A scoping review of human-AI collaboration patterns and task divisions in healthcare applications Authors: Yuxuan You, Xi Li Published: 2026 ABSTRACT AbstractWith the widespread use of Artificial Intelligence (AI) in healthcare, how human physicians work with AI, a new colleague, has become an issue of increasing concern. This review systematically explores the collaboration between humans and AI in the medical field, focusing on the prevalent human-AI collaboration patterns, the task division mechanisms, and the evaluation metrics. Following the PRISMA guidelines for scoping reviews, we screened the relevant literature and finally included 85 journal papers. Our analysis reveals that human-AI collaboration patterns in healthcare tend to be associated with the inherent risk and automation level of medical tasks, and the collaboration process demonstrates a clear division of labor between human physicians and AI. Furthermore, the evaluation of the human-AI collaboration effect is shifting from single technical metrics (e.g., diagnostic accuracy) to multidimensional aspects incorporating clinical efficiency and user experience. Overall, this review presents collaboration patterns in various medical tasks and serves as a theoretical basis for optimizing the clinical workflows for human-AI collaboration.Subject terms: Computational biology and bioinformatics, Health care, Mathematics and computing, Medical research Introduction In recent years, the application of Artificial Intelligence (AI) in the medical field has made significant progress, from medical image diagnosis to clinical decision support. AI systems have demonstrated a strong ability to process massive amounts of data. However, when medical tasks involve complex clinical situations, relying solely on AI may not ensure accurate diagnosis. By combining the expert experiences of human physicians with the analytical capabilities of AI, human-AI collaboration can further improve the safety and quality of healthcare services through interactions and supervision1. Therefore, human-AI collaboration has gradually become a research focus in healthcare. Currently, human-AI collaboration has been applied to a variety of healthcare tasks. For example, in the task of pathology analysis, AI can quickly identify the suspected area, which is then reviewed and diagnosed by physicians2. In the task of disease prediction, physicians can screen for clinical features and then leave it to AI to build intelligent prediction models3. In the task of cancer detection, an AI-generated detection model can be combined with the expertise of human experts to build a joint detection model4. In addition, human-AI collaboration plays significant roles in tasks such as medical report documentation and image annotation5,6. Although the above studies have widely explored the application of human-AI collaboration in healthcare, the research results are fragmented and lack systematic analysis. The motivation of this review is derived from three aspects: firstly, most studies focus on applying human-AI collaboration to a single task (e.g., medical diagnosis) to improve accuracy7,8, while neglecting the adaptations of collaboration patterns across different medical tasks. Secondly, the existing studies lack a description of the specific task division mechanism between AI and human physicians. It mostly remains in the description that AI provides suggestions and humans make the final decision, which does not reveal the interaction process and collaboration flow between humans and AI. Lastly, most studies focus on technical metrics such as diagnostic accuracy in the evaluation of human-AI collaboration9, without considering physicians’ workloads and patients’ acceptance of AI. The perspective from socio-technical systems emphasizes that the integration of AI into clinical workflows must consider not only technical performance but also the social, organizational, and cognitive dimensions of practice, and the importance of human-centered design in fostering trust and ensuring effective collaboration cannot be neglected10,11. To bridge the above gap, this review aims to answer three research questions (RQ for short): RQ1: What types of human-AI collaboration patterns are adopted for various medical tasks? RQ2: How are the medical tasks divided into work for humans and work for AI in the collaboration? RQ3: How to evaluate the effect of human-AI collaboration and what are its influencing factors? The above questions comply with the aims of a scoping review, “to search, select and synthesize the knowledge addressing an exploratory question”12. Therefore, we adopt the scoping review methodology to identify and process the available literature. We screen the relevant literature from January 2010 to July 2025 in the major databases of Web of Science, Scopus, and PubMed, and finally include 85 journal papers that meet the criteria. Combining with the existing theories of human-AI collaboration13,14, this review systematically summarizes seven common patterns of human-AI collaboration in the medical field (as shown in Fig. 1): AI-prior aid: AI’s suggestions are presented to human physicians, and then human physicians decide whether to adopt or not. AI-prior aid: AI’s suggestions are presented to human physicians, and then human physicians decide whether to adopt or not. AI-later aid: AI’s suggestions are shown to the human physician after the human physician has made the initial decision independently, and the human physician decides whether to adopt the AI’s suggestions. AI-later aid: AI’s suggestions are shown to the human physician after the human physician has made the initial decision independently, and the human physician decides whether to adopt the AI’s suggestions. AI-guided aid: AI needs the human physician to input the necessary information, and then outputs the suggestions. After that, the human physician decides whether to adopt the AI’s suggestions. AI-guided aid: AI needs the human physician to input the necessary information, and then outputs the suggestions. After that, the human physician decides whether to adopt the AI’s suggestions. Human aid: AI needs the human physician to input the necessary information, and then AI gives the decision result. Human aid: AI needs the human physician to input the necessary information, and then AI gives the decision result. Iterative interaction: Human physicians are involved in the iterative training process of AI models, and then AI provides the decision result. Iterative interaction: Human physicians are involved in the iterative training process of AI models, and then AI provides the decision result. Crowd integration by AI: AI is used to synthesize or reconcile the judgments provided by multiple human physicians, and then gives the unified decision result. Crowd integration by AI: AI is used to synthesize or reconcile the judgments provided by multiple human physicians, and then gives the unified decision result. Integration of human and AI: AI and human physicians give their respective decision results, and then the decision results are integrated by voting or other integration methods. Integration of human and AI: AI and human physicians give their respective decision results, and then the decision results are integrated by voting or other integration methods. The collaboration patterns of humans and AI in the medical field. a AI-prior aid. b AI-later aid. c AI-guided aid. d Human