Virginia Tech Study Finds ML Models Miss 66% of Critical Patient Deteriorations
A Virginia Tech study published in Communications Medicine found that machine learning models used to predict in-hospital mortality and detect critical patient deteriorations failed to recognize 66 percent of injuries. The research revealed serious deficiencies in the models' responsiveness to critical health events, raising concerns about their reliability for alerting physicians to patients with worsening conditions.
What happened
A Virginia Tech study published in Communications Medicine evaluated machine learning models designed to predict in-hospital mortality and alert physicians to patient deterioration. The study found that these models failed to recognize 66 percent of injuries in their assessment of in-hospital mortality prediction.
Reported impact
The failure to recognize critical health events poses a major problem for physicians trying to save lives in intensive care units, as predictions are only valuable if they accurately recognize critical patient conditions and can identify patients with worsening health conditions to alert doctors promptly.
Index analysis
The team developed multiple medical testing approaches, including a gradient ascent method and neural activation map, to systematically assess machine learning models' ability to respond to serious medical conditions. The evaluation involved multiple machine learning models, optimization techniques, and four data sets for two clinical prediction tasks, guided by medical doctors.
Governance controls
The study found that patient data alone is not enough to teach models how to determine future health risks, and calibrating health care models with test patients helps reveal the models' true ability and limitations.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.