Events Calendar

24 Mar
ISP AI Forum: "Using the SHAP Method to Produce Explanations of Predictions in Clinical Alerting Systems"
Event Type

Forums

Topic

Health & Wellness, Research, Technology

Target Audience

Faculty, Graduate Students

Website

https://pitt.co1.qualtrics.com/jfe/fo...

University Unit
Intelligent Systems Program
Subscribe
Google Calendar iCal Outlook

ISP AI Forum: "Using the SHAP Method to Produce Explanations of Predictions in Clinical Alerting Systems"

This is a past event.

AbstractMedical errors are a key challenge in healthcare and a leading cause of death in the U.S. One approach to error reduction involves automated alerting on clinical actions that are likely to be errors. Outlier-based clinical alerting performs this task by first constructing a model of typical patient care from an electronic medical record (EMR) system. It then uses that model to determine if the current care for a patient is unusual in some way; if so, an electronic alert about the outlier action(s) is sent to the clinician caring for the patient. The clinician may want to know why the system considers an action to be an outlier, and thus, possibly an error. To provide an explanation, we are exploring the use of the SHAP method to derive the relative importance of the model features in determining why the model labeled an action is unusual. In particular, the probability of a clinical action is explained in terms of a linear combination of Shapley values of the features in the model. Thus, the contribution of each feature toward deriving that probability is made explicit. A bedside clinician can use such explanations to help decide whether to heed an alert. Explanations may also provide insights into how to improve system performance through a deeper understanding of the features’ mapping to the output. This talk will describe preliminary work in using the SHAP method to explain alerts on clinical actions taken in the intensive care unit.

Bio: Joshua Anderson is a PhD student in the Intelligent Systems program. His research interests include Causal Inference, Interpretable AI, Computer Vision, and AI in Medicine. 

RSVP for Zoom meeting informationhttps://pitt.co1.qualtrics.com/jfe/form/SV_e2Udcl6qilgFq9o

Friday, March 24 at 1:00 p.m. to 1:30 p.m.

Virtual Event

ISP AI Forum: "Using the SHAP Method to Produce Explanations of Predictions in Clinical Alerting Systems"

AbstractMedical errors are a key challenge in healthcare and a leading cause of death in the U.S. One approach to error reduction involves automated alerting on clinical actions that are likely to be errors. Outlier-based clinical alerting performs this task by first constructing a model of typical patient care from an electronic medical record (EMR) system. It then uses that model to determine if the current care for a patient is unusual in some way; if so, an electronic alert about the outlier action(s) is sent to the clinician caring for the patient. The clinician may want to know why the system considers an action to be an outlier, and thus, possibly an error. To provide an explanation, we are exploring the use of the SHAP method to derive the relative importance of the model features in determining why the model labeled an action is unusual. In particular, the probability of a clinical action is explained in terms of a linear combination of Shapley values of the features in the model. Thus, the contribution of each feature toward deriving that probability is made explicit. A bedside clinician can use such explanations to help decide whether to heed an alert. Explanations may also provide insights into how to improve system performance through a deeper understanding of the features’ mapping to the output. This talk will describe preliminary work in using the SHAP method to explain alerts on clinical actions taken in the intensive care unit.

Bio: Joshua Anderson is a PhD student in the Intelligent Systems program. His research interests include Causal Inference, Interpretable AI, Computer Vision, and AI in Medicine. 

RSVP for Zoom meeting informationhttps://pitt.co1.qualtrics.com/jfe/form/SV_e2Udcl6qilgFq9o

Friday, March 24 at 1:00 p.m. to 1:30 p.m.

Virtual Event

Event Type

Forums

Target Audience

Faculty, Graduate Students

University Unit
Intelligent Systems Program

Powered by the Localist Community Events Calendar ©