kimi总结
这篇文章探讨了人工智能（AI）和机器学习（ML）在预防和管理静脉血栓栓塞症（VTE）方面的潜在用途，以及临床医生和医疗信息专家对此的看法。以下是文章的核心内容：
背景：
VTE是住院患者中可预防死亡的主要原因。
AI和ML可以支持个性化风险评估和预防措施的指南推荐。
研究方法：
进行了两项横断面电子调查，一项针对临床医生，另一项针对医疗信息专家。
调查由多学科团队设计和审核，并通过REDCap（一个安全的网络应用程序）进行管理。
调查结果：
共有101名信息专家和607名美国临床医生参与调查。
大多数信息专家认为AI/ML可以用于管理VTE（56.0%）。
超过三分之一的信息专家担心临床医生不会使用这项技术（38.9%）。
大多数临床医生认为AI/ML可能或肯定可以帮助预防VTE（70.1%）。
两组最常见的担忧是感知到的缺乏透明度（信息专家54.4%；临床医生25.4%）。
讨论：
尽管信息专家担心临床医生可能不愿意使用AI/ML，但大多数临床医生调查受访者认为AI/ML在VTE预防和管理领域是有帮助的。
透明度是临床医生和信息专家考虑AI/ML系统时的主要担忧。
解释性（interpretability）是AI/ML领域的一个重要概念，它指的是向人类以可理解的方式解释或呈现模型的能力。
结论：
大多数临床医生和医疗信息专家认为AI/ML可以帮助VTE管理。
两组都提出了关于AI/ML系统验证、临床设置中的部署方式以及临床判断重要性的问题。
需要对特定的AI/ML系统进行进一步研究，以了解它们如何影响临床护理。
数据可用性：
数据可以根据要求提供；有兴趣审查数据的研究人员应联系Rushad Patell。
资金来源和利益冲突：
研究由美国疾病控制和预防中心（CDC）支持。
文章中还列出了作者的利益冲突声明。
附加信息：
文章提供了补充材料的链接，并说明了如何获取材料和版权信息。
这篇文章提供了对AI/ML在VTE预防和管理中应用的初步看法和潜在障碍的见解，并强调了透明度和解释性在医疗领域中的重要性。
Abstract
Venous thromboembolism (VTE) is the leading cause of preventable death in hospitalized patients. Artificial intelligence (AI) and machine learning (ML) can support guidelines recommending an individualized approach to risk assessment and prophylaxis. We conducted electronic surveys asking clinician and healthcare informaticians about their perspectives on AI/ML for VTE prevention and management. Of 101 respondents to the informatician survey, most were 40 years or older, male, clinicians and data scientists, and had performed research on AI/ML. Of the 607 US-based respondents to the clinician survey, most were 40 years or younger, female, physicians, and had never used AI to inform clinical practice. Most informaticians agreed that AI/ML can be used to manage VTE (56.0%). Over one-third were concerned that clinicians would not use the technology (38.9%), but the majority of clinicians believed that AI/ML probably or definitely can help with VTE prevention (70.1%). The most common concern in both groups was a perceived lack of transparency (informaticians 54.4%; clinicians 25.4%). These two surveys revealed that key stakeholders are interested in AI/ML for VTE prevention and management, and identified potential barriers to address prior to implementation.
Similar content being viewed by others
Prediction of venous thromboembolism with machine learning techniques in young-middle-aged inpatients
Article Open access
18 June 2021
The potential of artificial intelligence to improve patient safety: a scoping review
Article Open access
19 March 2021
Preoperative prediction of the need for arterial and central venous catheterization using machine learning techniques
Article Open access
13 July 2022
Introduction
Venous thromboembolism (VTE) is the leading cause of preventable death in hospitalized patients1,2. This has long been recognized as a major public health challenge, with multiple organizations calling for increased awareness and thromboprophylaxis of patients3,4,5. Thromboprophylaxis is known to be safe and effective6, but numerous studies have shown that it is still misused or underused6,7,8,9.The first step in appropriate VTE prophylaxis involves identifying patients at risk for VTE. Prior guidelines recognized the burden of individual risk assessment and recommended a cohort-based approach to VTE prophylaxis5. However, these types of universal strategies have shown minimal impact on reducing VTE rates10,11. Contemporary guidelines emphasize the heterogeneity of hospitalized medical patients and recommend an individualized approach to risk assessment and prophylaxis12.