aid. e Iterative interaction. f Crowd integration by AI. g Integration of human and AI. Based on the classification of human-AI collaboration patterns in the medical field, this review further analyzes the questions of what types of human-AI collaboration patterns are adopted, how to divide work in the collaboration, and how to evaluate and improve the collaboration effect. Compared with the published literature review, the differences of this review are reflected in two aspects: (1) Different research domains. While existing reviews discuss human-AI collaboration in art or across various domains13,15, this review focuses on how human-AI partnership can be enhanced in the healthcare domain. Compared with other fields such as business and art, human-AI collaboration in healthcare is not only about the effectiveness of the technology, but also directly affects the safety and reliability of clinical practice. (2) Different research angles. While prior reviews have synthesized human-AI collaboration in healthcare from angles such as the antecedents and outcomes of human-AI collaboration16–18, few reviews have systematically synthesized collaboration patterns and task division mechanisms at the micro-level of workflow. For example, compared with the work18 that discusses the impact of AI on physicians’ skills in human-AI interaction, we analyze human-AI collaboration workflows from a micro level to ensure that AI can be better integrated into the existing healthcare system, and our methodological innovation lies in the inductive development of a typology based on collaboration patterns observed across a large and systematically collected studies. Overall, this review has the following contributions: First, this review reveals the associations between the types of medical tasks and the types of human-AI collaboration patterns in the existing literature. For example, it is reflected that the iterative interaction collaboration pattern tends to be associated with annotation tasks, while the human aid collaboration pattern appears to be associated with prediction tasks. These observations provide a new research perspective on understanding the current application status of medical human-AI collaboration. Second, this review provides a thorough analysis of the task division mechanism in medical human-AI collaboration. In most cases, AI dominates data-intensive operations, and human physicians provide expert experience and supervision. The two present a collaborative working relationship through AI preprocessing, human correction, and joint optimization. This finding provides a theoretical basis for optimizing the design of clinical workflows for human-AI collaboration. Third, this review observes a new trend toward multidimensional assessment for human-AI collaboration in the reviewed literature. Existing studies are shifting from technical benefit metrics (e.g., accuracy) toward incorporating social factors such as cognitive load, which echoes the perspective from socio-technical systems emphasizing the importance of social, organizational, and cognitive factors. In addition, through comprehensive analysis, it is found that collaboration patterns, AI’s interpretability and uncertainty, and the quality of AI’s suggestions may influence the effects of human-AI collaboration, laying the foundation for optimizing the human-AI collaboration strategy. The remainder of the paper is organized as follows: the “Results” section provides the main research results of the three research questions; the “Discussion” section discusses the research trends, future research opportunities, conclusions and limitations; the “Methods” section outlines the methodology and the literature screening process. Results Overview After literature search and screening (see Fig. 2 and Table 1), 85 papers were included in this study. The 85 included papers were published between 2016 and 2025, and more than half of the papers were published in the last three years (2022–2025). As shown in Fig. 3, the number of papers published per year is on the rise. Although the publication statistics for 2025 are only up to July 2025, they are already close to the publication statistics for the whole year of 2024. These data indicate that human-AI collaboration in healthcare has become an issue of increasing concern. Flow chart of literature search and screening. This figure describes the process of identification and screening. Table 1Inclusion and exclusion criteriaCategoryInclusion criteriaExclusion criteriaReasonArticle typeJournal article, published after peer reviewConference paper, letter, thesis and dissertation, commentaries, perspectives, book chapter, preprint paper (not peer-reviewed)Journal articles typically contain more comprehensive research content and resultsResearch typeQuantitative analysisNon-empirical research, qualitative analysis, literature reviewFor the research questions in this review, articles that employ quantitative analysis tend to derive more specific conclusionsMethodRelated to human-AI collaborationNot related to human-AI collaborationThe theme of this review is human-AI collaboration, i.e., the collaboration between humans and AI to accomplish a particular taskApplication scenariosApplications in the fields of healthcareNot in the fields of healthcareThis review focuses on the application of human-AI collaboration in the healthcare fieldResultsCompared with the benchmarksWithout comparisons with the benchmarksThe paper compared with the benchmark method (AI-only, human physicians-only) shows the effect of human-AI collaborationLanguageEnglishNot EnglishDue to language limitations, this review only includes papers published in English Inclusion and exclusion criteria The number of articles per year. This figure shows the distribution of the publication year for the included papers. We further utilized VOSviewer software to perform text analysis on the titles and abstracts of these papers. The high-frequency keywords and the keyword co-occurrence network are shown in Fig. 4. The color of the node was used to distinguish the average publication time of the paper where the keyword is located. The darker the color, the more attention the keyword has received in recent years. Based on Fig. 4, the keywords in the included papers can be divided into four categories: basis (e.g., study/dataset), technology (e.g., artificial intelligence/neural network), collaboration (e.g., human AI collaboration/interaction), and evaluation (e.g., diagnostic accuracy/sensitivity). The co-occurrence network and the trend of the keywords in titles and abstracts of the included papers. The darker the color, the more recent the keywords appeared. The temporal distribution of these keywords presents the following essential information: “Artificial intelligence” and other technical keywords are used throughout, indicating the role of AI in underlying support for medical human-AI collaboration. In these AI technologies, “neural network” and “deep learning” have received considerable attention. Among the co-occurring words that are closely related to these technological terms, it is worth noting that the heat of “interpretability” experienced a sudden increase in 2023, reflecting a trend in technology research changing from purely algorithmic approaches to a focus on algorithm reliability and interpretability. “Artificial intelligence” and other technical keywords are used throughout, indicating