Full Text

  1. Home
  2. Full Text
Research Article

Leveraging Advanced AI Agents for Unified Electronic Health Records: A Novel Approach to Healthcare Interoperability

Authors: Akhil Chaturvedi and Jasmine Kaur

Publication Date: December 30, 2023

DOI: https://doi.org/10.51219/JAIMLD/akhil-chaturvedi/334

Citation: Citation: Akhil Chaturvedi, et al. Leveraging Advanced AI Agents for Unified Electronic Health Records: A Novel Approach to Healthcare Interoperability. J Artif Intell Mach Learn & Data Sci, 2023, 1(4): 1-8.

Copyright:Copyright: ©2023 Akhil Chaturvedi, et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

View : PDF

Abstract

This paper presents an innovative solution to the persistent challenge of electronic health record (EHR) interoperability using advanced artificial intelligence (AI) agents. We propose a system that utilizes multiple AI agents with retrieval-augmented generation capabilities and multi-modal understanding to integrate disparate healthcare data into a unified, patient-centric database. By incorporating recent advancements in AI agent architectures, as explained by Lilian Weng and others, our approach aims to overcome the fragmentation of patient information across different EHR systems and data formats. We detail the current state of EHR systems, explain the concept of AI agents and their relevance to healthcare interoperability, present our proposed AI agent framework and analyze the potential impact on healthcare outcomes. Our system demonstrates the potential to significantly improve data integration, information retrieval and overall patient care by creating a standardized, vectorized representation of patient data that is both comprehensive and easily accessible.

 

Keywords: Electronic Health Records, AI Agents, Healthcare Interoperability, Data Integration, Patient-Centric Systems, Retrieval-Augmented Generation, Semantic Interoperability, Multi-modal Learning, Healthcare Data Standardization, Health Information Exchange.

 

1. Introduction

The introduction of Electronic Health Record (EHR) systems was heralded as a transformative development in healthcare, promising streamlined medical practices, enhanced communication among providers and improved patient outcomes1. However, the reality has often fallen short of these expectations due to the fragmentation and incompatibility of systems across different healthcare providers and institutions2.

 

Patient information is frequently dispersed across multiple EHR platforms, creating data silos that hinder the ability to obtain a comprehensive view of a patient's medical history3. This fragmentation is exacerbated by the diversity of data formats in healthcare, which includes text-based clinical notes, complex imaging studies like MRIs and CT scans and structured data such as laboratory results4. For instance, a patient's primary care records might be stored in an Epic EHR system, specialist visit notes in a Cerner EHR system, radiological images in a Picture Archiving and Communication System (PACS) and laboratory results in a Laboratory Information System (LIS)5.

 

The lack of a unified patient history can lead healthcare providers to make decisions based on incomplete information, potentially overlooking critical aspects of a patient's medical history and impacting the quality and safety of care provided6. A specialist might prescribe medication without awareness of potential drug interactions with medications prescribed by another provider, simply because that information is not readily available in their EHR system7.

 

Moreover, clinicians often spend considerable time searching for and reconciling patient information from multiple sources8. This administrative burden reduces the time available for direct patient care, potentially affecting the quality of healthcare delivery and patient satisfaction9. The fragmentation also increases the risk of medical errors, such as medication errors, unnecessary duplicate tests or missed diagnoses10. Patients might undergo the same diagnostic test multiple times because different providers are unaware that the test has already been conducted11.

 

Patients themselves bear a burden as well. They often have to repeatedly provide the same information to different providers or manually transfer their records between healthcare systems12. This not only leads to frustration but also increases the risk of important information being lost or misreported13. Furthermore, the fragmentation of health records poses a significant barrier to medical research and population health initiatives. The inability to easily aggregate and analyze data from multiple sources hampers efforts to identify trends and develop evidence-based practices that could benefit broader patient populations14.



Figure1: This figure visualizes the primary barriers to achieving full interoperability in electronic health records (EHR).

 systems. The central node represents the overall challenge of EHR interoperability. From this node, several key barriers branch out, including technical barriers, semantic differences, privacy and security concerns, economic disincentives and lack of standardization. Each of these barriers is further broken down into specific issues. For instance, technical barriers include proprietary data formats and incompatible communication protocols, while economic disincentives highlight vendor lock-in and reduced switching costs. This figure illustrates the complexity and multi-faceted nature of EHR interoperability challenges.


