Systems and methods for individualized alertness predictions

• US 8,781,796 B2
• Filed: 10/24/2008
• Issued: 07/15/2014
• Est. Priority Date: 10/25/2007
• Status: Active Grant

• Current Assignee: Pulsar Informatics, Inc., Trustees Of The University Of Pennsylvania (University of Pennsylvania), Washington State University
• Original Assignee: Trustees Of The University Of Pennsylvania (University of Pennsylvania)
• Inventors: Mott, Christopher, Mollicone, Daniel, Van Dongen, Hans, Huang, Jen-Kuang, Dinges, David
• Primary Examiner(s): JONES, HUGH M
• Application Number:
  US12/739,653
• Publication Number:
  US 20120191425A1
• Time in Patent Office: 2,090 Days
• Field of Search: 703/2
• US Class Current: 703/2
• CPC Class Codes: A61B 5/16 Devices for psychotechnics using teaching or educational appliances G09B1/00 - G09B7/00; Testing reaction times ; Devices for evaluating the psychological state
• CPC Class Codes: G16H 50/30 for calculating health indices; for individual health risk assessment

Abstract

Systems and methods are provided for generating individualized predictions of alertness or performance for human subjects. Alertness or performance predictions may be individualized to incorporate a subject'"'"'s individual traits and/or individual states. These individual traits and/or individual states (or parameters which represent these individual traits and/or individual states) may be represented by random variables in a mathematical model of human alertness. The mathematical model and/or prediction techniques may incorporate effects of the subject'"'"'s sleep timing, the subject'"'"'s intake of biologically active agents (e.g. caffeine) and/or the subject'"'"'s circadian rhythms. The mathematical model and/or prediction techniques may incorporate feedback from the subject'"'"'s measured alertness and/or performance.

Claims

1. A method for estimating alertness of a human subject implemented by a processor, the method comprising:

receiving initial values for a plurality of model variables of a mathematical model, one or more of the model values comprising variables which specify or estimate probability distributions, the model variables including;

one or more individual trait variables, each individual trait variable comprising a parameter that is unique to the subject and which is generally constant over time; and

one or more individual state variables, each individual state variable comprising a time varying parameter;

receiving a sleep history input indicative of the subject'"'"'s asleep and awake status between a first time and a second time;

identifying, by the processor, one or more transition time points within the sleep history, each transition time point corresponding to one of;

the subject'"'"'s transition from awake to asleep status and the subject'"'"'s transition from asleep to awake status;

dividing, by the processor, a time between the first time and the second time into a plurality of time segments, each time segment extending between;

a corresponding time segment start time which is one of;

the first time and one of the one or more transition time points;

a corresponding time segment end time which is one of;

the one or more transition time points and the second time;

each time segment also associated with a sleep status value as indicated by the sleep history;

initializing, by the processor, the model variables at the first time to be the received initial values;

for each time segment, starting at a first time segment whose time segment start time corresponds to the first time through to a last time segment whose time segment end time corresponds to the second time;

using the model, by the processor, to estimate values of the model variables at the time segment end time, using the model comprising;

configuring the model based at least in part on the sleep status value of the time segment, and

basing the estimated values of the model variables at the time segment end time on the values of the model variables at the time segment start time and a duration of the time segment; and

for time segments other than the last time segment, setting, by the processor, the estimated values of the model variables at the time segment end time to be the values for the model variables at the time segment start time of a next time segment; and

estimating, by the processor, alertness values of the subject at the second time based at least in part on applying the model to the values of the model variables at the second time.

2. A method according to claim 1 wherein the alertness values comprise variables which specify or estimate probability distributions.

3. A method according to claim 1 wherein using the model to estimate values of the model variables at the time segment end time comprises receiving one or more inputs, other than the sleep history, which permit estimation of updated values of the individual state variables and using the model to estimate values of the model variables at the time segment end time based at least in part on the one or more inputs.