the role of AI in underlying support for medical human-AI collaboration. In these AI technologies, “neural network” and “deep learning” have received considerable attention. Among the co-occurring words that are closely related to these technological terms, it is worth noting that the heat of “interpretability” experienced a sudden increase in 2023, reflecting a trend in technology research changing from purely algorithmic approaches to a focus on algorithm reliability and interpretability. Keywords in the category of collaboration grow significantly after 2023, especially “human AI collaboration” and “interaction”, which are closely related to the research topic of this review. This evidences that human-AI collaboration has gradually become a new research focus in recent years. Keywords in the category of collaboration grow significantly after 2023, especially “human AI collaboration” and “interaction”, which are closely related to the research topic of this review. This evidences that human-AI collaboration has gradually become a new research focus in recent years. Among the keywords in the category of evaluation, early studies mainly focused on technical metrics such as “diagnostic accuracy”, while recent studies shifted to social metrics such as “trust” and “bias”. This may suggest that the research perspective is shifting from pure algorithm performance evaluation to the multidimensional assessment of the human-AI collaboration process. Among the keywords in the category of evaluation, early studies mainly focused on technical metrics such as “diagnostic accuracy”, while recent studies shifted to social metrics such as “trust” and “bias”. This may suggest that the research perspective is shifting from pure algorithm performance evaluation to the multidimensional assessment of the human-AI collaboration process. In summary, based on the overall research trend, the current focus of medical human-AI collaboration research has transformed from AI providing technical support to human-AI collaboration. The importance of interpretability, trust, and other factors in the process of medical human-AI collaboration has been highlighted. However, there is still a lack of systematic integration of specific collaboration patterns and task division between humans and AI. In this review, we systematically combine the collaboration patterns, task division, and evaluation of the collaboration effect to bridge the existing research gap and provide a theoretical base for building an efficient medical human-AI collaboration system. RQ1: medical tasks and collaboration patterns Based on the 85 papers included in this study, we categorize medical tasks into 14 types: annotation, anonymization, classification, coding, detection, diagnosis, documentation, evaluation, interaction, prediction, scoring, segmentation, feature selection, and treatment. The categories were inductively generated from the reviewed literature. Based on the task descriptions originally framed by the authors, we grouped similar descriptions into corresponding categories. Note that diagnosis, classification, and detection look similar but have different emphases in the medical field. Diagnosis is a decision-making process that integrates information from multiple sources (symptoms, history, test results). Classification is a task that classifies objects into discrete categories based on predefined criteria, e.g., classifying pathological images. Detection is a task that identifies the presence of a specific target. The categorization of papers into diagnosis, classification, and detection is based on the vocabulary used in the papers. The distribution of the number of medical tasks in the 85 included papers is shown in Fig. 5a. Diagnosis, classification, annotation, detection, treatment, and prediction are the six most frequently occurring tasks, which are also the high-frequency scenarios in medical practice. The number of articles grouped by tasks and collaboration patterns. a The number of article grouped by tasks. b The number of article grouped by collaboration patterns. Figure 5b shows the distribution of the number of human-AI collaboration patterns in the 85 included papers. The classification of collaboration patterns was based on the human-AI interaction flow described in each paper. For example, if AI’s suggestions are presented to human physicians and then human physicians decide whether to adopt or not, it is classified as AI-prior aid; if AI’s suggestions are shown to the human physician after the human physician has made the initial decision independently and the human physician decides whether to adopt the AI’s suggestions, it is classified as AI-later aid. In addition to the seven collaboration patterns we have mentioned, “Multiple collaboration patterns” indicates that the paper considered more than two collaboration patterns, and “Not clarified” suggests that the paper did not specify which collaboration pattern was used. Overall, it is found that iterative interaction, in which humans are involved in the iterative training of AI models, is one of the most dominant collaboration patterns. Top-ranked collaboration patterns also include AI-prior aid, human aid, and AI-later aid. To answer RQ1, a heat map is used to quantify the correspondence between the medical tasks and the collaboration patterns, as shown in Fig. 6. The number in each cell in the figure represents the number of articles that adopt the collaboration pattern under the medical task. The heat map of the distribution of collaboration patterns adopted in various medical tasks. The number in each cell in the figure represents the number of articles that adopt the collaboration pattern under the medical task. By analyzing the distribution of collaboration patterns, we find that: Among the 11 papers with annotation tasks, 10 papers adopted the iterative interaction as the human-AI collaboration pattern6,19–27. Annotation requires multiple rounds of human-AI information exchange. Through iterative interaction between AI and human physicians, the automatic annotation model can be trained through continuous adjustment from the human physicians. Among the 11 papers with annotation tasks, 10 papers adopted the iterative interaction as the human-AI collaboration pattern6,19–27. Annotation requires multiple rounds of human-AI information exchange. Through iterative interaction between AI and human physicians, the automatic annotation model can be trained through continuous adjustment from the human physicians. Among the 17 papers with classification tasks, 5 papers adopted AI-prior aid as the human-AI collaboration pattern8,28–31, and 5 papers adopted iterative interaction as the human-AI collaboration pattern32–36. By AI-prior aid, the AI will present the automatic classification results first, and then the human physician will review the results. By iterative interaction, some samples will be given to the human physician for labeling in the training process of the classification model, or the feedback of the classification results from human physicians will be utilized for iterative training of AI. Among the 17 papers with classification tasks, 5 papers adopted AI-prior aid as the human-AI collaboration pattern8,28–31, and 5 papers adopted iterative interaction as the human-AI collaboration pattern32–36. By AI-prior aid, the AI will present the automatic classification results first, and then the human physician will review