Predicting critical alarms
Embodiments propose methods and system for predicting the occurrence of critical alarms in response to the occurrence of less severe, non-critical alarms. It is proposed to use a machine-learning model trained to discern whether a non-critical alarm will be followed by a critical alarm within a particular time period, e.g. whether the non-critical alarm will develop into a critical alarm. Unlike existing alarm systems which are merely threshold based, this approach uses physiological data from a window of data. This window of data can be expected to carry more information than a simple breach of the threshold.
1. A computer-implemented method of predicting the occurrence of a critical alarm for a subject undergoing physiological parameter monitoring, the computer-implemented method comprising:
receiving a non-critical alarm signal indicating that the subject has entered a clinically undesirable state at a time of occurrence; and
in response to the non-critical alarm signal:
obtaining a set of one or more pre-alarm values derived from and/or comprising one or more values of at least one monitored physiological parameter of the subject collected within a pre-alarm time window (TW 1 ) of a first predetermined length, wherein the start time (t tw ) of the pre-alarm time window depends upon the time of occurrence (t o ) of the non-critical alarm signal and wherein the pre-alarm time window ends before or at the time of occurrence of the non-critical alarm signal; and
processing, using a machine-learning algorithm, the set of one or more pre-alarm values to generate a predictive indicator indicating a probability that the non-critical alarm signal will be followed, within a post-alarm time window (TW 2 ), by a critical alarm signal indicating that the subject has entered a clinically actionable state,
wherein the post-alarm time window is of a second predetermined length and begins at the time of occurrence of the non-critical alarm signal.
2. The computer-implemented method of claim 1 , wherein the pre-alarm time window (TW 1 ) ends at the time of occurrence (t o ) of the non-critical alarm signal.
3. The computer-implemented method of claim 1 , wherein the first predetermined length is greater than the second predetermined length.
4. The computer-implemented method of claim 1 , wherein the first predetermined length is from 1 to 3 minutes, and the second predetermined length is from 1 to 3 minutes.
5. The computer-implemented method of claim 1 , further comprising, in response to the non-critical alarm, obtaining information about any other alarm for the subject occurring during the pre-alarm time window,
wherein the step of processing the obtained values comprises processing at least the obtained set of one or more pre-alarm values and the information about any other alarm using the machine-learning algorithm to thereby generate the predictive indicator.
6. The computer-implemented method of claim 1 , further comprising, in response to the non-critical alarm, obtaining metadata of the subject undergoing physiological parameter monitoring,
wherein the step of processing the obtained values comprises processing at least the obtained set of one or more pre-alarm values and the metadata of the subject using the machine-learning algorithm to thereby generate the predictive indicator.
7. The computer-implemented method of claim 1 , further comprising, in response to the non-critical alarm, determining at least one correlation measure indicative of a correlation between two or more values of at least one monitored physiological parameter,
wherein the step of obtaining a set of one or more pre-alarm values comprises including the at least one correlation measure in the set of one or more pre-alarm values.
8. The computer-implemented method of claim 1 , wherein:
the predictive indicator is a binary output indicating a prediction of whether or not the non-critical alarm will be followed by a critical alarm within the post-alarm time window; and
the machine-learning algorithm is configured to have a specificity of no less than 0.95.
9. The computer-implemented method of claim 1 , further comprising, in response to the predictive indicator indicating that a likelihood that the non-critical alarm will develop into a critical alarm is at or above a predetermined threshold, generating a first clinician perceptible alert.
10. The computer-implemented method of claim 9 , further comprising, in response to the predictive indicator indicating that a likelihood that the non-critical alarm will develop into a critical alarm is below a predetermined threshold, not generating the first clinician perceptible alert.
11. The computer-implemented method of any of claim 1 , further comprising:
obtaining metadata of the subject undergoing physiological parameter monitoring; and
setting the first predetermined length based on the obtained metadata of the subject.
12. The computer-implemented method of claim 1 , wherein the non-critical alarm signal indicates that at least one physiological parameter of the subject has entered a clinically undesirable state, the method further comprising, in response to the non-critical alarm:
obtaining one or more values for the at least one physiological parameter that triggers the non-critical alarm at a time the non-critical alarm is triggered;
modifying the first predetermined length based on the obtained one or more values for the at least one physiological parameter that triggers the non-critical alarm at a time the non-critical alarm is triggered.
13. A non-transitory computer-readable medium that stores therein a computer program product, which, when executed on a processor, causes the method of claim 1 to be performed.
14. A system for predicting the occurrence of a critical alarm for a subject undergoing physiological parameter monitoring, the system comprising:
an alarm receiving module adapted to receive a non-critical alarm signal indicating that the subject has entered a clinically undesirable state; and
an alarm predicting module adapted to, in response to the non-critical alarm:
obtain a set of one or more pre-alarm values derived from and/or comprising values of at least one monitored physiological parameter of the subject collected within a pre-alarm time window (TW 1 ) of a first predetermined length, wherein the start time (t tw ) of the pre-alarm time window depends upon the time of occurrence (t o ) of the non-critical alarm signal and wherein the pre-alarm time window ends before or at the time of occurrence of the non-critical alarm signal; and
process, using a machine-learning algorithm, the set of one or more pre-alarm values to generate a predictive indicator indicating a probability that the non-critical alarm signal will be followed, within a post-alarm time window (TW 2 ), by a critical alarm signal indicating that the subject has entered a clinically actionable state,
wherein the post-alarm time window is of a second predetermined length and begins at the time of occurrence of the non-critical alarm signal.
15. The system of claim 14 , further comprising a user interface arranged to generate a clinician perceptible alert in response to the predictive indicator indicating that a likelihood that the non-critical alarm will develop into a critical alarm is above a predetermined threshold.