To address these pressing challenges, we propose leveraging advanced AI agents to create an interoperable, patient-centric health information system. Our approach aims to bridge the gaps between disparate EHR systems, unify diverse data formats and provide a comprehensive, accessible view of patient health information. By developing specialized AI agents capable of understanding and processing various data formats, extracting relevant information, standardizing it according to established healthcare protocols and integrating it into a unified, vectorized patient record, we aim to ensure that healthcare providers have access to the right information at the right time15.

 

By incorporating advanced AI technologies, including large language models with retrieval-augmented generation capabilities and multi-modal learning, we envision a system that not only integrates existing health data but also enhances its usability and accessibility for both healthcare providers and patients16. This system has the potential to transform healthcare delivery by providing a more complete picture of patient health, reducing administrative burdens, minimizing the risk of medical errors and facilitating more comprehensive medical research17.

 

2. Background: The State of EHR Interoperability

The concept of electronic health records emerged in the 1960s and 1970s, with early systems developed by academic medical centers and the Veterans Administration18. Widespread adoption, however, did not occur until the 2000s, propelled by government initiatives and the promise of improved healthcare quality and efficiency19. The Health Information Technology for Economic and Clinical Health (HITECH) Act of 2009 was a significant milestone, providing financial incentives for healthcare providers to adopt and demonstrate "meaningful use" of EHRs20. Consequently, the percentage of non-federal acute care hospitals using basic EHR systems increased dramatically21.

 

Despite widespread adoption, seamless information exchange between different EHR systems remains elusive. The Office of the National Coordinator for Health Information Technology (ONC) defines interoperability at three levels: foundational, structural and semantic22. Foundational interoperability allows data exchange without requiring the receiver to have knowledge of the data's origin. Structural interoperability ensures data exchanges have unaltered meaning at the data field level. Semantic interoperability enables systems to exchange information and interpret it meaningfully using defined domain models23.

 

Achieving true semantic interoperability has proven challenging due to several factors. Technical barriers exist because different EHR systems often use proprietary data formats and communication protocols24. Semantic differences, such as varied terminologies and inconsistent data representations, can lead to misinterpretation25. Privacy and security concerns, including compliance with regulations like HIPAA, can conflict with data-sharing efforts26. Economic disincentives may also play a role, as some EHR vendors have incentives to maintain closed systems27. The lack of widespread adoption and consistent implementation of data exchange standards like HL7's FHIR further complicates interoperability efforts28.

 

The lack of interoperability has significant implications for healthcare delivery and patient outcomes. Fragmented care arises when providers lack access to a patient's complete medical history, potentially leading to medical errors or unnecessary procedures29. Healthcare providers spend valuable time reconciling patient information, leading to inefficiencies30. Patients may need to repeatedly provide the same information or manually transfer records, causing frustration and increasing the risk of errors31. Research and population health initiatives are impeded due to difficulties in aggregating and analyzing data32, contributing to increased healthcare costs33.

 

3. Advancements in AI Relevant to EHR Interoperability

3.1. Understanding AI Agents

Artificial Intelligence (AI) agents are autonomous systems that perceive their environment through sensors and act upon it using actuators to achieve specific goals34. They perform complex tasks by learning from data and making decisions. According to Lilian Weng35, AI agents can be built upon large language models (LLMs) and augmented with tools, reasoning capabilities and memory to enhance performance.

 Key components of AI agents include:

Ø     Perception: Processing and interpreting data from various sources.

Ø     Reasoning and Planning: Using logic and learned knowledge to make decisions.

Ø     Action: Executing tasks based on decisions.

Ø     Learning: Improving performance over time through data and experience.

Ø     Memory: Storing and retrieving information to support decision-making.

 

Figure2: This figure demonstrates the basic architecture of an AI agent.

 

The central "AI Agent" block interacts with its environment by processing percepts and generating actions. Inside the AI agent, various internal components are responsible for performing different tasks. These include perception (processing data), reasoning and planning (making decisions based on learned knowledge), learning (improving over time through experience) and memory (storing and retrieving information). This structure showcases how AI agents, particularly in healthcare settings, can process multiple types of input data, make informed decisions and take appropriate actions.