Multiple risk assessment models (RAMs) have been developed over the years to assist clinicians in risk stratifying patients. These RAMs are effective tools for identifying at-risk patients13, and some studies have shown that incorporating a RAM into the clinical workflow can reduce VTE events14,15. However, there are dozens of validated RAMs with variable risk predictors, making it difficult to determine which RAM a clinician should choose, and in which populations. An ideal RAM would be applicable to diverse patients and recognize that the risk of VTE can change over the course of a hospitalization. The RAM would be easy to use and consider both the risk of bleeding and clotting. These lofty goals make new technologies such as artificial intelligence (AI) and its subbranch, machine learning (ML), particularly promising in the field of VTE prevention. AI/ML is an area of computer science that trains computer algorithms to make predictions based on available data. An effective AI/ML tool would search the medical chart for relevant risk factors, weigh the risk of VTE as well as the risk of bleeding, alert the clinician when that risk changes, and assist the clinician in making a decision about appropriate VTE prophylaxis.
A recent systematic review and meta-analysis of AI/ML tools for VTE prevention showed overall high performance16, but few AI/ML-based models have been implemented in clinical settings17. More research is needed to understand the safe and effective implementation of AI/ML 17. Understanding stakeholder attitudes is an important part of the path to implementation. In this paper, we present two surveys, one directed at clinicians and the other at informaticians, to better understand their views on AI/ML in the field of VTE prevention and management.
Methods
Survey design
We conducted two cross-sectional electronic surveys about perceptions on the use of AI for VTE prevention among clinicians (Supplementary data 1) and among healthcare informaticians (Supplementary data 2). Both study instruments were designed and administered in REDCap, a secure web-based application18. Both were vetted by a multidisciplinary team with expertise in hematology, informatics, and public health, and refined for clarity, length, and relevance with a survey expert. They were individually approved by the Beth Israel Deaconess Medical Center IRB as exempt protocols and all research was performed in accordance with relevant guidelines and regulations. Informed consent was obtained at the start of each survey via a statement about the voluntary nature of the study.
We defined VTE as deep vein thrombosis (DVT) and/or pulmonary embolism (PE). We defined prophylaxis as including pharmacologic and/or mechanical measures for VTE prevention, and we used the terms prophylaxis and prevention interchangeably. In the healthcare informatician survey, we used the term “blood clot” instead of VTE because not all respondents may be familiar with the term VTE. We defined AI/ML as technologies that mimic human intelligence, and gave an example of technology using a large dataset to train a model that can predict outcomes or help make decisions. We used the terms AI and ML interchangeably.
The clinician survey explored attitudes and practices around VTE prevention, with a specific section exploring the use of AI/ML for VTE prevention (3 questions). We report the findings of that section here. The healthcare informatician survey explored attitudes and practices around AI/ML in healthcare in general, with a specific section on the use of AI/ML for VTE prevention (21 questions total). We report the findings of the survey in its entirety here.
Survey items included both quantitative and qualitative questions. Quantitative questions used Likert scales and Yes/No/Do not know responses. Several questions provided respondents with a field to write a free-text response. Both surveys asked respondents about general demographics such as age, sex, ethnicity, role, and time in practice (Supplementary data 1 and 2).
Respondents
The surveys were administered to targeted convenience samples between August 2021 and December 2022. Clinical respondents were recruited through social media, professional organizations, and program directors across the United States (US). Clinical trainees could opt to enter a raffle drawing for one of 100 $10 gift cards. There was no incentive provided for other groups. Healthcare informaticians were recruited through professional organizations and a publicly available database listing recipients of National Institutes of Health informatics grant awardees. Due to these methods of recruitment, a response rate could not be accurately calculated. As a descriptive survey utilizing a sample of convenience, we did not perform any a priori sample size calculations.