4. A method according to claim 3 wherein the model comprises a process noise component, the process noise component comprising a probability distribution representing an uncertainty associated with the one or more inputs.

5. A method according to claim 3 wherein using the model to estimate current values of the model variables based at least in part on the one or more inputs comprises using the model to estimate current values of the individual state variables based at least in part on the one or more inputs and maintaining the current values of the individual trait variables constant.

6. A method according to claim 1 wherein the first time corresponds to a current time, the second time corresponds to a future time and the sleep history input is representative of a simulated future sleep history for the subject.

7. A method according to claim 1 comprising receiving one or more alertness inputs, each alertness input comprising a measured value of the alertness of the subject and a corresponding measurement time;

wherein identifying, by the processor, one or more transition time points, further comprises for each alertness input identifying a transition time point at the corresponding measurement time if a transition point is not already identified at that time; and

wherein using the model to estimate values of the model variables at the time segment end time comprises determining if an alertness input exists for the time segment end time and, if it does, updating at least one of the model variables using the alertness input.

8. A method according to claim 7 wherein the model comprises a measurement noise component, the measurement noise component comprising a probability distribution representing an uncertainty associated with the one or more measured alertness values.

9. A method according to claim 2 wherein estimating alertness values of the subject comprises generating expected values for the future alertness values and one or more measures of uncertainty for the future alertness values.

10. A method according to claim 9 wherein the one or more measures for uncertainty comprises one or more of:

a percentile-based confidence interval;

a distribution function; and

a standard deviation of a distribution.

11. A method according to claim 1 wherein using the model to estimate values of the model variables and estimating alertness values of the subject comprise performing Bayesian recursive estimation.

12. A method according to claim 11 wherein using the model to estimate values of the model variables comprises approximating a solution to a prediction update equation of the form

p(x_k|U_k,Y_k-1,x₀)=∫

_-∞

^∞

p(x_k|x_k-1,u_k)p(x_k-1|U_k-1,Y_k-1,x₀)dx_k-1

where p(x_k-1|U_k-1, Y_k-1, x₀) is a prior probability distribution which describes probability distributions of the model variables x_k-1 at time t_k-1, given;

all prior inputs U_k-1, all prior alertness measurements Y_k-1, and initialized model variable probability distributions x₀; and

where p(x_k|x_k-1, u_k) is a transitional probability distribution which describes probability distributions of the model variables at time t_k, given;

inputs u_k between t_k and t_k-1 and the model variables x_k-1 at the time t_k-1.

13. A method according to claim 11 wherein estimating alertness values of the subject comprises approximating a solution to a measurement update equation of the form

p(x_k|U_k,Y_k,x₀)=Cp(y_k|x_k)p(x_k|U_k,Y_k-1,x₀)

where C is a normalization constant, p(x_k|U_k, Y_k-1, x₀) is a prior probability distribution of the model variables x_k-1 which describes the probability distribution of the model variables x_k at time t_k, given:

all prior inputs U_k up to time t_k, all prior alertness measurements Y_k-1 up to time t_k-1, and initialized model variable probability distribution x₀; and

p(y_k|x_k) is a measurement likelihood distribution which describes a probability distribution of observing alertness measurements y_k at time t_k.

14. A method according to claim 11 wherein performing Bayesian recursive estimation comprises using at least one of:

an Unscented Kalman Filter;

a Markov Chain Monte Carlo particle filter;

an Extended Kalman Filter;

Bayesian forecasting; and

a Bayesian grid search.

15. A method according to claim 1 wherein receiving initial values for the plurality of model variables comprises initializing the individual trait values with individual trait values previously determined for the subject.

16. A method according to claim 15 wherein the individual trait values previously determined for the subject comprise individual trait values determined from a previous application of the method.

17. A method according to claim 1 wherein receiving initial values for the plurality of model variables comprises initializing the individual trait values based on alertness data obtained from a sample population.