3.2. Large Language Models and Retrieval Augmented Generation

Large language models like GPT-4 have demonstrated exceptional capabilities in understanding and generating human-like text36. Retrieval-augmented generation (RAG) combines the generative abilities of LLMs with external knowledge sources, allowing models to access and incorporate up-to-date information37. This approach enhances the model's ability to generate accurate and contextually relevant outputs. In the context of EHR interoperability, RAG enhances information extraction, contextual understanding and knowledge-grounded generation of standardized medical data38-40.

 

3.3. Multi-modal Learning

Healthcare data includes various formats such as text, images and structured data. Multi-modal learning techniques enable AI systems to process and integrate information from these diverse data types41. Recent advancements include vision-language models like CLIP, which understand relationships between text and images42 and multi-modal transformers capable of processing multiple data modalities simultaneously43. These technologies facilitate the integration of clinical notes, medical imaging, laboratory results and other data types.

 

3.4. Few Shot Learning and Prompt Engineering

Few-shot learning enables models to perform new tasks with minimal examples44. Prompt engineering involves designing prompts to guide AI model behavior, allowing adaptation to specific tasks without extensive fine-tuning45. These techniques are crucial for developing AI agents that can quickly adapt to the specific data formats and requirements of different EHR systems.

 

4. Proposed AI Agent Framework for HER Interoperability

4.1. System Architecture

Building on these advancements, we propose an AI agent-based system designed to address EHR interoperability challenges. The system architecture includes the following components:

Ø     Data Ingestion Layer: Securely accesses and ingests data from various EHR systems.

Ø     AI Agent Network: The core of the system, consisting of specialized AI agents.

Ø     Knowledge Base: A comprehensive repository of medical knowledge.

Ø     Unified Patient Record Database: Stores integrated and standardized patient information in a vectorized format.

Ø     API Layer: Provides secure access to integrated patient records for healthcare providers and patient portals.

 

Figure3: This figure outlines the high-level architecture of the proposed AI agent-based system for EHR integration.

 

At the foundation, EHR systems feed into the data ingestion layer, where data from multiple systems is securely accessed. The AI agent network is the system's core, where specialized AI agents work together to process and integrate data. The knowledge base stores medical knowledge, which the agents access as needed. Integrated patient records are stored in the unified patient record database and these records are made accessible through the API layer, ensuring both healthcare providers and patient portals have secure, standardized access to comprehensive patient data.

 

4.2. Specialized AI Agents

The AI agent network consists of several specialized agents:

Ø     Data Ingestion Agent: Interfaces with various EHR systems, handling different data formats and ensuring secure data transfer.

Ø     Multi-Modal Processing Agent: Processes different data types using specialized models for each modality.

Ø     Information Extraction Agent: Extracts relevant medical information from processed data using advanced NLP and computer vision techniques.

Ø     Standardization Agent: Converts extracted information into standardized formats like SNOMED CT or LOINC46, addressing semantic interoperability challenges.

Ø     Integration Agent: Combines standardized information from various sources into a comprehensive patient record, resolving conflicts and ensuring consistency.

Ø     Quality Assurance Agent: Monitors the integration process, validating data integrity and flagging potential inconsistencies or errors for review.

4.3. Multimodal Processing

To handle diverse data types, the system employs multi-modal transformers capable of processing different modalities simultaneously. This integration includes:

Ø     Clinical Notes and Reports (text data)

Ø     Medical Imaging Studies (image data)

Ø     Laboratory Results and Vital Signs (structured data)

Ø     Medication Lists and Prescriptions (semi-structured data)

Table 1: Multi-Modal Data Processing Approaches

Data Type

Processing Approach

Key Technologies

Text

Natural Language Processing

Large Language Models, Retrieval-Augmented Generation (RAG)

Images

Computer Vision

Vision Transformers, Multi-Modal Models

Structured Data

Data Analysis

Tabular Data Models, Graph Neural Networks

Time Series Data

Sequence Modeling

Recurrent Neural Networks, Temporal Convolutional Networks

 

4.4. Standardization and Integration

The Standardization Agent plays a crucial role by mapping diverse terminologies and data representations to standardized formats46. Utilizing NLP capabilities and the medical knowledge base, it ensures that information extracted from various sources adheres to established healthcare standards. The Integration Agent then combines this standardized information, resolving any discrepancies and creating a consistent patient record stored in a vectorized format for efficient retrieval and querying.

 

4.5. Privacy and Security

Privacy and security are fundamental considerations in the system design.

The system implements:

Ø     End-to-End Encryption: Ensures data is securely transmitted and stored.