Analysis
We used descriptive statistics to present responses and respondent characteristics, summarizing data as proportions. Binomial 95% confidence intervals (CI) was included for primary outcomes. We used the chi square test to compare categorical responses between groups. We performed all analyses using SAS, version 9.4 (SAS Institute, Cary, NC, USA).
All clinicians were asked if AI/ML can help with VTE prevention in hospitalized patients; respondents who answered anything other than “definitely yes” were asked what concerns they have about AI/ML and provided a list of options derived from the literature, along with space to write a free-text response. Two researchers (BDL and RP) conducted a qualitative analysis of these answers to batch responses that matched existing categories and to categorize new themes derived from the data. All conflicts were settled by consensus.
All healthcare informaticians were asked if AI/ML can help with management of blood clots. Respondents who answered yes were asked what aspects of blood clot management AI/ML could be useful for. All respondents were asked about potential barriers and what else to consider when using AI/ML to assist with clinical management of blood clots. Two researchers (BDL and RP) conducted a qualitative analysis of these answers. All qualitative analysis for both surveys was done by reviewing responses, developing a codebook, and using the codebook to independently code all free text responses. Themes were derived from the data. Up to two codes were assigned to each response, and the researchers compared their codes and reached consensus for every response. All qualitative coding was completed with Microsoft Excel (Microsoft Corp., Seattle, WA, USA).
Results
Healthcare informatician—respondent characteristics
Of 101 respondents to the healthcare informatician survey, a majority were greater than 40 years old (54.5%), male (62.0%), and white (70.1%) (Table 1). Respondents could identify as more than one role; most were clinicians (44.6%), followed by data scientists (36.6%). Most respondents reported that they had been practicing in informatics for more than 10 years (54.5%).
Table 1 Healthcare informatician demographic characteristics.
Full size table
Most felt they were very well informed (41.6%) or sufficiently informed (48.5%) about AI/ML, with a majority reporting that they had taken coursework on the topic (61.4%) or were doing research on the topic (68.3%). A large portion also reported that they had worked on deploying AI/ML (42.6%).
Clinician survey—respondent characteristics
Of 607 US-based respondents to the clinician survey, a majority were 40 years old or younger (69.9%), female (55.7%), and white (68.2%) (Table 2). Physicians made up 70.7% of respondents, of which 45.4% were trainees. Hospital medicine was the most common specialty (52.1%) followed by hematology (20.8%). A majority of respondents (65.8%) reported making a decision about whether a patient needs VTE prophylaxis every day. Only 20.5% of respondents reported that they had used AI/ML to inform their clinical practice; a majority had not (57.9%) or were unsure (21.6%).
Table 2 Clinician demographic characteristics.
Full size table
Healthcare informatician—experiences with AI/ML
Most informaticians (62.6%) reported that their organization is using or developing AI/ML for healthcare. Of these 62 respondents, a majority described the status of AI/ML at their organization as implemented with at least one model in use (82.3%), and that their organizations primarily develop the models themselves (81.4%). Less than half (45.8%) reported using third-party vendors or partnering with local universities (28.8%).
Respondents who reported developing AI/ML systems used Python (76.6%), R (45.3%), and toolkits (42.2%). Of the respondents who described what toolkit they prefer, the most commonly cited ones were Scikit-learn and TensorFlow.
Healthcare informatician—attitudes towards AI/ML
A majority of informaticians agreed that AI/ML can have a positive impact on the care of patients (95.0%), can help healthcare organizations meet regulatory requirements (95.0%), and can have a positive economic impact on their healthcare organization (81.0%) (Fig. 1, Supplementary Table 1). Informaticians mostly agreed that AI/ML has the potential to perform better than humans (76.3%) and will replace human employees in some jobs (60.4%). Respondents found AI/ML to be overall reliable (58.5%) and would trust their own care to an AI/ML system (49.5%). However, less than half would trust a closed proprietary system (39.7%). Most informaticians agreed that AI/ML should be independently vetted and standardized prior to use in a clinical setting (96.0%), regulated (95.6%), and evaluated in randomized controlled trials (81.2%).