the results. By iterative interaction, some samples will be given to the human physician for labeling in the training process of the classification model, or the feedback of the classification results from human physicians will be utilized for iterative training of AI. Among the 6 papers with prediction tasks, 4 papers adopted human-aid as the human-AI collaboration pattern3,37–39. In human-AI collaboration by human aid, the expert experience of human physicians is integrated into the AI prediction algorithm to improve the performance. Among the 6 papers with prediction tasks, 4 papers adopted human-aid as the human-AI collaboration pattern3,37–39. In human-AI collaboration by human aid, the expert experience of human physicians is integrated into the AI prediction algorithm to improve the performance. In addition, Fig. 6 enables us to draw some valuable conclusions using the collaboration pattern as a vantage point. For example, although there are only 7 papers adopting the AI-later aid, 4 of them are related explicitly to diagnostic tasks7,9,40,41. AI-later aid requires human physicians to first rely on their own experience to reach a conclusion before receiving the suggestions from AI, which avoids the risk of preconceived notions from AI. The adoption of AI-later aid may relate to the risk profile of such tasks. To further clarify the relevance of medical tasks to the human-AI collaboration patterns, we classify the seven human-AI collaboration patterns into three main groups based on which entity primarily produces the final output of the collaboration process: (a) AI-driven collaboration pattern, including human aid, iterative interaction, and crowd integration by AI; (b) Human-dominant collaboration pattern, including AI-prior aid, AI-later aid, and AI-guided aid; (c) Others, including integration of human and AI, multiple collaboration patterns, and not clarified. We take iterative interaction as an example to illustrate the classification rule. In most cases of this pattern, several rounds of training are performed using human expert feedback to optimize ideal performance of the AI model, and once the AI model is trained, the AI model autonomously performs the target task on new data23. Therefore, the final task output is primarily generated by the AI, with human involvement mainly contributed to the model training process rather than the operational output phase. Based on the above reasons, iterative interaction is grouped under AI-driven collaboration pattern in this review. The percentages of the three groups of collaboration patterns in various medical tasks are shown in Fig. 7. According to Fig. 7, we have the following findings: Human-dominant collaboration pattern is mainly adopted in the treatment tasks. Since treatment is a high-risk clinical task that needs to be accomplished by human experts, AI can’t completely replace human experts in this task area. Human-dominant collaboration pattern is mainly adopted in the treatment tasks. Since treatment is a high-risk clinical task that needs to be accomplished by human experts, AI can’t completely replace human experts in this task area. AI-driven collaboration pattern is widely adopted in annotation, anonymization, interaction, prediction, segmentation, and feature selection tasks. The common point of these tasks is that they have been developed with a high degree of automation and can be completed by AI in most scenarios. For example, in the annotation tasks, the automatic annotation model can replace humans to complete most of the annotation work. AI-driven collaboration pattern is widely adopted in annotation, anonymization, interaction, prediction, segmentation, and feature selection tasks. The common point of these tasks is that they have been developed with a high degree of automation and can be completed by AI in most scenarios. For example, in the annotation tasks, the automatic annotation model can replace humans to complete most of the annotation work. The ratio of human-dominant collaboration patterns and AI-driven collaboration patterns is close on the tasks of classification, coding, detection, diagnosis, evaluation, and scoring. It is reflected that although AI plays an enhanced role in disease classification, initial screening, diagnosis, and evaluation, the results of AI rely on the supervision of human physicians. The ratio of human-dominant collaboration patterns and AI-driven collaboration patterns is close on the tasks of classification, coding, detection, diagnosis, evaluation, and scoring. It is reflected that although AI plays an enhanced role in disease classification, initial screening, diagnosis, and evaluation, the results of AI rely on the supervision of human physicians. The percentage of three types of collaboration patterns in various medical tasks. AI-driven collaboration pattern includes human aid, iterative interaction, and crowd integration by AI. Human-dominant collaboration pattern includes AI-prior aid, AI-later aid, and AI-guided aid. Others includes integration of human and AI, multiple collaboration patterns, and not clarified. In summary, for RQ1, the core conclusion is that human-AI collaboration patterns tend to be associated with the risk attributes and automation of medical tasks. High-risk, low-automation tasks retain human dominance, while low-risk, high-automation technical operations expand the AI’s roles in improving efficiency. RQ2: division of tasks between humans and AI RQ1 and RQ2 operate at different levels: collaboration patterns shape the overall interaction logic, whereas task division specifies the distribution of work content. To answer RQ2, we summarize the task divisions in human-AI collaboration based on the 85 included papers. The works undertaken by AI and the works undertaken by humans in various medical tasks are shown in Table 2. Note that the sub-tasks represent the specific functional components that constitute the division of labor within the collaboration process. For example, in a study where the overall task is annotation, the AI may be assigned the sub-task of “predict label” for clear cases, while the human physician is assigned the sub-task of “review” for uncertain cases. Thus, the sub-tasks are not distinct task types but rather the work units that together accomplish the medical task. By analyzing Table 2, we find that the task division mechanism in medical human-AI collaboration presents three significant features: Human physicians and AI have a clear division of labor. AI dominates data-intensive operations, while human physicians focus on expert experience and supervision. This division of tasks is essentially an adaptation of technical computability and clinical cognitive complexity. For example, in diagnostic tasks, AI is responsible for identifying lesion regions and providing preliminary diagnostic conclusions, and then human physicians conduct professional diagnoses in conjunction with the lesion regions identified by AI. This collaboration combined the advantages of AI and humans to improve efficiency and reduce the missed diagnoses2. Human physicians and AI have a clear division of labor. AI dominates data-intensive operations, while human physicians focus on expert experience and supervision. This division of tasks is essentially an adaptation of technical computability and clinical cognitive complexity. For example, in diagnostic tasks, AI is responsible for identifying lesion regions and