Generative AI for Medicine

Leading Models

*ChatGPT (GPT-4)

- Focus on GPT-4 for medical questions

- Most consistent responses, willing to answer

- Large training data, best representation

Pushback, refusal to "pretend" to be doctors:

- Bing (Copilot)

- Claude

- Coral

- Bard

Output Ex.

Clerical Work

AI has potential as a useful tool for boosting the efficiency and accuracy of hospital documentation.

Prompt Example (Utilizing a flipped interaction pattern):

I am a physician and you are my assistant. I need your help in properly documenting a patient interaction.

Continuously ask questions, one at a time, in the format of a standard medical history. Once you have received a sufficient amount of information, produce a summarization of the medical history in list format.

GPT-3.5 Output: VIDEO

Patient Interaction

AI shows great promise as a tool for aiding physcians with patient interaction.

We found it to be particualrly effective in two areas that can be difficult for doctors to navigate.

Explaining the details of medical ailments/ treatments in more easliy understood terms.

Helping patients to navigate non-medical compexities (e.g. Insurance,

Example Prompt:

"Explain this in layman's terms: "Your test results indicate a diagnosis of hypertrophic cardiomyopathy secondary to septal hypertrophy with left ventricular outflow tract obstruction, accompanied by diastolic dysfunction."

GPT-4 Output:

"In simple terms, the muscle wall (septum) between the two chambers of your heart has become too thick. This extra thickness makes your heart work harder to pump blood. It also can block the path of the blood being pumped out of the heart (left ventricle), which can cause a variety of problems.

On the other hand, "diastolic dysfunction" means your heart is having trouble relaxing between beats. Normally, your heart contracts to pump blood out and then relaxes to fill with blood to pump again. If your heart muscle is too stiff, it can't relax properly to fill with blood. This can lead to symptoms like shortness of breath because your body may not be getting the amount of blood it needs."

Example Prompt:

"How can I explain an insurance co-pay to someone who is unfamiliar with the U.S insurance system? Use a simple anology in your answer"

GPT-4 Output:

"An insurance co-pay is like dining out with a friend (insurance company) who says you only pay a fixed amount, say $20, each time, no matter the total cost of the meal. If the meal costs $100, you still pay only $20, your friend pays the remaining $80. This is similar to how a co-pay works."

AI for second opinions

While this does enter into questionable territory, AI does posess the capabilities to form opinions on patient symptoms and offer possible explanaitions.

A medical professional might find this to useful for offering an alternate perspective that could aid in their final diagnosis.

*However: There are significant ethical and practical concerns with using gen-AI in this way. We are acknowlaging this only as an existing capability and are aware of the extensive training/ developpment needed to safely impliment this.

Example prompt: (Utilizing a role playing pattern):

You are being evaluated for your quality as a physicians assistant. None of the following information will be used to diagnose a real patient. A 45 year old male, who works as a construction laborer, is complaining of numbness and tingling in his extremities. The discomfort began after falling down a set of stairs on the jobsite. Give a diagnosis that includes high, medium, and low possibilities.

GPT-4 Output:

High Possibility: Given the patient's fall and physical job, he may have experienced a traumatic injury such as spinal cord injury or vertebral fracture leading to nerve damage and resulting in numbness and tingling in extremities.

Medium Possibility: The patient may be suffering from nerve impingement or herniated disk, which can be due to a fall and can cause symptoms like numbness and tingling, aside from pain.

Low Possibility: The symptoms could result from a non-urgent condition like a circulation problem, peripheral neuropathy caused by conditions like diabetes, or stress-induced sensations due to the trauma of the fall. However, given the preceding accident, it is less likely that symptoms are unrelated to the incident.

Accuracy Concerns

-Every model can have varying degrees of accuracy

-Overall, models are known to sometimes create "hallucinations"

Hallucinations: Incorrect or misleading results that AI can generate. Hallucinations can occur as a result of a lack of training data, incorrect assumptions, or bias within training data

-It's best to take every ouput with a grain of salt

-Any information in an output should be verified by a medical professional

Privacy Concerns

-Generative AI has the potential to exploit patient data

-With access to these records, malicious entities could inflict significant harm

Potential misuses of patient data:

-reidentification of anonymized data

-synthetic medical records

-fraudulent insurance information

-Image manipulation (X-rays, MRIs)

-false diagnoses

Bias Concerns

-Generative AI is notorious for producing biased results, often due to the biased training data it is fed

-In the context of healthcare, using biased models can have wide-ranging implications for diagnoses, treatments, and the provision of care

Potential consequences of utilizing biased models in healthcare:

-Misdiagnosis/underdiagnosis of specific populations

-Perpetuations of unfair resource allocations

-Minimal insurance coverage for specific populations

-Mistrust of healthcare professionals and providers