he advent of artificial intelligence (AI) and its application in healthcare has opened new frontiers in medical practice and research. Large Language Models (LLMs) like GPT-4 are at the forefront of this revolution, offering unprecedented capabilities in processing and generating human-like text. When integrated with Digital Health Technologies (DHTs), LLMs can significantly enhance clinical decision-making, patient engagement, and overall healthcare delivery. There are numerous multifaceted applications of LLMs in healthcare, including their integration with DHTs and the implications for clinical and patient care.

LLMs can enhance clinical decision support systems, patient engagement platforms, telehealth services, and complex condition management. But there are also ethical, privacy, and regulatory challenges associated with deploying LLMs in healthcare settings. Through real-world examples and case studies, this article provides a comprehensive overview of the transformative potential of LLMs and DHTs in improving patient outcomes and healthcare delivery.

Clinical Decision Support

According to the 21st Century Cures Act, Clinical Decision Support (CDS) is defined as tools that supply healthcare providers and patients with clinical knowledge and patient-specific information, intelligently filtered or presented at appropriate times, to enhance health and healthcare. CDS systems are pivotal in modern healthcare, providing clinicians with evidence-based recommendations to improve patient outcomes. The integration of LLMs into CDS systems can revolutionize this domain by analyzing vast amounts of clinical data, including notes, research papers, and guidelines. LLMs can provide personalized, evidence-based recommendations, thereby reducing diagnostic errors and enhancing treatment plans. One of the critical advantages is the ability to trace recommendations back to specific guidelines or studies, fostering a sense of trust and control among clinicians. Furthermore, through the use of vector databases using Retrieval Augmented Generation (RAG), a system where contextually relevant documents are fed into a generative model to use this context to generate responses for that text that are informed by the retrieved information, LLMs enable more accurate clinical decision support with interactive chatbots that facilitate user communication.

One such tool that demonstrates this application is SKIN GPT. SKIN GPT is a specialized LLM application designed to assist in the diagnosis and management of dermatologic conditions. By analyzing images of skin lesions and incorporating patient history and other relevant data, SKIN GPT provides detailed diagnostic suggestions and treatment recommendations. This tool exemplifies the use of LLMs in analyzing multimodal data inputs and delivering personalized healthcare information in a user-friendly manner, enhancing both patient and clinician experiences.

Patient Monitoring and Management

The integration of LLMs into DHTs represents a significant advance in patient monitoring and management by facilitating a more proactive and patient-centric approach to healthcare. LLMs, with their robust natural language processing capabilities, enable digital health platforms to analyze vast amounts of multimodal data from reputable knowledge bases. This integration has allowed for real-time, personalized health insights and recommendations. Moreover, the continuous learning capability of LLMs ensures that these systems stay current with the latest medical research and clinical guidelines, thereby enhancing the overall quality of care.

One such tool with these capabilities is AutoHealth for Parkinson’s disease (PD) management, which represents an advanced Internet of Medical Things (IoMT) system leveraging LLM-powered wearable technology to manage PD. The system integrates smartwatches equipped with biosensors and AI chatbots to continuously monitor movement patterns and speech data of PD patients. Utilizing vector-based learning, AutoHealth provides early detection, continuous tracking, and personalized rehabilitation management for PD. The AI chatbot offers interactive communication with patients, responding to text and speech inquiries with tailored guidance, thus enhancing patient engagement and proactive health management.

Electronic Health Records

Electronic Health Records (EHRs) are fundamental to patient care, yet their management often imposes a significant administrative burden on healthcare providers. Ambient listening players who capture and interpret spoken language in real time within clinical settings are using LLMs to streamline this process by converting physician dictations into structured EHR entries, thus reducing the time spent on documentation. Moreover, LLMs like GatorTron have been utilized to analyze EHR data to identify health trends, comorbidities, and treatment responses, bringing the field closer to personalized care. This capability will undoubtedly not only improve efficiency but also enhance the quality of patient care. As we look towards the near future, integrated LLMs will also begin to use multimodal data input, gathering information from various external sources and reputable knowledge bases to analyze sensor data like voice and movement, which are used to communicate interactively with personalized, multimodal, and multilingual chatbots.