Ø     Role-Based Access Controls: Limits data access to authorized personnel based on their role.

Ø     Audit Trails: Monitors access and changes to patient records, providing transparency and accountability.

Ø     De-Identification Techniques: Protects patient identity in research and analytics use cases by removing or obfuscating personally identifiable information.

 

5. Discussion
5.1. Challenges and Limitations

Implementing the proposed AI agent-based framework presents several challenges that must be addressed to ensure its success.
Data Privacy and Security is paramount. Ensuring robust protection of sensitive health information while allowing necessary data sharing requires ongoing vigilance against evolving security threats and strict adherence to privacy regulations52. The system must employ advanced encryption methods and access controls to prevent unauthorized access or data breaches.

Ethical Considerations arise regarding data ownership, algorithmic bias and the potential for automated decision-making to inadvertently harm patients. Careful consideration and ongoing ethical oversight are necessary to address these issues, ensuring that AI agents operate transparently and fairly53. Establishing ethical guidelines and involving multidisciplinary teams can help mitigate these concerns.

 

Integration with Existing Systems poses technical and organizational challenges. The proposed system needs to work alongside existing EHR systems and workflows, which may require significant effort in terms of technical integration, staff training and change management54. Collaborating with EHR vendors and healthcare organizations to develop interoperable interfaces is essential.

 

Data Quality and Completeness significantly affect the effectiveness of AI agents. Variations in data quality across diverse EHR systems can impede accurate information extraction and integration [55]. Implementing data validation procedures and working towards improving data entry practices can help address these challenges.

 

Regulatory Compliance is complex, given the need to navigate healthcare regulations across different jurisdictions, such as HIPAA in the United States and GDPR in Europe56. Compliance may limit certain aspects of data sharing and integration, necessitating careful legal review and possibly influencing system design choices to accommodate regional regulations.

 

Trust and Adoption are critical for the system's success. Building trust among healthcare providers, patients and other stakeholders requires transparency in the AI's decision-making processes and robust validation of the system's performance57. Demonstrating the system's benefits through pilot programs and clinical studies can encourage wider adoption.

 

6. Future Directions

Several key areas for future research and development emerge from this work. Advancing Natural Language Understanding is crucial, particularly developing domain-specific models that can better comprehend and interpret complex medical terminology and context58. Such models would enhance the AI agents' ability to accurately process and standardize unstructured text data.

 

Improving Multi-Modal Integration techniques is essential for handling complex or uncommon medical data formats. Research into more sophisticated methods for integrating diverse data types will facilitate the inclusion of a broader range of healthcare data59. This could include integrating genomics data, wearable device data and other emerging data sources.

 

Exploring Federated Learning Approaches offers a promising avenue for allowing AI models to learn from distributed datasets without centralizing sensitive patient data60. This approach can enhance privacy and comply with regulations while still benefiting from large-scale data for training AI agents.

 

Advancements in Explainable AI in Healthcare are necessary to make AI decision-making processes more transparent and interpretable61. Developing methods that allow clinicians and patients to understand how AI agents reach their conclusions will foster trust and facilitate acceptance of AI-driven systems.

 

Implementing Dynamic Knowledge Integration methods will enable the system's knowledge base to be continuously updated with the latest medical research and best practices62. This ensures that AI agents make decisions based on the most current information, improving patient care quality.

 

Finally, developing Patient-Centered Interoperability solutions, such as patient-facing interfaces that empower individuals to understand and manage their comprehensive health data, will promote patient engagement and autonomy63. Incorporating patient input and preferences into the system design can enhance its usability and effectiveness.

 

7. Conclusion

The persistent challenge of EHR interoperability continues to impact healthcare delivery, patient outcomes and medical research. Our proposed AI agent-based framework represents a novel approach that leverages advancements in AI technologies to create truly interoperable electronic health records. By employing specialized AI agents capable of understanding, processing and integrating diverse healthcare data, this system has the potential to transform healthcare delivery.

 

The benefits of this approach include comprehensive patient records, improved efficiency, enhanced decision support, facilitated research and increased patient empowerment. Realizing these benefits requires addressing challenges in data privacy, ethical AI use, system integration and regulatory compliance. A collaborative approach involving healthcare providers, AI researchers, policymakers and patients is essential for the ongoing development and refinement of AI-driven healthcare interoperability solutions.