Figure 1
figure 1
Informatician attitudes towards AI/ML (n = 101).
Full size image
The three most common reasons respondents identified as barriers to the successful development of AI/ML in healthcare were data quality (67.3%), lack of standardization (39.8%), and difficulty of acceptance by healthcare providers (35.7%).
Healthcare informatician—attitudes towards AI/ML for the management of blood clots
A majority of informaticians agreed that AI/ML can be used for clinical management of blood clots (56.0% 95% CI 46–66%). Of these 56 respondents, most agreed that AI/ML can be used for risk stratification (94.6%), radiologic accuracy (87.5%), surveillance (80.4%), diagnosis (73.2%), and treatment (73.2%) (Fig. 2). Four respondents proposed potential additional uses for AI/ML: monitoring the process of clot dissolution, warfarin dosing, shared decision-making, and treatment during acute versus chronic recovery phases. All respondents were asked about perceived barriers, and the most commonly cited barriers were a lack of transparency with AI/ML systems (48.5%), concern that clinicians would not use an AI/ML system (34.7%), and concerns around liability (24.8%) (Supplementary Table 2). Informaticians who identified as clinicians were more likely to think that AI can help with VTE compared to those who did not identify as clinicians, though the difference was not statistically significant (66.7% vs. 47.3%, respectively; p = 0.052). Respondents at organizations that have implemented AI were not significantly more likely to think that AI can help with VTE compared to those at organizations that had not implemented AI (59.0% vs. 48.7%; p = 0.32).
Figure 2
figure 2
Potential applications of AI/ML for management of blood clots (n = 56).
Full size image
All respondents were asked the free-text question, “What else should be considered when using AI/ML to assist in the clinical management of blood clots?” There were 37 responses, six of which were removed from analysis because they did not answer the questions. Of the remaining 31 responses, nearly all were related to validation of the system and discussed factors related to testing, bias, and transparency. Several responses discussed deployment, and a few responses touched on the importance of clinician oversight (themes in Table 3; coding tree in Supplementary Table 3).
Table 3 Informatician recommendations regarding AI/ML for management of blood clots.
Full size table
Clinician attitudes towards AI/ML for the management of blood clots
Most respondents (70.1%, 95% CI: 66.6–73.6) believed that AI/ML probably or definitely can help with VTE prevention. Respondents who had used AI in clinical practice were more likely to agree that AI can help (RR: 1.45, 95% CI: 1.33–1.58) compared to non-AI users. Respondent age and physician training status were not associated with the belief that AI can help (both p ≥ 0.05).
Respondents who answered anything other than “definitely yes” to whether AI/ML can help with VTE prevention were asked about their concerns. The three most common concerns were a lack of transparency (25.4%), that AI/ML is not ready for clinical use (25.0%), and that AI/ML is not accurate enough (22.7%). Seventy respondents provided free-text responses. A total of 34 comments were removed (16 responses were removed for comments such as “I don’t have enough knowledge in this area,” 13 responses were removed for comments such as “I have no concerns,” and 5 responses were removed for comments requesting more information about a specific AI/ML system.) The remaining 36 comments were categorized into the following themes: the importance of clinical judgment, concerns around how the AI/ML system is validated, and concerns around deployment of the system and the possibility of technology fatigue (themes in Table 4; coding tree in Supplementary Table 4).
Table 4 Clinician concerns around AI/ML for VTE prevention.
