Tim O’Connell of emtelligent on Structuring Healthcare Data for Responsible AI

Tim O’Connell, M.Eng., M.D., co-founder and CEO of emtelligent, offers a rare blend of clinical insight and technical acumen when it comes to the challenges of responsible AI in healthcare. As the industry confronts a surge of generative AI deployments, health-system leaders must reconcile the push for rapid automation with ethical imperatives around data governance, algorithmic fairness, and patient safety.

In the Q&A below, O’Connell outlines the root causes of unreliable AI outcomes: generic model architectures that lack domain fluency, the notorious propensity of generative AI to hallucinate, and the absence of structured, verifiable inputs. He underscores that unstructured clinical text, accounting for roughly 80 percent of an electronic health record (EHR), cannot be treated as a drop-in feed for off-the-shelf AI engines. Without human-in-the-loop validation and rigorous alignment to ontologies such as SNOMED CT and RxNorm, algorithmic outputs risk propagating inaccuracies into chart notes, care plans, and billing workflows.

O’Connell also addresses the ethical dimension: providers and payers carry a duty to protect patient privacy, educate users on AI limitations, and clearly flag AI-generated content. His recommended best practices include mandatory accuracy scoring for end users, regular training whenever models change, and transparent documentation of decision pathways back to source sentences. These measures do more than satisfy auditors; they build clinician and patient trust that AI-driven insights are truly safe to act upon.

Finally, O’Connell previews how domain-trained NLP technologies can revolutionize operational efficiency, such as automating chart abstraction, reducing clinician burnout, preemptively catching billing errors, and powering predictive analytics. Looking ahead, emtelligent’s eight years of model development serve as a blueprint for any organization seeking to elevate raw EHR data into decision-grade intelligence.

What are the primary challenges in adopting AI-driven solutions for structuring unstructured healthcare data, and how can organizations address concerns around data accuracy and system reliability?

Most AI solutions are based on generic, general-use models that are either not equipped to handle the complexities of medical language or fall short in terms of accuracy and use case relevance. As a result, they fail to address the specific needs of healthcare.

To make matters worse, many of these solutions are developed by companies lacking  firsthand knowledge of the unique challenges healthcare organizations face when implementing AI. The result is a mismatch between the promises of AI and the capabilities needed to extract actionable insights and drive meaningful change.

The accuracy of data provided healthcare organizations by AI has been a concern due in large part to the propensity of GenAI to “hallucinate,” or make up answers. If inaccurate data is used by a clinician to guide treatment, patient safety is at risk. The best way to address those concerns is to have human experts “in the loop” to verify the accuracy of AI-generated data. On a broad level, the healthcare industry must adopt and adhere to a set of principles for the ethical use of AI. 

How can healthcare organizations balance the push for AI-driven automation with ethical imperatives such as data governance, algorithmic fairness, and compliance with evolving regulations?

Providers and payers that implement AI have a responsibility to protect patient privacy. As the AMA argues in its proposed guidelines for ethical use of AI in healthcare, “the unique nature of Large Language Models (LLMs) and generative AI warrant special emphasis on entity responsibility and user education.”

Training and education for users should cover the legal, ethical, and equity issues surrounding generative AI, security and patient privacy concerns, the hazards of feeding personal and sensitive information to the algorithm, and why patients should be informed when generative AI is being used to process their data. 

Education also should include presenting accuracy scores to end users so they get a reality check on AI accuracy. Further, training should be conducted whenever there are changes to the AI algorithm. 

In what ways can AI and NLP technologies enhance operational efficiency across healthcare workflows—including administration, compliance, and analytics—while also supporting improved decision-making?

There are a number of ways Medical AI and NLP can improve operational efficiency for clinicians. First, automating the process of extracting and formatting unstructured clinical data – which comprises 80% of all EHR data – can reduce to seconds a process that can take hours for humans to complete. 

Second, Medical AI can identify patterns and provide insights that enhance decision support at the point of care. Third, by automating charting and time-consuming administrative burdens, Medical AI can reduce clinician burnout, a condition that leads to fatigue and lack of focus. 

AI also can quickly identify and fix billing errors prior to claims submissions, reducing claims denials that prevent or delay payments. Similarly, AI automates routine tasks for reviewers, enabling them to be more productive. Further, Medical AI can conduct  automated documentation review to ensure compliance with regulations such as HIPAA and automate regulatory reporting.

Finally, AI tools can be used to power predictive analytics to support clinical decision-making, manage resource allocation, improve operational efficiency, and inform financial decisions. 

What approaches can healthcare organizations take to foster trust in AI-powered tools for managing health information and drive adoption across diverse user groups and sectors?

To gain trust in AI’s ability to manage clinical data, transparency is imperative. Patients and clinicians have a right to know when data and communications have been generated by AI. 

It also is important that clinicians, coders, and reviewers have confidence in the quality of data being generated by AI. Health data that is incomplete, indecipherable, or inaccurate undermines the ability of the LLMs and machine learning algorithms that power medical AI to extract value. Medical AI is most effective for payers and providers when their data is accurate, complete, and standardized. 

Finally, neither clinicians nor patients would be comfortable with AI unilaterally making clinical decisions. Thus, it is essential that healthcare organizations implementing AI ensure there is a “human in the loop” to check the accuracy and relevance of AI outputs, which should include links to the source material for verification. 

Looking ahead, how do you see AI and NLP transforming the landscape of healthcare data management, and what strategic initiatives is your organization pursuing to lead in this space?

Neither AI nor NLP can impact healthcare data utilization if they’re not trained on the right data. Our approach is grounded in deep domain expertise. Over the past eight years, we’ve perfected proprietary language models that are engineered for healthcare. These models are the result of collaboration between a practicing physician with firsthand clinical experience and a computer scientist specializing in NLP and machine learning. 

Our models are not just built for healthcare; they are fluent in it. Leveraging a robust library of medical ontologies and designed for the nuances of clinical language, our Medical AI consistently outperforms tech giants like Google and AWS and our healthcare-specific competitors. 

This performance superiority enables us to address any use case, transforming medical language in any form into actionable insights. From structured data creation to seamless integration into workflows, our technology surpasses the limitations of open-source platforms to deliver the accuracy and precision healthcare demands.