 

By working together, we can move toward a future where comprehensive, accessible and actionable health information is available at the point of care, ultimately leading to better health outcomes for all.

 

8. References

  1. Blumenthal D, Tavenner M. The "meaningful use" regulation for electronic health records. New England Journal of Medicine 2010;363(6):501-504.
  2. Adler-Milstein J, Pfeifer E. Information blocking: is it occurring and what policy strategies can address it? Milbank Quarterly 2017;95(1):117-135.
  3. Kruse CS, Kristof C, Jones B, Mitchell E, Martinez A. Barriers to electronic health record adoption: a systematic literature review. Journal of Medical Systems 2016;40(12):252.
  4. Ratwani RM, et al. A usability and safety analysis of electronic health records: a multi-center study. Journal of the American Medical Informatics Association 2018;25(9):1197-1201.
  5. Fernández-Alemán JL, et al. Security and privacy in electronic health records: a systematic literature review. Journal of Biomedical Informatics 2013;46(3):541-562.
  6. Institute of Medicine. Health IT and Patient Safety: Building Safer Systems for Better Care. The National Academies Press 2012.
  7. Bates DW, et al. Ten commandments for effective clinical decision support: making the practice of evidence-based medicine a reality. Journal of the American Medical Informatics Association 2003;10(6):523-530.
  8. Sinsky C, et al. Allocation of physician time in ambulatory practice: a time and motion study in 4 specialties. Annals of Internal Medicine 2016;165(11):753-760.
  9. Overhage JM, et al. A randomized trial of "corollary orders" to prevent errors of omission. Journal of the American Medical Informatics Association 1997;4(5):364-375.
  10. Kohn LT, Corrigan JM, Donaldson MS(Eds.). To Err is Human: Building a Safer Health System. National Academies Press 2000.
  11. Bates DW, et al. Effect of computerized physician order entry and a team intervention on prevention of serious medication errors. JAMA 1998;280(15):1311-1316.
  12. Walker J, et al. The value of health care information exchange and interoperability. Health Affairs 2005;24(1):5-10.
  13. Tang PC, et al. Personal health records: definitions, benefits and strategies for overcoming barriers to adoption. Journal of the American Medical Informatics Association 2006;13(2):121-126.
  14. Jensen PB, Jensen LJ, Brunak S. Mining electronic health records: towards better research applications and clinical care. Nature Reviews Genetics 2012;13(6):395-405.
  15. Rajkomar A, et al. Scalable and accurate deep learning with electronic health records. NPJ Digital Medicine, 2018;1(1):18.
  16. https://lilianweng.github.io/posts/2023-06-23-agent/
  17. Miotto R, et al. Deep patient: an unsupervised representation to predict the future of patients from the electronic health records. Scientific Reports 2016;6:26094.
  18. McDonald CJ, et al. The Regenstrief medical record system: a quarter century experience. International Journal of Medical Informatics 1998;54(3):225-253.
  19. Jha AK, et al. Use of electronic health records in U.S. hospitals. New England Journal of Medicine 2009;360(16):1628-1638.
  20. Blumenthal D. Launching HITECH. New England Journal of Medicine 2010;362(5):382-385.
  21. Office of the National Coordinator for Health Information Technology. Report to Congress on Health IT Progress 2016.
  22. Office of the National Coordinator for Health Information Technology. Connecting Health and Care for the Nation: A Shared Nationwide Interoperability Roadmap 2015.
  23. Lehne M, et al. Why digital medicine depends on interoperability. NPJ Digital Medicine 2019;2(1):79.
  24. Dixon BE, et al. A framework for evaluating the costs, effort and value of nationwide health information exchange. Journal of the American Medical Informatics Association 2010;17(3):295-301.
  25. Moreno-Conde A, et al. Clinical information modeling processes for semantic interoperability of electronic health records: systematic review and inductive analysis. Journal of the American Medical Informatics Association 2015;22(4):925-934.
  26. Goldstein MM, Rein AL. Consumer Consent Options for Electronic Health Information Exchange: Policy Considerations and Analysis. Office of the National Coordinator for Health Information Technology 2010.
  27. Adler-Milstein J, Pfeifer, E. Information blocking: is it occurring and what policy strategies can address it? Milbank Quarterly 2017;95(1):117-135.
  28. Bender D, Sartipi K. HL7 FHIR: An agile and restful approach to healthcare information exchange. In 2013 26th IEEE International Symposium on Computer-Based Medical Systems 2013:326-331.
