Physicians’ behaviors are nearly all AI needs to head off faulty drug prescriptions

Contrary to intuitive expectations, many errors in drug ordering are caused or worsened by the intricacies of the EHR. These problems often stem from mistakes made at the point of the clinician–computer interface.

To avert the resulting risk, researchers have designed an AI-based system that can flag errors going just by the ordering clinician’s behavior and the context in which it occurs.

The innovation may not only reduce prescription-drug errors but also relieve pharmacists’ workloads while better ensuring patient privacy and security, the team suggests.

The work was carried out at New York University and is described in a study published Oct. 5 in JAMIA Open.

To build and test their system, Martina Balestra, PhD, of NYU’s Center for Urban Science and Progress in Brooklyn and colleagues collected data on drug prescribers’ actions over a two-week period at an academic medical center.

From this data they built a machine learning classification model to predict drug orders questionable enough that they needed a pharmacist to stop and examine.

Internally validating the model, they found it had an area under the receiver-operator characteristic curve of 0.91 and an area under the precision-recall curve of 0.44.

Balestra and co-authors note that conventional EHR drug alerting has entailed reviewing drug orders alongside the patient’s medical records.

The shortcoming in this approach, they maintain, is its lack of attention to the EHR’s inherent vulnerabilities to error and how these potential fail points may show up in the orderer’s actions.

By contrast, the NYU team’s machine learning modeling

offers us a novel perspective on the factors influencing order entry by focusing on the behavior of the provider and errors that arise from the workflow around the EHR. Whereas previous models predicting errors ingest patients’ medical records, by focusing on the behavior of the clinician, we also reduce the risk to the privacy and security of these patients’ data while still being useful to pharmacists.”

The study is available in full for free.

Around the web

The AI approach may prove more effective than PET/CT, which can only diagnose N2 disease with 70-85% sensitivity, researchers reported Tuesday in Radiology.

The results could lead to expanded use of chest radiography in interstitial lung disease clinics, experts reported in the American Journal of Roentgenology.

Experts say this is the first study to assess how preprocessing techniques impact providers' ability to routinely extract similar quantitative imaging features from US scans.

Trimed Popup
Trimed Popup