providing preliminary diagnostic conclusions, and then human physicians conduct professional diagnoses in conjunction with the lesion regions identified by AI. This collaboration combined the advantages of AI and humans to improve efficiency and reduce the missed diagnoses2. Human physicians and AI present a collaborative working relationship through AI preprocessing, human correction, and joint optimization. Taking the classification task as an example, AI first performs preliminary classification on all samples and selects a small number of samples with high uncertainty for human physicians to review, and then incorporates the corrected labels by the human physician into the model for iterative training. The above operations are repeated until the classification model achieves the expected accuracy32,33. This division of tasks highlights the workflow of AI preprocessing, human correction, and joint optimization. Human physicians and AI present a collaborative working relationship through AI preprocessing, human correction, and joint optimization. Taking the classification task as an example, AI first performs preliminary classification on all samples and selects a small number of samples with high uncertainty for human physicians to review, and then incorporates the corrected labels by the human physician into the model for iterative training. The above operations are repeated until the classification model achieves the expected accuracy32,33. This division of tasks highlights the workflow of AI preprocessing, human correction, and joint optimization. Human physicians retain control and authority in high-risk tasks. When the task involves high risks to life and health (e.g., treatment tasks), humans strictly retain final control, and the AI only acts as a proposal generator42. Conversely, in low-risk technical operations (e.g., documentation), AI can autonomously complete the first draft of the writing, and then a physician reviews the results of AI5. Human physicians retain control and authority in high-risk tasks. When the task involves high risks to life and health (e.g., treatment tasks), humans strictly retain final control, and the AI only acts as a proposal generator42. Conversely, in low-risk technical operations (e.g., documentation), AI can autonomously complete the first draft of the writing, and then a physician reviews the results of AI5. Table 2The work divisions of humans and AITaskWork of AIWork of HumanReferencesAnnotation• Select samples for human to annotate• Predict label• Annotate• Annotate• Review6,19–27,58Anonymization• Anonymize• Annotate59Classification• Classify• Select samples for human to annotate• Segment• Encode• Learning prototype• Collect data• Review• Classify• Annotate• Edit attention map• Review prototype8,28–36,60–66Coding• Select samples for human to annotate• Code• Highlight• Annotate• Input expertise• Adjust thresholds• Code43,67Detection• Detect• Derive variables• Select samples for human to annotate• Review• Derive variables• Annotate• Input expertise4,44,68–75Diagnosis• Diagnose• Identify regions• Combine the input of human group• Predict disagreement• Categorize clinical significance• Diagnose• Review• Select features• Annotate• Vote2,7,9,40,41,46,47,49,54,76–86Documentation• Write• Review• Regenerate by AI5Evaluation• Evaluate• Select areas for AI to evaluate• Review• Evaluate• Identify the optimal AI model design87–89Interaction• Interact• Input expertise45Prediction• Predict• Cluster• Aggregation rules• Review• Input agreement scores• Select features• Estimate change points in model• Integrate cluster• evaluate rules3,37–39,90,91Scoring• Score• Refine by human annotation• Review• Annotate92,93Segmentation• Segment• Annotate• Score94,95Feature selection• Suggest features• Select features96Treatment• Treat• Suggest treatment• Optimize treatment• Provide options• Treat• Review• Input initial treatment• Choose42,48,50,53,97–100 The work divisions of humans and AI • Select samples for human to annotate • Predict label • Annotate • Annotate • Review • Classify • Select samples for human to annotate • Segment • Encode • Learning prototype • Collect data • Review • Classify • Annotate • Edit attention map • Review prototype • Select samples for human to annotate • Code • Highlight • Annotate • Input expertise • Adjust thresholds • Code • Detect • Derive variables • Select samples for human to annotate • Review • Derive variables • Annotate • Input expertise • Diagnose • Identify regions • Combine the input of human group • Predict disagreement • Categorize clinical significance • Diagnose • Review • Select features • Annotate • Vote • Review • Regenerate by AI • Select areas for AI to evaluate • Review • Evaluate • Identify the optimal AI model design • Predict • Cluster • Aggregation rules • Review • Input agreement scores • Select features • Estimate change points in model • Integrate cluster • evaluate rules • Score • Refine by human annotation • Review • Annotate • Annotate • Score • Treat • Suggest treatment • Optimize treatment • Provide options • Treat • Review • Input initial treatment • Choose To visualize the division of tasks between humans and AI, we use a Sankey diagram (Fig. 8) to show the flow relationship from task to segmentation of subtasks, and then assigning subtasks to humans and AI. In Fig. 8, the thickness of the lines between the nodes represents the magnitude of the flow. It is shown that the task division reflects a joint cognitive system where humans and AI are interdependent, each compensating for the other’s limitations. The main sub-tasks assigned to AI include “Predict label”, “Select samples for human to annotate” (i.e., the uncertain samples are asked human for help), etc. (bold blue lines in Fig. 8). The main sub-tasks assigned to humans include “Review”, “Input expertise”, “Select features”, etc. (bold red line in Fig.8). The experience of experts can help the AI to correct errors, and also help the AI to select appropriate features to improve the model. The above findings once again demonstrate the clear division of labor between humans and AI. The work of AI, the work of humans, and the work of humans and AI. The thickness of the lines between the nodes represents the magnitude of the flow. RQ3: metrics and influencing factors of collaboration performance To address the question of RQ3 that how to evaluate the effect of human-AI collaboration, we collected the metrics used in the papers that compared the experimental results of three working groups (human alone, AI alone, and human-AI collaboration). In most of these comparative experiments, although the main metrics are work effectiveness (diagnosis accuracy, detection accuracy, treatment performance, etc.), there is a new trend toward multidimensional assessment for human-AI collaboration in the reviewed literature. For example, some comparative experiments use the time cost in addition to the effectiveness5,43,44, since time cost is also an important metric for healthcare operation management. Apart from the technical metrics above, non-technical metrics (e.g., the relation and emotion of humans towards AI) are gradually gaining attention in the adoption of human-AI collaboration. Some non-technical factors, such as patients’ acceptance of AI and physicians’ psychological load in AI, have been investigated in existing studies45. To address the question of RQ3, which factors affect human-AI collaboration performance, we