Machine learning models lack in predicting in-hospital mortality rates

It can be extremely beneficial for doctors looking to save lives in intensive care units when a patient's condition deteriorates rapidly or shows vitals to a very unusual extent.

While current machine learning models are trying to achieve that goal, a Virginia Tech study published in Communication Medicine recently refers to predicting the likelihood of death in a hospital death patient, and not recognizing 66% of injuries.

Prediction is only valuable if you can accurately recognize the condition of a critical patient. You need to be able to identify patients with poor health and alert doctors promptly. ”

Dangfen “Daphne” Yao, Professor of Computer Science and Affiliate Faculty at Sangani Artificial Intelligence and Data Analysis Center

“Our research has discovered a serious flaw in the responsiveness of current machine learning models,” Yao said. “Most of the models we assess cannot recognize important health events, which creates a major problem.”

To conduct their research, Yao and PhD student Tanmoy Sarkar Pierces are as follows:

• Charmin Afroz, Oak Ridge National Laboratory, Tennessee
• Greenlife Medical College Hospital, Moondasturi, Dhaka, Bangladesh
• Ipsita Hamid Trisha, Banner University Medical Center, Tucson, University of Arizona School of Medicine
• Xinwei Deng, Virginia Tech's Bureau of Statistics
• Charles B. Nemerov, Faculty of Psychiatry and Behavioral Sciences, University of Texas Austindel Medical University

Their paper, “Low Responsiveness of Machine Learning Models to Serious or Deteriorating Health Conditions,” shows that patient data is not sufficient to teach models how to determine future health risks. Calibrating healthcare models using “test patients” helps to reveal the true capabilities and limitations of the model.

The team has developed multiple medical testing approaches, including gradient elevation methods and neural activation maps. The color change in the neuroactivation map shows how well a machine learning model responds to worsening patient conditions. The gradient upgrading method can automatically generate special test cases, making it easier to make the model quality easier.

“We systematically evaluated the ability of machine learning models to respond to serious medical conditions using new test cases, some of which are time series. That is, we use a set of observations collected regularly to predict future value,” PIAS said. “Our physician-guided assessment included four datasets: multiple machine learning models, optimization techniques, and two clinical prediction tasks.”

In addition to models that did not recognize 66% of injuries in hospital mortality forecasts, the models, in some cases, were unable to generate appropriate death risk scores in all test cases. This study identified similar defects in the reactivity of a 5-year breast cancer prognosis model.

These findings inform future healthcare research using machine learning and artificial intelligence (AI), as Yao shows that statistical machine learning models trained from patient data alone are grossly insufficient and have many dangerous blind spots. Strategically developed synthetic samples can be utilized to diversify training data. This is an approach that Yao's team will consider in 2022 to increase predictive equity for minority patients.

“The more basic design is to incorporate medical knowledge into clinical machine learning models,” she said. “It's a very interdisciplinary job and requires a large team with both computing and medical expertise.”

In the meantime, Yao's group has actively tested other medical models, including large language models, for safety and efficacy in time-dependent clinical tasks such as sepsis detection.

“AI safety testing is a competition from time as businesses are putting their products into the healthcare sector,” she said. “A transparent and objective test is a must. AI testing helps protect people's lives, and that's what my group is committed to.”