A Health Record That Flags Partner Violence Three Years Before a Patient Enrolls in a Program
Artificial intelligence can now flag people who may be at elevated risk of intimate partner violence years before they disclose the violence. On average, the models identified risk of IPV more than three years before patients enrolled in a hospital-based domestic abuse intervention program.
The researchers added a caveat that is easy to skim past: careful performance evaluation on the more general population of female patients should be conducted before the models are implemented for clinical use. A prediction that is statistically early is not automatically a prediction a clinic should act on.
There is a second question that accuracy scores don't answer. Early knowledge can give a provider the chance to offer support; it also means a system holds an inference about a patient she has not chosen to share, and the first reader of that flag is a clinician, not her.
AI-generated discussion · Different perspectives
The asymmetry is the crux. The record can hold an inference about a patient that the patient has not decided to make explicit anywhere.
The researchers want the models evaluated on the more general population of female patients before clinical use, which is their own way of saying the current evidence isn't deployment-ready.
Even a cautious flag changes the room. "Elevated risk" is a probability attached to a person, and it can shape how the next appointment goes before anyone has offered a service.