screened 10 papers out of 85 included papers that compare human-AI collaboration performance under different conditions. We summarized the context, task, collaboration pattern, comparative factors, metrics, and outcomes of these papers. The results are presented in Table 3. Through summarization, we found that the factors affecting human-AI collaboration performance include the following: Collaboration pattern. Three papers compared the effects of different collaboration patterns on human-AI collaboration performance, but the results were inconsistent. For example, in an acute abdominal pain diagnostic experiment, the results showed that AI-later aid was superior to AI-prior aid46; however, in another diagnostic experiment on knee MRI (Magnetic Resonance Imaging) and ECG (Electrocardiogram) images, the results showed that AI-prior aid was superior to AI-later aid47. The different experimental conclusions may be related to the experimental context. AI has achieved remarkable results in the diagnosis of medical images, while human physicians may be more experienced in exploring the etiology of acute abdominal pain. Besides, some studies have found that adopting AI-later aid has a lower adjustment cost than adopting crowd integration by AI and integrating human and AI48. Collaboration pattern. Three papers compared the effects of different collaboration patterns on human-AI collaboration performance, but the results were inconsistent. For example, in an acute abdominal pain diagnostic experiment, the results showed that AI-later aid was superior to AI-prior aid46; however, in another diagnostic experiment on knee MRI (Magnetic Resonance Imaging) and ECG (Electrocardiogram) images, the results showed that AI-prior aid was superior to AI-later aid47. The different experimental conclusions may be related to the experimental context. AI has achieved remarkable results in the diagnosis of medical images, while human physicians may be more experienced in exploring the etiology of acute abdominal pain. Besides, some studies have found that adopting AI-later aid has a lower adjustment cost than adopting crowd integration by AI and integrating human and AI48. Interpretability and uncertainty of AI. Existing studies have demonstrated that human-AI collaboration in the presence of AI with interpretability leads to better task outcomes49, and different forms of interpretation may affect the effectiveness of human-AI collaboration50. The interpretability of AI plays an important role51, since it is not merely a technical feature but a prerequisite for clinician trust and appropriate reliance. In addition, whether AI presents the uncertainty of the output results can affect the effectiveness of human-AI collaboration. In cases where AI provides human physicians with a level of uncertainty about the outcome of the AI’s decision, human-AI collaboration can achieve superior performance40. Interpretability and uncertainty of AI. Existing studies have demonstrated that human-AI collaboration in the presence of AI with interpretability leads to better task outcomes49, and different forms of interpretation may affect the effectiveness of human-AI collaboration50. The interpretability of AI plays an important role51, since it is not merely a technical feature but a prerequisite for clinician trust and appropriate reliance. In addition, whether AI presents the uncertainty of the output results can affect the effectiveness of human-AI collaboration. In cases where AI provides human physicians with a level of uncertainty about the outcome of the AI’s decision, human-AI collaboration can achieve superior performance40. Incorrect, biased, or unsafe suggestions from AI. Several studies have investigated whether human physicians can effectively filter out incorrect, biased, or unsafe suggestions from AI during human-AI collaboration. The experimental results were found as follows: (a) When AI presents incorrect suggestions in human-AI collaboration, the diagnostic accuracy will be reduced31, and the decision time of physicians is significantly longer in the context of incorrect AI assessments29. Existing work52 also finds that incorrect AI suggestions may adversely impact clinician treatment selections. (b) When AI presents biased suggestions in human-AI collaboration, physicians may not be sufficiently in control to correct the agent’s bias30. (c) When AI presents unsafe suggestions in human-AI collaboration, physicians stop unsafe AI advice more frequently than safe ones, though not all unsafe AI advice was stopped53. In summary, the quality of AI’s suggestions will directly affect the effectiveness of human-AI collaboration, and the algorithmic accountability and ethical responsibility are in need of attention for implementing AI. Incorrect, biased, or unsafe suggestions from AI. Several studies have investigated whether human physicians can effectively filter out incorrect, biased, or unsafe suggestions from AI during human-AI collaboration. The experimental results were found as follows: (a) When AI presents incorrect suggestions in human-AI collaboration, the diagnostic accuracy will be reduced31, and the decision time of physicians is significantly longer in the context of incorrect AI assessments29. Existing work52 also finds that incorrect AI suggestions may adversely impact clinician treatment selections. (b) When AI presents biased suggestions in human-AI collaboration, physicians may not be sufficiently in control to correct the agent’s bias30. (c) When AI presents unsafe suggestions in human-AI collaboration, physicians stop unsafe AI advice more frequently than safe ones, though not all unsafe AI advice was stopped53. In summary, the quality of AI’s suggestions will directly affect the effectiveness of human-AI collaboration, and the algorithmic accountability and ethical responsibility are in need of attention for implementing AI. Table 3The factors affecting human-AI collaboration performance verified by publicationsReferencesContextTaskCollaboration patternComparative factorsMetricsOutcomes (which group wins)46Acute abdominal painDiagnosisMultiple collaboration patternsHuman vs. Human-AI (AI-later aid) vs. Human-AI(AI-prior aid)Diagnosis accuracyHuman-AI (AI-later aid)47Diagnosis by knee MRI, ECGDiagnosisMultiple collaboration patternsHuman vs. AI vs. Human-AI (AI-later aid) vs. Human-AI(AI-prior aid)Diagnosis accuracyHuman-AI (AI-prior aid)48Chronic obstructive pulmonary diseaseTreatmentMultiple collaboration patternsHuman-AI (Crowd integration by AI) vs. Human-AI (Integrate human and AI) vs. Human-AI (AI-later aid)Adjustment costHuman-AI (AI-later aid)49Chest X-ray diagnosisDiagnosisAI-prior aidHuman-AI vs. Human-AI(with explanation)Diagnosis accuracyHuman-AI (with explanation)50Clinical decision supportTreatmentAI-prior aidHuman-AI vs. Human-AI(with different types of explanation)Trust, performanceHuman-AI (with different types of explanation)40Chest X-ray diagnosisDiagnosisAI-later aidHuman-AI vs. Human-AI(with uncertainty)Diagnosis accuracyHuman-AI (with uncertainty)31Cell classificationClassificationAI-prior aidHuman vs. Human-AI(incorrect)Classify accuracyHuman29ColonoscopyClassificationAI-prior aidHuman-AI (correct) vs. Human-AI(incorrect)The decision time of humansThe time is significantly longer in the context of incorrect AI assessments.30Triage choicesClassificationAI-prior aidHuman vs. Human-AI (unbiased) vs. Human-AI(biased)Whether participants