Machine learning models cannot detect key addicts, research shows

It would be very beneficial for doctors who try to save lives in intensive care units if they can be warned when the condition of a patient quickly deteriorates or shows vitals in very abnormal reach.

While the current models for machine learning try to achieve that goal, a Virginia Tech study has been published in Communication medication Show that they fail with models for predictions in the hospital, which refers to predicting the probability that a patient dies in the hospital and does not recognize 66% of the injuries.

“Predictions are only valuable if they can accurately recognize critical patient disorders. They must be able to identify patients with deteriorating health problems and warn doctors immediately,” said Dafeng “Daphne” Yao, professor in the Department of Computer Science and Affiliate Faculty Member at the Sangani -Palysis.

“Our study found serious shortcomings in the responsiveness of current models for machine learning,” said Yao. “Most models that we have evaluated cannot recognize critical health events and that is a major problem.”

To conduct their research, Yao and Computer Science Ph.D. Student Tanmoy Sarkar Pias collaborated with a number of researchers.

Their article, “low responsiveness of models for machine learning on critical or deteriorating health problems”, shows that patient data is not sufficient to learn models how they can determine future health risks. Calibrating health care models with “test patients” helps to reveal the true power and limitations of the models.

The team developed several medical test approaches, including a gradient -rising method and neural activation card. Color changes in the neural activation card indicate how well models for machine learning react to worse patient conditions. The Gradient Ascent method can automatically generate special test cases, making it easier to evaluate the quality of a model.

“We have systematically assessed the ability of Machine Learning models to respond to serious medical disorders using new test cases, some of which mean they use a series of observations that are collected with regular intervals to predict future values,” Pias said.

“Led by doctors, our evaluation included several models for machine learning, optimization techniques and four data sets for two clinical prediction tasks.”

In addition to models that do not recognize 66% of the injury in the hospital, in some cases the models have not generated adequate mortality risk scores for all test cases in some cases. The study identified similar shortcomings in the reaction capacity of the five-year breast and lung cancer prognosis models.

These findings inform future health care research using machine learning and artificial intelligence (AI), said Yao, because they demonstrate that statistical models for machine learning that have been trained exclusively from patient data are not sufficiently sufficient and have many dangerous blind spots.

To diversify training data, strategically developed synthetic samples can be used, An approach from YAO investigated in 2022 to improve the fairness of the prediction for minority patients.

“A more fundamental design is to take medical knowledge deep in models for clinical machine learning,” she said. “This is very interdisciplinary work and requires a large team with both computing and medical expertise.”

In the meantime, the YAO group actively tests other medical models, including large language models, on their safety and efficacy in time -sensitive clinical tasks, such as sepsis detection.

“AI security tests is a race by time, because companies pour products into the medical space,” she said. “Transparent and objective testing is a must. AI tests helps to protect people’s lives and that is where my group strikes.”