Executive Summary 

• Artificial Intelligence is increasingly integrated into medical practice, combining traditional AI, such as machine learning, and generative AI to improve efficiency and decrease research costs. 

• Abridge AI, Freed AI, GE’s Critical Care Suite, and AlphaFold are FDA approved HIPAA compliant tools which automate clinical documentation, aid diagnostic imaging, and accelerate biomedical research. 

• Despite its benefits, AI in medicine faces significant challenges including data privacy restrictions under HIPAA, potential biases from limited training datasets, transcription inaccuracies especially with accented speech, and unresolved legal liability in cases of misdiagnosis. 

Introduction 

Artificial intelligence (AI) has had a large impact on healthcare by streamlining clinical workflows as well as medical research. Unlike many consumer-facing AI models such as ChatGPT, medical AI solutions are designed to assist with specific purposes such as electronic health record (EHR) documentation, radiological image analysis, and protein structure prediction. A number of successful medical AI tools received FDA approval, HIPAA clearance, and have been deployed to different hospitals and medical research facilities.  

However, the application of AI in medicine presents unique challenges. Training data availability is restricted by privacy laws such as HIPAA, which may introduce biases or inaccuracies. Furthermore, complicated medical language or accented speech may be difficult for transcription technologies to understand, increasing the possibility of mistakes in patient records. Finally, the legal landscape remains uncertain regarding liability when AI tools contribute to diagnostic errors. 

This research analyzes the application of AI in medicine, provides examples of existing AI tools, and examines the risks and concerns connected to deployment of AI in the medical context.   

AI and Medicine 

Artificail Intelligence is a branch of computer science exploring simulation of human intelligence and behavior. AI can be split into two major components: traditional AI and generative AI. Traditional AI focuses on modeling reasoning and logic, focusing on identifying patterns and statistical techniques. These AI models are excellent at classifying information or forecasting future events based on past trends by learning from existing data. Generative AI builds upon those capabilities as it uses user input, called prompts, and trained data to generate output, including text, media, or analysis. 

Medical applications of AI differ from a traditional day-to-day use of ChatGPT or Claude. Rather than focusing on answering basic user queries, medical AI tools are designed to streamline note taking during appointments, generate charts for electronic health record (EHR), support identification of X-ray imaging, or predict biological structures, based on finite data which is harder to obtain due to existing HIPAA laws. This requires a mix of traditional and generative AI capabilities. Traditional AI excels at analyzing historical data to forecast future probabilities and risks, identify possible treatment plans, and detect and transcribe languages using machine learning. On the other hand, Gen AI can simulate different future scenarios, personalize treatment plans, add nuances and context to different languages, and generate any necessary products based on traditional AI output.  

Medical Application Examples 

There are several successful AI tools that have carved out a niche in medical applications. They differ by focus, application, and HIPAA compliance, however, they are united by the purpose of supporting medical personnel and researchers by applying AI to areas traditionally seen as time consuming and mundane. 

Abridge AI 

Abridge’s AI Scribe is a HIPAA compliant EHR built-in tool that allows clinicians to decrease the need for administrative assistants. The tool transcribes the conversation between the clinician and patient and generates accurate clinical notes for HER. According to the Abridge, this cuts down on the two hours a day spent catching up on clinical documentation, decreasing the likelihood of burnout in medical providers. By taking away the need to write down notes during the session, it allows doctors to focus on the patient, increasing patient satisfaction and ability to spend more time checking and discussing issues. On the back end, Abridge is trained on healthcare-specific data from over fifty specialties and 28 languages, is monitored and updated regularly based on clinicians’ feedback, and requires providers to go over notes and approve them prior to EHR upload. Users of Abridge AI include Johns Hopkins, Duke, and Mayo Clinic.  

Freed AI  

A competitor to Abridge AI, Freed is a ready-to-use AI-powered scribe for clinicians, used to transcribe and generate notes from a doctor-patient meeting. Freed allows clinicians to use AI to transcribe patient meetings and directly paste generated chart notes into the EHR system. As it is not directly integrated into EHR and lets doctors delete notes following the appointment, it allows the company to circumvent any HIPAA rules. Similarly to Abridge, any large language models which support medical chart creation allow doctors to save time and build more meaningful connections with patients.  