Full size table
Discussion
Although informaticians expressed concern that clinicians would not be willing to use AI/ML, a majority of clinician respondents in our survey believed that AI/ML can be helpful in the field of VTE prevention and management. This is consistent with surveys evaluating clinician attitudes towards AI/ML in other specialties; generally, clinicians have reported that AI/ML will have a positive impact on their field19,20,21. Surveys have found that some specialists may be less likely to perceive benefits from AI/ML22. As AI/ML technology advances and becomes more accessible, further research into clinician and other stakeholder views is critical in order to understand how AI/ML can positively and safely impact care.
Both clinicians and informaticians cited transparency as their number one concern when considering AI/ML systems for the management of blood clots. Though transparency has been cited as a barrier in other clinician surveys, it was not the main concern. As AI/ML moves closer to the bedside, concerns around the “black box” nature of AI/ML may increase. In a landmark paper from 2017, Doshi-Velez and Kim discuss how different measures of success have emerged as AI/ML becomes more ubiquitous23. Initially, models were optimized for performance, but now there are efforts to optimize models for other important criteria such as safety and nondiscrimination. In order to evaluate for and reinforce these criteria, experts have called for increased interpretability of AI/ML. Doshi-Velez and Kim define interpretability as “the ability to explain or to present in understandable terms to a human” and discuss how interpretability can allow humans to learn, ensure the model is safe and ethical, and weigh competing objectives23. These factors may be particularly important in the field of medicine, where care needs to be personalized not only to each individual’s health history, but also to their values and preferences. Interpretability may allow clinicians to readily weigh their own clinical judgment, a factor that both clinicians and informaticians identified as important. Efforts are being made to further define what it means for AI/ML systems to be interpretable, and how that can be technically accomplished24.
Both informaticians and clinicians discussed the importance of how AI/ML systems are validated and deployed, touching on subjects such as the quality of electronic health record data, whether the system can be applied to patient populations other than which it was trained , how to avoid perpetuating bias, and how the system will fit into the clinical workflow. Organizations are working to define standards for how AI/ML studies report their data and how these models might be evaluated in clinical trials25,26. Leaders also recognize the numerous factors that need to be addressed before AI/ML is brought to the bedside including safety, ethical, and security considerations27,28,29, and guidelines are being developed to support healthcare organizations looking to deploy these systems30.
When asked about potential concerns around the use of AI/ML in VTE prevention, several clinicians shared that they did not know enough about the field of AI/ML to comment. This self-identified knowledge gap has been voiced in other studies20, and some experts argue that clinicians will need to learn more about the field given AI/ML’s inevitable arrival in healthcare31. Clinicians will be needed to serve as peer reviewers and clinical trial leaders, as well as to make the ultimate decision as to whether a new system has a clinically relevant impact on their patients31. While clinicians cannot be expected to become experts in every aspect of medicine, it is clear that we will need leaders in this area in order to ensure that AI/ML enters healthcare in a safe, clinically effective, and ethical manner.
There are several limitations to our study. We used online platforms such as message boards and social media websites for survey distribution, which allowed for wide dissemination but did not allow for an accurate response rate calculation, which can underestimate or obscure sampling bias. Given participation was voluntary, the respondents who elected to participate may be more comfortable with technology and more likely to view AI/ML positively. The majority of clinician respondents were younger than 40 years old, and while these findings may reflect the views of a new generation of practitioners, they may not reflect the views of older clinicians in leadership roles who have the ability to effect change today. We used broad terms such as “blood clot” and “AI/ML” in our surveys in order to assess general attitudes and account for respondents who may not have a clinical or technical background, respectively. Several respondents commented on the difficulty of assessing these broad fields, and further research will need to be done evaluating specific algorithms in specific clinical contexts. Additionally, further studies evaluating the opinions and preferences of patients will need to be done21.
Conclusion
A majority of clinicians and healthcare informaticians believe AI/ML can help with VTE management. Both groups raised concerns around how the AI/ML system is validated, how the system is ultimately deployed in the clinical setting, and the importance of clinical judgment. Further studies of specific AI/ML systems are needed to understand how they can impact clinical care.
Data availability
Data is available upon request; researchers interested in reviewing our data should contact Rushad Patell at rpatell@bidmc.harvard.edu.