neutralized bias from the biased artificial agentHuman participants may not be sufficiently in control to correct the agent’s bias.53ICUTreatmentAI-later aidHuman vs. Human-AI (safe) vs. Human-AI(unsafe)SafetyPhysicians stop unsafe AI advice more frequently than safe ones, though not all unsafe AI advice was stopped. The factors affecting human-AI collaboration performance verified by publications Furthermore, some studies have proposed approaches to enhance the performance of human-AI collaboration. For example, in human-AI collaboration involved in chest radiograph diagnosis, allowing AI to predict the degree of disagreement between AI and human physicians while providing conclusions will effectively reduce the clinician’s burden54. Studies in this area are few but have critical clinical applications. Discussion The number of published papers related to human-AI collaboration in healthcare has continued to grow over the past few years, showing the increasing attention in this field. Despite the above progress in existing research, there are several research gaps to be resolved in future research: First, there is a lack of universal guidance on the applicability of human-AI collaboration patterns. Existing studies have found that the effectiveness of human-AI collaboration patterns varies depending on the task contexts, e.g., AI-prior aid is more effective than AI-later aid in knee MRI, ECG medical image diagnosis47, and AI-later aid is more effective than AI-prior aid in acute abdominal pain diagnosis46. However, the deep mechanisms of such differences have not been thoroughly analyzed. Future research can develop an adaptation model of human-AI collaboration tailored to the characteristics of specific diseases and clarify the applicable scenarios of different collaboration patterns to guide clinical practice. The balance between automation and clinician autonomy is also important in clinical practice, and the field lacks clear guidelines for balancing these trade-offs. Future research could investigate hybrid models that dynamically adjust the level of clinician autonomy based on task complexity, clinician expertise, and real-time performance feedback. Additionally, the factors that influence the reliance of human physicians on AI should be carefully considered and investigated in the implementation of AI55. Second, existing studies are limited in analyzing multi-objective conflict situations in human-AI collaboration. For example, some conflicts may occur between the metrics of time cost and effectiveness. In the tasks of coding and documentation, although the adoption of human-AI collaboration may improve the quality of work compared to the adoption of AI alone, the time spent on the tasks increased. Future research can propose optimization methods for balancing human-AI collaborative work performance and work efficiency. A systematic and comprehensive assessment framework for human-AI collaboration performance has yet to be developed. Future research can integrate technical performance, clinical efficiency, and the experiences of physicians and patients within a unified system, providing a more comprehensive basis for human-AI collaboration in practice. Third, the interpretability and safety of AI’s suggestions in human-AI collaboration need to be deeply studied. Existing research confirms that AI’s interpretability will directly affect the effectiveness of collaboration. However, most studies focus solely on whether to provide explanations, and only a few propose methods for how to present explanations50. Future studies can integrate cognitive psychology into human-AI collaboration to further improve collaboration workflow. In addition, when AI provides incorrect, biased, or unsafe suggestions in human-AI collaboration, human physicians have limited filtering capabilities30,31,53. The risk prevention and control mechanisms in human-AI collaboration are inadequate. Future research can develop a real-time risk warning system with multiple intelligences to filter incorrect, biased, or unsafe suggestions from AI in the human-AI collaboration system. Fourth, the institutional factors that shape the adoption of AI in healthcare settings should be fully considered. Institutional factors such as hospital size, resource availability, regulatory frameworks, liability concerns, and training infrastructure influence the adoption and success of human-AI collaboration. For example, resource-rich academic medical centers may more readily implement human-AI collaboration, while under-resourced primary care settings may struggle with integration. The institutional policies that define roles, responsibilities, and accountability in human-AI collaboration need to be proposed. These institutional dimensions represent an important avenue for future research. Finally, while our review focuses on the collaboration between physicians and AI, patients are often the ultimate beneficiaries or potential recipients of AI-influenced decisions. Few of the included studies considered patient acceptance, trust in AI-assisted diagnoses, or preferences for human oversight. Patient perspectives are critical because low patient trust in AI may undermine adherence to treatment or satisfaction, even if technical performance is high. Future research is encouraged to include patient-reported outcomes and to investigate how to communicate with patients for AI involvement in a proper manner. This review systematically explores the collaboration between humans and AI in the medical field, focusing on three research questions: (1) the human-AI collaboration pattern, (2) the division of tasks between humans and AI, (3) the evaluation metrics and influencing factors of human-AI collaboration. In summary, we have the following findings: First, human-AI collaboration patterns in healthcare tend to be associated with the inherent risk and automation level of medical tasks. High-risk, low-automation tasks are more inclined to adopt human-dominant collaboration patterns, while low-risk, high-automation tasks are more likely to adopt AI-driven modes. This adaptability strikes a balance between safety and efficiency. Second, human-AI collaboration demonstrates a clear division of labor between human physicians and AI. AI focuses on data-intensive operations, while humans focus on injecting expert knowledge into the AI model and reviewing the AI’s results. Lastly, existing studies are shifting from a single-dimensional to a multidimensional assessment. Meanwhile, the collaboration pattern, the interpretability and uncertainty of the AI, and the quality of the suggestions made by the AI have an impact on human-AI collaboration. Overall, this review provides a systematic research perspective on understanding the operation mechanism of medical human-AI collaboration, laying the foundation for constructing a more efficient and credible framework for medical human-AI collaboration. There are several limitations in this review: First, while our use of precise terminology in the search terms enhances the accuracy of identifying papers that focus on human-AI collaboration as a partnership, it may sacrifice comprehensiveness. We encourage future research to adopt broader search terms as the field matures and terminology becomes more standardized. Second, the findings of our review are based on observed patterns in the literature. Due to the methodological constraints, these findings cannot directly