Break “The Empire of Intellectual”: How artificial intelligence makes an unbalanced account in health care

Whenever the patient gets a CT scan in Texas Medical University branch (UTMB), the resulting images are automatically sent to the heart disease department, analyzed by artificial intelligence and set the heart risk.

Within just a few months, thanks to a simple algorithm, artificial intelligence informed many patients at risk of high cardiovascular and blood vessels. CT scan does not have to be linked to the heart; The patient does not have to have heart problems. Each examination automatically leads to evaluation.

It is clear and direct Preventive care Enabled by artificial intelligence, allowing the medical facility to finally start using huge amounts of data.

“The data is sitting there,” Peter Antarri, the chief artificial intelligence officer at UTMB. “What I love in this matter is that Amnesty International should not do anything super. It performs a low thinking task, but in a very large size, and this still provides a lot of value, because we constantly find things we miss.”

I admit, “We know we miss things. Before, we did not have the tools to return and find them.”

How to help artificial intelligence UTMB to determine the risk of cardiovascular

Like a lot Healthcare facilitiesUTMB applies artificial intelligence across a number of areas. One of the first cases of use is to examine the heart risk. Models have been trained on scanning for casual coronary artery (ICAC), which is a strong indicator of cardiovascular risk. McFari explained that the goal is to identify patients exposed to heart disease who may have been ignored in another way because they do not show any clear symptoms.

Through the examination program, each CT scan is analyzed in the facility automatically using AI to detect coronary calcification. The examination should not have any relationship with heart disease; This can be requested due to a fracture of the spine or an abnormal lung doctrine.

Surveying operations are feed in a CNN, which calculates the AGATSTON degree, which represents the accumulation of plaque in the arteries of the patient. Astrate explained that this is usually calculated by the human radiologist.

From there, artificial intelligence specializes patients who get an ICAC at or above 100 to three “risk levels” based on additional information (such as whether they are on two or they visited with a cardiologist). Astrate explained that this task depends on the rules and can be derived from separate values ​​within the electronic health registry (EHR), or artificial intelligence can define values ​​by treating a free text such as Clinical visit notes Using GPT-4O.

Digital messages are sent automatically from 100 or more than 100 or more, with no known date to visit heart or treatment, digital messages are sent. The system also sends a note to the main doctor. Patients identified as ICAC are more than 300 or also higher phone call.

Astrate explained that almost everything is automated, except for the phone call; However, the facility is also experimenting with tools in the hope of automating sound calls. The only area where humans are in the episode to confirm the degree of calcium derived from artificial intelligence and risk plane before continuing with an automatic notification.

Since the launch of the program in late 2024, the medical facility has evaluated about 450 examinations per month, with five to ten of these cases as highly dangerous every month, which requires intervention, according to Astari.

He pointed out that “GIST here is not anyone to suspect that you have this disease, and no one must order studying this disease.”

There is another critical use of AI to discover stroke and pulmonary blockage. UTMB uses specialized algorithms trained to discover specific symptoms and science care teams within seconds of photography to accelerate treatment.

As with the ICAC registration tool, CNNS, which is trained in a row on stroke and pulmonary blockage, automatically receives computerized tomography tests and searching for indicators such as gave blood flows or sudden blood vessels.

“Human radiologists can discover these visual characteristics, but here is an automatic detection and occurs in only seconds,” said Azzetri.

Any CT scan “under doubt” is sent to stroke or pulmonary blockage automatically to artificial intelligence – for example, the doctor may determine in the arishing of the face or fading and release the “CT stroke”, which leads to algorithm.

Both algorithms include a messaging application that notifies the entire care team as soon as the result is made. This will include a photo screen shot with an intersection on the lesion site.

“These are special cases of emergency use, as the speed of the start of treatment reaches.” “We have seen situations where we can get several minutes of intervention because we had faster heads of artificial intelligence.”

Reducing hallucinations, bias bonding

To ensure the performance of the models as perfectly as possible, the UTMB defines them with allergies, privacy, F-1 degree, bias and other factors alike before publication and after repeated publication.

Therefore, for example, the ICAC algorithm is validated before publishing by turning on the model on a balanced group of CT scans while the radiologists are manually recorded-then the two are compared. In the post -publication review, at the same time, radiologists are given a random sub -segment of AI’s tomatoes and carrying out a complete ICAC action on the degree of artificial intelligence. Astrate explained that this allows his team to calculate the form of the model frequently and also discover the potential bias (which will be seen as a shift in the size and/or the direction of the error).

To help prevent biasing bias – artificial and human intelligence depends heavily on the first part of the information they face, thus losing important details when making a decision – UTMB employ “peer learning” technique. A random sub -group of radiology tests is chosen, mixed, unidentified and distributed to different radiologists, and their answers are compared.

This not only helps to evaluate the performance of an individual radiologist, but also discovers whether the average results lost are higher in studies in which artificial intelligence was used to highlight specific views specifically (which leads to bias in connection).

For example, if artificial intelligence is used to determine bone fractures and inform them of X -rays, the team will look at whether studies containing bone fracture flaws have increased Miss rates for other factors such as narrowing the joint space (common in arthritis).

MCCAFAFREY and his team found that both consecutive versions in the seasons (different versions of the GPT-4O) and through the layers (GPT-4.5 versus 3.5) tend to have a low hallucinatory rate. He said: “But this is not zero and not specified, although it is nice-we can only ignore the possibility and repercussions of hallucinations.”

Therefore, they are usually attracted to the AI’s given tools that do a good job in citing their sources. For example, a model that summarizes the patient’s medical path with clinical observations that served as a basis for their output.

“This allows the provider to serve as protection against hallucinations,” said McAveri.

The “basic things” mark to enhance health care

UTMB also uses artificial intelligence in many other areas, including an automatic system that helps the medical staff determine whether internal admission operations are justified. The system works as a participant pilot, and automatically extracts all patient notes from EHR and use ClaudeAnd GPT and Gemini to summarize and examine them before submitting evaluations to employees.

“This allows our employees to look at all patients and candidates/sorting them,” I explained. The tool also helps employees to formulate documents to support admission or observation.

In other areas, artificial intelligence is used to re -examine reports such as heart science explanations or clinical notes and identify gaps in care. In many cases, “it is simply a sign of the basic things,” said Macafri.

He pointed out that health care is complicated, with data extracts from everywhere, as he indicated – photos, doctor’s notes, laboratory results – but very little of these data was calculated because there was not enough human human forces.

This led to what he described as “the huge and huge intellectual bottle neck.” Not a lot of data is simply calculated, although there are great potentials that are proactive and find things earlier.

“It is not an indictment for any specific place,” McAveri confirmed. “It is a state of health care in general.” In the absence of artificial intelligence, “You cannot spread intelligence, scrutiny, and work on the measure required to arrest everything.”