18. A method according to claim 17 wherein the alertness data obtained from the sample population is at least one of:

represented using a metric substantially similar to that of the alertness values estimated by the model; and

convertible to a metric substantially similar to that of the alertness values estimated by the model.

19. A method according to claim 17 wherein initializing the individual trait values based on alertness data obtained from the sample population comprises performing a maximum likelihood estimation which estimates parameter values making it maximally likely to observe the alertness data obtained from the sample population.

20. A method according to claim 1 wherein receiving initial values for the plurality of model variables comprises initializing the individual trait values to have at least one of:

a uniform probability distribution; and

a Gaussian probability distribution.

21. A method according to claim 1 wherein receiving initial values for the plurality of model variables comprises initializing the individual trait values using a combination of two or more of:

(i) individual trait values previously determined for the subject;

(ii) individual trait values based on alertness data obtained from a sample population;

(iii) individual trait values having a uniform probability distribution; and

(iv) individual trait value having a Gaussian probability distribution.

22. A method according to claim 1 wherein receiving initial values for the plurality of model variables comprises initializing the individual state values with initial estimates of a homeostatic state S and a circadian phase φ

of the subject, the homeostatic state S and the circadian phase φ

being parts of a two-process model which involves a homeostatic process that increases during periods of the subject being awake and decreases during periods of the subject being asleep and a oscillatory circadian process which varies with a period of approximately 24 hours.

23. A method according to claim 22 wherein initializing the individual state values with initial estimates of the homeostatic state S and the circadian phase φ

of the subject comprise initializing the individual state values with one or more variables which specify or estimate probability distributions of the homeostatic state S and the circadian phase φ

of the subject.

24. A method according to claim 22 wherein initializing the individual state values with initial estimates of the homeostatic state S and the circadian phase φ

of the subject is based on at least one of;

a light exposure history of the subject;

a history of administration of biologically active agents to the subject; and

a movement history of the subject.

25. A method according to claim 1 wherein receiving initial values for the plurality of model variables comprises initializing the individual state values based on at least one of:

a light exposure history of the subject;

a history of administration of biologically active agents to the subject; and

a movement history of the subject.

26. A method according to claim 3 wherein the one or more inputs which permit estimation of the updated values of the individual state variables are based on at least one of:

a light exposure history of the subject;

a history of administration of biologically active agents to the subject; and

a movement history of the subject.

27. A method according to claim 7 wherein receiving the one or more alertness inputs comprises one or more of:

measuring objective reaction-time tasks;

measuring cognitive tasks;

performing a Psychomotor Vigilance Task test;

performing a Digit Symbol Substitution test;

measuring subjective alertness based on questionnaires;

measuring subjective alertness based on a scale;

measuring subjective alertness based on a Stanford Sleepiness Scale;

measuring subjective alertness based on a Epworth Sleepiness Scale;

measuring subjective alertness based on a Karolinska Sleepiness Scale;

measuring electroencephalography (EEG) data from the subject;

performing a sleep-onset-test on the subject;

performing a Karolinska drowsiness test on the subject;

performing a Multiple Sleep Latency Test (MSLT) on the subject;

performing a Maintenance of Wakefullness Test (MWT) on the subject;

performing a blood pressure test on the subject;

performing a heart rate test on the subject;

performing a pupillography test on the subject;

performing an electrodermal activity test on the subject;

performing a hand-eye coordination performance test on the subject; and

performing a virtual task performance test on the subject.

28. A method according to claim 1 wherein the estimated alertness values of the subject comprise alertness values having metrics associated with one or more of:

objective reaction-time tasks;

cognitive tasks;

performing a Psychomotor Vigilance Task test;

a Digit Symbol Substitution test;

subjective alertness based on questionnaires;

subjective alertness based on a scale;

subjective alertness based on a Stanford Sleepiness Scale;

subjective alertness based on a Epworth Sleepiness Scale;

subjective alertness based on a Karolinska Sleepiness Scale;

electroencephalography (EEG) data from the subject;

a sleep-onset-test on the subject;

a Karolinska drowsiness test;

a Multiple Sleep Latency Test (MSLT);

a Maintenance of Wakefullness Test (MWT);

a blood pressure test;

a heart rate test;

a pupillography test;

an electrodermal activity test;

a hand-eye coordination performance test; and

a virtual task performance test.