Patient Engagement Platforms (PEPs)

Patient engagement is a cornerstone of effective healthcare delivery. LLMs are starting to be integrated into patient portals and mobile health apps to provide personalized health education, answer patient queries, and send reminders for medication adherence. By using natural language processing, these models offer explanations in plain language, making complex medical information more accessible to patients. This greater engagement will empower patients to take a proactive role in managing their health, thereby improving health literacy and outcomes. The continuous and current collection of quality data that LLMs can pull from is expected to improve the “groundedness” of knowledge, integrating better with other tools and data sources to provide more accurate alerts and recommendations.

A great example of this application is Conversational Health Agents (CHAs) like openCHA that demonstrate the potential of LLM-powered frameworks to handle complex healthcare tasks. OpenCHA integrates external data sources, knowledge bases, and analysis models into LLM-based solutions to deliver personalized healthcare responses. By incorporating multimodal data analysis, such as physiological and lifestyle data, CHAs can perform real-time, multistep problem-solving and provide personalized, multilingual, and multimodal interactions with users. This framework exemplifies how LLMs can be used to create more effective, personalized, and trustworthy healthcare solutions for patients.

Transforming Telehealth Services

Telehealth services which support remote healthcare access will also use this technology by connecting disparate pieces of information and refining multiple iterative paths of reasoning (e.g., using chain-of-thought reasoning, agentic AI which dynamically makes decisions and performs actions autonomously based on their programming and learning capabilities, etc.) to arrive at more accurate and informative responses for users. What’s more, LLMs will also enhance telehealth platforms by offering real-time transcription services, language translation, and preliminary assessments based on patient symptoms. These features will make telehealth more accessible and efficient, particularly for patients in underserved areas. By improving the quality and accessibility of telehealth services, LLMs hold promise to contribute to a more equitable healthcare ecosystem.

Addressing Ethical, Privacy, and Regulatory Challenges

The deployment of LLMs in healthcare raises several ethical, privacy, and regulatory challenges. Ensuring fairness and mitigating bias in LLM outputs is critical, as these models can inadvertently propagate biases present in their training data. Transparency and explainability are also essential to maintain trust in AI-driven decisions. In terms of privacy, safeguarding patient data and ensuring robust deidentification methods are paramount. Regulatory compliance with laws like HIPAA and GDPR is crucial, and standardization and certification processes for AI technologies must be developed to ensure safety and efficacy.

Bias and fairness in LLM recommendations are still ongoing concerns. Healthcare providers must ensure that LLMs are interpretable and do not exacerbate health disparities or discrimination. Transparency in how LLMs arrive at conclusions is also vital for maintaining accountability and trust. Additionally, patient autonomy must be respected, ensuring that LLMs support informed consent processes and patient choice.

Data protection and confidentiality are also critical when using LLMs in healthcare. Ensuring the security of patient data and mitigating the risks of reidentification are essential. Compliance with privacy regulations, such as HIPAA and GDPR, is necessary to protect patient information and maintain trust.

Lastly, navigating the complex regulatory environment of healthcare is a significant challenge for deploying LLMs. Ensuring compliance with healthcare regulations and determining liability in cases of adverse outcomes are critical considerations. Developing standards and certification processes for AI technologies in healthcare is necessary to ensure their safety and efficacy.

The Coalition for Health AI (CHAI) is an organization dedicated to promoting the responsible use of AI in healthcare and is currently focused on developing frameworks and standards to address all of these challenges by bringing together a coalition of key stakeholders including patient-community advocates, technology companies, start ups, public sector organizations, medical device manufacturers, payers, and healthcare organizations to come up with consensus definitions and metrics to define what responsible AI in health looks like. More specifically, CHAI is developing technical standards and implementation guides around bias, safety, and testing and evaluation best practices. They are also assembling a nationwide network of certified quality assurance labs and a national registry platform featuring model report cards for algorithms that have been through an assurance lab.

Conclusions

The integration of LLMs into DHTs presents transformative opportunities for clinical and patient care. By enhancing clinical decision support systems, optimizing electronic health records, improving patient engagement, and managing complex conditions like Parkinson’s disease, LLMs can significantly improve healthcare delivery and patient outcomes. However, addressing ethical, privacy, and regulatory challenges is essential to realizing the full potential of these technologies. Collaborative efforts among AI developers, healthcare providers, regulatory bodies, and patients are crucial to navigate these challenges effectively. The future of healthcare, enabled by LLMs in DHTs, promises to be a more efficient, personalized, and equitable healthcare system if this technology is implemented responsibly, continuously involving a diverse set of stakeholders in their development and deployment.