Critical Illness
Conditions
Brief summary
Alerts related to outlier clinician behavior are generated in real-time by an intelligent system continuously scraping EHR (electronic health record) data. These alerts are passed to the bedside and their potential impact on bedside clinical behavior is evaluated.
Detailed description
A clinician-informed AI model will generate outlier alerts from real-time review of the EHR (electronic health record) of UPMC Presbyterian/Montefiore ICU patients. These alerts will first be reviewed by an ICU clinician, along with the patients' EHR, for clinical relevance. For those alerts deemed potentially relevant, the ICU clinician will contact the treating ICU clinician (eg, an ICU pharmacist, physician, advanced practice provider) and discuss the alert. The treating ICU clinician will take whatever action, including no action, they deem best.
Interventions
DEVICERevealed Alerts
Bedside reveal of alerts generated by the alerting system
DEVICEUnrevealed Alerts
Alerts will be generated but not revealed.
Sponsors
David T Huang
National Institute for Biomedical Imaging and Bioengineering (NIBIB)
Study design
Allocation
RANDOMIZED
Intervention model
SEQUENTIAL
Primary purpose
TREATMENT
Masking
NONE
Intervention model description
Step wedge RCT
Eligibility
Sex/Gender
ALL
Age
18 Years to 100 Years
Healthy volunteers
No
Inclusion criteria
* All patients in the Presbyterian and Montefiore ICUs
Exclusion criteria
* None
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Rate of Clinical Actions Following Revealed vs. Non-Revealed Alerts | From time of ICU admission until ICU discharge | The primary outcome compares the proportion of alerts that lead to documented clinical actions when revealed to treating ICU clinicians versus when not revealed. Alerts are generated by a decision support system and reviewed daily by ICU study clinicians. On average, 30 alerts are reviewed per ICU per day, with approximately 5 alerts revealed to the treating clinicians. The analysis uses a stepped wedge design with ICU beds as the unit of analysis, where each ICU acts as its own control. The outcome will assess whether revealing alerts increases the rate of appropriate clinical actions taken, as compared to when alerts are withheld. |
Secondary
| Measure | Time frame | Description |
|---|---|---|
| Rate of Any Clinically Responsive Action Following Alerts | Up to 90 days after ICU admission | For each alert, a range of clinical actions may be deemed "responsive," including but not limited to the specific recommended action. This outcome assesses the differential rate of any clinically appropriate response (whether or not it matches the recommended action) between alerts that are revealed versus not revealed to treating ICU clinicians. |
| True Positive Alert Rate (TPAR) by Study Group and Alert Type | Up to 90 days after ICU admission | This outcome evaluates the overall and alert-specific True Positive Alert Rate (TPAR), defined as the proportion of alerts that are associated with a clinically appropriate action. TPARs will be calculated separately for the intervention group (alerts revealed), the control group (alerts not revealed), and the combined population. Comparisons will assess whether revealing alerts is associated with a higher TPAR across alert categories. |
| Rate of Non-Responsive Actions Following Alerts | Up to 90 days after ICU admission | This outcome assesses the difference in the rate of clinical actions that are not considered responsive to the alert (i.e., actions taken that do not address the alert's content or recommended intervention) between the intervention group (alerts revealed) and the control group (alerts not revealed). This helps evaluate potential unintended or off-target responses to alerting. |
| Overall Alert Rate | Through study completion, an average of 2 years | Total number of alerts generated per ICU per day. |
| Alert Rate per Alerting Model | Daily, up to 90 days | Number of alerts generated per model type per ICU per day. |
| Delay Between Alert Generation and First Responsive Action | Measured continuously through study completion, an average of 2 years | Median time from alert generation to the first documented responsive clinical action. |
| ICU Length of Stay | Up to 90 days after ICU admission | Total number of days each participant spends in the ICU during the index hospitalization. |
| Hospital Length of Stay | up to 90 days after hospital admission | Total number of days each participant spends in the hospital during the index hospitalization. |
| In-Hospital Mortality | Up to 90 days after hospital admission | Proportion of participants who die during the index hospital stay. |
| Time Trend of True Positive Alert Rate (TPAR) | Through study completion, an average of 2 years | Evaluate changes in the True Positive Alert Rate (TPAR) over time during the study period to assess model performance stability and potential temporal variation in responsiveness. |
Countries
United States
Contacts
PRINCIPAL_INVESTIGATORDavid Huang, MD
University of Pittsburgh
Outcome results
None listed