Critical Care Suite 

GE HealthCare’s Critical Care Suite is an on-device AI suite used to support the detection and triaging of critical conditions. It uses AI algorithms to identify pneumothorax, known as collapsed lung, and to provide more accurate positioning for a endotracheal tube. The tool, trained on x-ray imaging, is able to identify a possible collapsed lung and sends it to the top of the review queue of radiologists. Its integration directly into the X-ray device allows to cut down on detection time and make intubation outside of operating room more accurate, decreasing severe complication.  

AlphaFold 

AI does not have to be used directly in patient interactions or diagnosis. Google DeepMind’s AlphaFold is an AI system that predicts a protein’s 3D structure from its amino acid sequence. Prior to launch of AlphaFold identifying singular protein structures required years of dedicated analysis and incurred significant costs. Since 2020, AlphaFold predicted over 200 million catalogued proteins, including their possible interaction predictions with other biomolecules, and made them publicly available to researchers and medical personnel. AlphaFold’s protein outputs have been used in case studies including stopping malaria, Parkinson’s treatment, and designing more effective drugs. Given that AlphaFold identified almost all known proteins, is actively used by reputable institutions, and is open access for researchers, it provides the benefits of AI and ML in research, comes from a vetted source using least biased data, and does not interact directly with patients, possibly violating HIPAA laws.  

Concerns 

While AI tools can cut down on documentation time, decrease burnout among doctors, or predict the structures of almost all existing proteins, its implementation into a hospital setting needs to gradual and conscious, as there are numerous concerns and risks.  

As all AI tools, medical AI tools require significant amount of data to train and learn patterns and structures. Given American HIPAA laws, that requires the companies to partner with “covered entities,” to use health data for AI training. The data can be used only if patients gave explicit consent and the data has been anonymized. However, this limits the training dataset, which could result in biased output and could result in hallucinations or wrong diagnosis. While companies try to circumvent it by requiring providers to go over outputs and approve them, a high rate of correct output could lead to clinicians learning to trust the tool and approve everything without a detailed check.  

Additionally, while AI transcribing tools are trained on multiple languages, they are not as accurate when dealing with heavily accented speech. Given that a number of patients in the US speak English as second language, with varying levels of accents, medical chart notes riddled with mistakes could significantly impact providers. Similarly, even native language speakers could struggle with the names of medications, known to be complicated. A nurse clinician who uses Freed AI noted that “a patient said they took their medication Ubrelvy and Freed AI did not correctly abstract the information.” While doctors are able to manually edit the inputs, if they were focusing on the patient instead of writing down notes, they might have a lower recollection of details or could let issues slip. Given that EHR records are used in subsequent visits and treatments, any mistakes or misspeaks could become deadly.  

The legality of diagnosis AIs is also a significant concern for patients, doctors, and AI tool developers. There is an ongoing discussion on liability for mistakes made by AI tools. Detection and diagnosis tools, especially ones like GE’s Critical Care Suit, can move X-rays identified as high risk to the top of the review queue for radiologists. While the suite is approved by the FDA and has an over 94 percent specificity and an AUC of 0.96, there is still a possibility of a false positive or false negative result which could lead to permanent health damage, either for the patient with false negative, or to patients who were dropped down in a queue after another person’s false positive. Without existing precedence cases, this could constitute a medical malpractice lawsuit against the doctors, the medical facility which purchased and deployed the AI tool, or the tool developer. This would be a new territory for judicial proceedings, one that could make or break the medical AI space. 

Conclusion 

AI is reshaping medicine and healthcare by improving efficiency, supporting diagnosis, and accelerating research. AI tools offer significant support to clinicians, benefiting both healthcare providers and patients. However, challenges such as data privacy, transcription errors, and legal uncertainty must be carefully addressed.