  29. Bodenheimer T. Coordinating care-a perilous journey through the health care system. New England Journal of Medicine 2008;358(10):1064-1071.
  30. Hill RG, et al. 4000 clicks: a productivity analysis of electronic medical records in a community hospital ED. The American Journal of Emergency Medicine 2013;31(11):1591-1594.
  31. Patel VN, et al. Individuals’ access and use of their online medical record nationwide 2015.
  32. Weber GM, et al. Finding the missing link for big biomedical data. JAMA 2014;311(24):2479-2480.
  33. Walker J, et al. The value of health care information exchange and interoperability. Health Affairs 2005;24(1):5-10.
  34. Russell S, Norvig, P. Artificial Intelligence: A Modern Approach (3rd ed.). Pearson 2016.
  35. https://lilianweng.github.io/posts/2023-06-23-agent/
  36. https://openai.com/research/gpt-4
  37. Lewis P, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks. Advances in Neural Information Processing Systems 2020;33:9459-9474.
  38. Zhang Z, et al. Semantics-aware BERT for language understanding. In Proceedings of the AAAI Conference on Artificial Intelligence 2020;34(05):9628-9635.
  39. Alsentzer E, et al. Publicly available clinical BERT embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop 2019;1:72-78.
  40. Guu K, et al. Retrieval augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning 2020:3929-3938.
  41. Baltrušaitis T, Ahuja C, Morency LP. Multimodal machine learning: A survey and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence 2019;41(2):23-443.
  42. Radford A, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning 2021:8748-8763.
  43. Lu J, et al. ViLBERT: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in Neural Information Processing Systems 2019.
  44. Wang Y, et al. Generalizing from a few examples: A survey on few-shot learning. ACM Computing Surveys 2020;53(3):1-34.
  45. Liu P, et al. Pre-train, prompt and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys 2023;55(9):1-35.
  46. https://www.snomed.org/
  47. Nguyen A, et al. Artificial intelligence in the intensive care unit. Critical Care 2020;24(1):1-13.
  48. Syrowatka A, et al. Leveraging artificial intelligence to improve quality and safety in nursing care: A systematic review. International Journal of Medical Informatics 2021;152:104394.
  49. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nature Medicine 2019;25(1):44-56.
  50. Rajkomar A, et al. Scalable and accurate deep learning with electronic health records. NPJ Digital Medicine 2018;1(1):18.
  51. Mandl KD, Kohane IS. Time for a patient-driven health information economy? New England Journal of Medicine 2016;374(3):205-208.
  52. Price WN, Cohen IG. Privacy in the age of medical big data. Nature Medicine 2019;25(1):37-43.
  53. Char DS, Shah NH, Magnus D. Implementing machine learning in health care-addressing ethical challenges. New England Journal of Medicine 2018;378(11):981-983.
  54. Sittig DF, Singh H. A new sociotechnical model for studying health information technology in complex adaptive healthcare systems. Quality and Safety in Health Care 2010;19(3):i68-i74.
  55. Weiskopf NG, Weng C. Methods and dimensions of electronic health record data quality assessment: enabling reuse for clinical research. Journal of the American Medical Informatics Association 2013;20(1):144-151.
  56. Cohen IG, et al. The legal and ethical concerns that arise from using complex predictive analytics in health care. Health Affairs 2014;33(7):1139-1147.
  57. Jacobs M, et al. How machine-learning recommendations influence clinician treatment selections: the example of antidepressant selection. Translational Psychiatry 2021;11(1):108.
  58. Alsentzer E, et al. Publicly available clinical BERT embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop 2019:72-78.
  59. Huang SC, et al. Fusion of medical imaging and electronic health records using deep learning: a systematic review and implementation guidelines. NPJ Digital Medicine 2020;3(1):136.
  60. Rieke N, et al. The future of digital health with federated learning. NPJ Digital Medicine 2020;3(1):119.
  61. Holzinger A, et al. Causability and explainability of artificial intelligence in medicine. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 2019;9(4):e1312.
  62. Beam AL, Kohane IS. Big data and machine learning in health care. JAMA 2018;319(13):1317-1318.
  63. Mandl KD, Kohane IS. Time for a patient-driven health information economy? New England Journal of Medicine 2016;374(3):205-208.