move into causal conclusions or normative guidance without empirical validation. Third, to prioritize breadth of coverage, this review includes research papers covering various types of medical tasks, making it difficult to conduct a formal quality appraisal. Studies with varying methodological rigor were included alongside more robust studies. The observed associations we present should be interpreted as descriptive patterns derived from the existing literature rather than prescriptive recommendations supported by a synthesized evidence base of uniform quality. Finally, since the research questions of this review require numerical experimental validation, only papers using quantitative analysis methods were retained. However, qualitative insights are invaluable for understanding the “why” and “how” behind collaboration patterns, including physicians’ perceived trust and the contextual factors that shape successful collaboration. We encourage future research to integrate qualitative synthesis to complement the quantitative findings. Methods Literature search This review was pre-registered on the PROSPERO website (registration number CRD420251118870). The methodology of this review was conducted following the PRISMA guidelines for scoping reviews (PRISMA-ScR)56, which is a systematic approach to synthesize the literature on a topic. The literature search flowchart is shown in Fig. 2. The search was conducted across three large, widely used research databases: Web of Science, Scopus, and PubMed. The pre-search test and the formal search were performed on July 31, 2025 and August 1, 2025. The scope of the search was limited to journal papers published within the past 15 years (January 2010 to July 2025), excluding shorter conference papers, letters, and other types of publications, as papers published in journals typically contain more comprehensive research content. Before the formal search, a pre-search test was conducted. Two researchers proposed three groups of terms for a combined search: human, AI, and healthcare. However, we found that the selected articles based on these words were broad in scope and failed to focus on the research theme, i.e., the cooperation between humans and AI. Therefore, we modified the search strategy by dividing the search terms into two groups, one group being terms related to human-AI collaboration, including “human-AI collaboration”, “human-AI interaction”, “human-AI team”, “human in the loop”, etc., and another group of terms related to healthcare, including “health”, “healthcare”, “medical”, “medicine”, “disease”, “patient”, “hospital”, and so on. Variations in abbreviations and full names (i.e., AI or Artificial Intelligence), as well as noun-to-verb conversions and noun singular-to-plural conversions, were considered in the formal search process to incorporate as much literature as possible. The search strings used for the three databases are presented in Supplementary Table 1. We also provide a PRISMA-ScR Checklist in Supplementary Table 2. Literature screening A total of 2229 papers were retrieved from the three databases. After removing duplicates, 1262 papers proceeded to the screening stage. We established strict inclusion and exclusion criteria, as shown in Table 1. First, the paper needs to be a formally published, peer-reviewed journal article to ensure the paper’s quality. Second, the paper should present an original experiment that utilizes quantitative analysis methods to evaluate the performance of the human-AI collaboration. Based on this requirement, we also excluded purely theoretical work, qualitative analysis, and literature review. Third, the paper needed examples of humans and AI systems collaborating to perform a particular task, so we excluded studies that used only AI without human involvement. Fourth, the paper needs to be applied to the medical field, so we excluded papers with non-medical applications. Fifth, the paper should compare the experimental results of the human-AI collaboration system with those of a benchmark approach (AI-only, human physicians-only) in the same context to show the effect of human-AI collaboration. These criteria helped ensure a baseline level of methodological rigor across the included papers. Additionally, the paper must be written in English, excluding those not written in English. We performed two rounds of screening using the web tool Rayyan57. The first round of initial screening was conducted by filtering through the title, abstract, and publication type of the literature, and the second round of fine screening was performed by reading the full text. To minimize the impact of individual assessment bias, the literature screening process was conducted independently by two researchers. Only papers that both researchers concurrently agreed to exclude were excluded, and controversial papers were referred to a third expert for judgment. After two rounds of screening, 85 papers remained for analysis in this study. Data extraction Based on the 85 papers screened, we conducted a detailed data extraction. The extracted items are dependent on the three research questions. According to RQ1, we extracted the context, medical task, and human-AI collaboration pattern adopted in each paper. According to RQ2, we extracted the specific task divisions in human-AI collaboration of each paper, which includes works undertaken by AI and works undertaken by humans. According to RQ3, we extracted the comparative methods, metrics, and outcomes of each paper. In addition to the above key information, basic details such as the title, abstract, and year of publication will also be extracted and included in the analysis. The main coding process in the data extraction (medical task type coding, collaboration pattern coding) was carried out independently by two researchers, and the results will then be compared to resolve the inconsistencies using the following procedure: first, the two researchers discussed the inconsistent items to try to reach a consensus. If consensus could not be reached after discussion, a third senior expert served as an independent adjudicator, and the expert’s judgment was taken as the final decision. The inter-rater agreement was also assessed. Cohen’s Kappa for the first round of screening, the second round of screening, medical task type coding, and collaboration pattern coding are 0.820, 0.831, 0.863, and 0.826, respectively. Krippendorff’s Alpha for the first round of screening, the second round of screening, medical task type coding, and collaboration pattern coding are 0.821, 0.831, 0.864, and 0.827, respectively. Supplementary information Supplementary Information Acknowledgements This work was funded by Zhejiang Provincial Philosophy and Social Sciences Planning Project (26NDJC013YBM) and a Project Supported by Scientific Research Fund of Zhejiang Provincial Education Department (Y202558061). Author contributions Conceptualization: You Y., collecting and synthesizing data: You Y. and Li X., drafting the manuscript: You Y., revising and finalizing the manuscript: You Y. and Li X. All authors read and approved the final manuscript. Data availability No datasets were generated or analyzed during the current study. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41746-026-02912-y. References Associated Data Supplementary Materials Supplementary Information Data Availability Statement No datasets were generated or analyzed during the current study.