A live emergency-call alert may not improve dispatcher recognition despite higher standalone model sensitivity.
Collections: Negative results
Executive brief
A better classifier is not a better operating model. The alert's low precision and insertion into a time-critical conversation prevented technical sensitivity from becoming human performance.
AI value · Dispatcher recognition of confirmed out-of-hospital cardiac arrest
Verified93.1% with the alert, P=.15; no statistically significant improvement
High-quality randomized negative result. The model alone was more sensitive but materially less specific and had lower positive predictive value than dispatchers.
Before
Listen to the caller, recognize suspected cardiac arrest, and initiate the protocol. → Coach CPR and dispatch resources when cardiac arrest is recognized.
After
Listens to ongoing calls and flags suspected cardiac arrest. → Considers the alert and decides whether to recognize OHCA and initiate the standard response.
Human boundary
The model alerts; the dispatcher decides whether to declare suspected OHCA and start CPR coaching and dispatch.
Why it matters
A live emergency-call alert may not improve dispatcher recognition despite higher standalone model sensitivity.
Before
How the work ran before the change.
Medical dispatcher
Listen to the caller, recognize suspected cardiac arrest, and initiate the protocol.
ControlStandard dispatch protocol
Dispatcher
Coach CPR and dispatch resources when cardiac arrest is recognized.
ControlHuman clinical judgment
What changed
A live emergency-call alert may not improve dispatcher recognition despite higher standalone model sensitivity.
Decision rightHuman moves from creator to judge
After
How the same work runs now.
Speech-recognition ML model
Listens to ongoing calls and flags suspected cardiac arrest.
ControlModel alert
Medical dispatcher
Considers the alert and decides whether to recognize OHCA and initiate the standard response.
ControlDispatcher retains all response rights
Exception path
Dispatchers ignore false alerts and follow standard protocols; no response is automatically triggered by the model.
Decision authority
The model alerts; the dispatcher decides whether to declare suspected OHCA and start CPR coaching and dispatch.
| # | Actor | Action | Control |
|---|---|---|---|
| 01 | Medical dispatcher | Listen to the caller, recognize suspected cardiac arrest, and initiate the protocol. | Standard dispatch protocol |
| 02 | Dispatcher | Coach CPR and dispatch resources when cardiac arrest is recognized. | Human clinical judgment |
| # | Actor | Action | Control |
|---|---|---|---|
| 01 | Speech-recognition ML model | Listens to ongoing calls and flags suspected cardiac arrest. | Model alert |
| 02 | Medical dispatcher | Considers the alert and decides whether to recognize OHCA and initiate the standard response. | Dispatcher retains all response rights |
Work that left the path
Human role before
Dispatchers recognized cardiac arrest using the caller conversation and protocol.
Human role after
Dispatchers receive an additional alert but remain responsible for recognition and response.
AI role
Speech-recognition and classification model that identifies suspected cardiac arrest during live calls.
90.5% with standard protocol and no model alert93.1% with the alert, P=.15; no statistically significant improvement
2018-09-01 through 2019-12-31 · 169,049 calls screened; 5,242 suspected calls randomized; 654 confirmed OHCA cases
High-quality randomized negative result. The model alone was more sensitive but materially less specific and had lower positive predictive value than dispatchers.
Anti-pattern
Selecting on model sensitivity while ignoring alert precision and operator response.
Questions
Portability conditions
Reputation risk
low