surveyed hospitals (38/160, 23.75%) had deployed or tested an AI+CDSSs. There were significant statistical differences on hospital scale and volume (including hospital grade, actual opening beds, number of doctors and nurses, annual quantity of outpatient, surgery, and discharged patients) (p-value see Table 1) between the two groups of hospitals which had or had not an AI+CDSSs implemented.

Attitude towards deployed AI+CDSSs
The general attitude toward AI+CDSSs among the respondents was positive. On the 5point Likert scale, 81.58% (31/38) of respondents rated their overall satisfaction as 3 or 4.

Concerns regarding AI+CDSSs
We explored a total of 6 major concerns ( Table 4) and their resolution priority.
Respondents were asked to select which three concerns should be solved as an urgent need or in future implantations. The most common concern was system functions improvement and integration into the clinical process. Majority of the respondents (131/160, 81.88%) considered that this was the top-priority concern to be solved in existing or future implementation of AI+CDSSs. The next concern selected by over threequarters of the respondents was regarding data quality and data sharing mechanism improvement (124/160, 77.50%). The third concern selected by over one-third of the respondents (70/160, 43.75%) was methodological bias, including explainability, transparency, adaptability of algorithms ( Table 4).

Overview
Breakthrough emergence of new technologies and launch of new software products draw attention and publicity, but often the status of their clinical application is less studied.
There are global studies (22) about AI+CDSSs implementations, and healthcare professionals' attitudes and experiences of AI application in healthcare delivery. Studies investigating the implementation status of AI+CDSSs in clinical practice from Chinese perspective are scarce, and our work helps address this existing knowledge gap.
Our study found that AI+CDSSs are yet to become widespread in clinical practice in Chinese tertiary hospitals. Current systems were rarely embedded in multi-disciplinary or remote consultation systems and one of the most successfully executed functions of AI+CDSSs was imaging, followed by the previewing triage and risk factors screening. Our finding is consistent with a previous prediction that AI could foremost enter widespread use in medical imaging services (23,24). Chinese hospitals which have clinically tested or deployed AI+CDSSs tend to have higher grade, larger scale (actual opening beds, number of doctors and nurses) and bigger medical volume (quantity of outpatients, surgery, and discharged patients). A feasible explanation to this fact is that such hospitals may have larger capital investments into AI+CDSSs construction in order to meet their large needs and clinical workload.

Concerns and challenges
The implementation of AI+CDSSs in Chinese clinical practice are not yet widespread and implemented are often not evaluated as most satisfactory. Some underlying issues and constraints were already acknowledged in prior work (25,26). In our study we explored the prioritized concerns and challenges which accompany the implementation of AI+CDSSs in China. Herein we discuss: 1) system functions improvement and integration into the clinical process, 2) data quality and data sharing mechanism improvement, and 3) methodological bias, e.g. explainability, transparency, adaptability of algorithms, which were identified as the three most important areas for the surveyed group.
The top challenge for current and future implementation of AI+CDSSs identified in the survey is system functions improvement and integration into the clinical process. Our interpretation of this result is that the AI+CDSSs in Chinese hospitals, similarly to systems explored in other studies (27,28), may suffer from a mismatch between users' requirements/needs and system functions thus making the absence of translating users' requirements into system function design and clinical pathway integration is one of the major obstacles faced by AI+CDSSs implementation. Another challenge related to the implementation of AI+CDSSs highlighted by our analysis is the data quality and data sharing mechanism improvement. This is a common global problem where the lack high quality and available data creates a barrier for practitioners to receive the "right information" (26) from an AI+CDSSs system. The link between data quality and value can be illustrated by the Garbage In, Garbage Out (GIGO) phenomenon; data quality is also a prerequisite to data sharing. The third major concern about implementations of AI+CDSSs in Chinese hospitals is regarding algorithm bias, e.g. transparency, adaptability, and explainability. The lack of methodological transparency inherent in machine learning methods ("black-box") can impair user trust in the outputs produced by an AI+CDSS (29).
This is a phenomenon is distinctive in the healthcare field (22) where medical practitioners are less likely to adopt AI in clinical practice if they don't trust the technology or understand how it was used to support process of care or patient outcomes (30).

Perspectives
What can an AI+CDSS do and where will it work best, what are the clinicians' needs to deliver the best health care, and how to maximize the success of an AI+CDSS implementation in Chinese hospitals? While not an exhaustive recommendation set, the following key behaviors, informed by previous and this work's China-specific analysis, can help alleviate the concerns and challenges of developing and implementing a usable and well-received AI+CDSSs.
Perform comprehensive requirements definition involving all stakeholders (medical professionals, hospital administration, payers, government agencies, and patient groups) to ensure that system functionality addresses the clinicians' perception of greatest need and value for intelligent decision support.
Design the functions of AI+CDSSs in light of the multifaceted and complex workflow of the specific hospital setting where it is to be deployed in order to generate the right recommendations to the right person at the right time and location.
Develop and implement data standards, data quality measures, and data sharing mechanisms.
Technology and user rules should be combined to improve data-focused procedures such as storage, governance, patient record portability and security.
Ensure end-user training on system functionality and logic takes place among clinicians to explain how AI+CDSSs suggestions were reached in ways that are understandable to clinical users and reduce the 'black-box' perception.

Limitations of the study
There are two limitations in our study: non-response bias control and the the quantitative study design. Firstly, the non-response bias of this survey was not controlled and analyzed. In the implementation stage, the electronic questionnaires were received efficiently in the Midlands region (Henan), and the planned sample size was achieved.
Comparatively, the response efficiency was lower in the Eastern and Western area.
Therefore, investigators assigned more than one province in the Eastern (Shanghai, Guangdong) and Western area (Qinghai, Chongqing, Yunnan). In the data analysis stage, the impact of non-respondent bias was not analyzed due to the fact that detailed

Conclusions
In this work, we aimed to explore the gap between the excitement about the potential of AI+CDSSs and their actual implementation in Chinese tertiary hospitals. While AI+CDSSs were not yet wide-spread in Chinese clinical settings, hospital professionals recognize the potential benefits and challenges regarding in-hospital AI+CDSSs.

Consent for publication
Not applicable.

Availability of data and materials
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Competing interests
The authors declare that they have no competing interests.

Authors' contributions
All authors contributed to conception and design of the work, or acquisition of data, or analysis and interpretation of data; and drafting or revising the article. All authors have approved the submitted version; and have agreed both to be personally accountable for the author's own contributions and to ensure that questions related to the accuracy or integrity of any part of the work, even ones in which the author was not personally involved, are appropriately investigated, resolved, and the resolution documented in the literature.  Note: # Overall satisfaction with AI+CDSSs on a 5-point Likert scale, from 1 being "not at all satisfied" to 5 being "very satisfied".
* Mann-Whitney U test. Note: *Hierarchical diagnosis and treatmentOne of the medical system reforms in China which was aimed to allocate medical resources properly and to promote the equalization of basic medical services. Shanghai is the first pilot area. The main content of this reform is: 1) encouraging patients with common diseases to visit community health centers first, 2) redirect referral patients with chronic diseases or in recovery stage from tertiary hospitals to community health centers, 3) establishment of cooperation mechanism of different grades of hospitals to make full use of medical resources.  Figure 1 Characteristics of the study respondents (N=160) Figure 2 Clinical implementation of AI-enabled CDSSs AI+CDSSs embedded in consultation systems