29. A method according to claim 1 wherein the estimated alertness values of the subject comprise performance estimates relating to the subject'"'"'s performance of a specific task.

30. A method according to claim 1 wherein the model comprises a two-process model which involves a homeostatic process that increases during periods of the subject being awake and decreases during periods of the subject being asleep and a oscillatory circadian process which varies with a period of approximately 24 hours.

31. A method according to claim 30 wherein the model is cast as a dynamic state space model.

32. A method according to claim 3 wherein receiving the one or more inputs which permit estimation of the updated values of the individual state variables comprises receiving the one or more inputs over a communication network.

33. A method according to claim 7 wherein receiving the one or more alertness inputs comprises obtaining the one or more measured values of the alertness of the subject over a communication network.

34. A method according to claim 1 wherein estimating the alertness values of the subject comprises transmitting the alertness values of the subject over a communication network.

35. A non-transitory computer readable medium carrying instructions which when executed by a suitably configured processor cause the processor to perform the method of claim 1.

36. A method for estimating alertness of a human subject implemented by a processor, the method comprising:

receiving initial values for a plurality of model variables of a dynamic mathematical model, the plurality of model variables specifying or estimating probability distributions and the model including a process noise component comprising a probability distribution representing an uncertainty associated with the plurality of model variables;

receiving a sleep history input indicative of the subject'"'"'s asleep and awake status between a first time and a second time;

identifying, by the processor, one or more transition time points within the sleep history, each transition time point corresponding to one of;

the subject'"'"'s transition from awake to asleep status and the subject'"'"'s transition from asleep to awake status;

dividing, by the processor, a time between the first time and the second time into a plurality of time segments, each time segment extending between;

a corresponding time segment start time which is one of;

the first time and one of the one or more transition time points;

a corresponding time segment end time which is one of;

the one or more transition time points and the second time;

each time segment also associated with a sleep status value as indicated by the sleep history;

initializing, by the processor, the model variables at the first time to be the received initial values;

for each time segment, starting at a first time segment whose time segment start time corresponds to the first time through to a last time segment whose time segment end time corresponds to the second time;

using the model, by the processor, to estimate values of the model variables at the time segment end time, using the model comprising;

configuring the model based at least in part on the sleep status value of the time segment, and

basing the estimated values of the model variables at the time segment end time on the values of the model variables at the time segment start time, the process noise component, and a duration of the time segment; and

for time segments other than the last time segment, setting, by the processor, the estimated values of the model variables at the time segment end time to be the values for the model variables at the time segment start time of the next time segment; and

estimating, by the processor, alertness values of the subject at the second time based at least in part on applying the model to the values of the model variables at the second time.

37. A method according to claim 36 wherein the first time corresponds to a current time, the second time corresponds to a future time and the sleep history input is representative of a simulated future sleep history for the subject.

38. A method according to claim 36 comprising receiving one or more alertness inputs, each alertness input comprising a measured value of the alertness of the subject and a corresponding measurement time;

wherein identifying, by the processor, one or more transition time points, further comprises for each alertness input identifying a transition time point at the corresponding measurement time if a transition point is not already identified at that time; and

wherein using the model to estimate values of the model variables at the time segment end time comprises determining if an alertness input exists for the time segment end time and, if it does, updating at least one of the model variables using the alertness input.

39. A method according to claim 38 wherein the model comprises a measurement noise component, the measurement noise component comprising a probability distribution representing an uncertainty associated with the one or more measured alertness values.

40. A method for predicting alertness of a human subject implemented by a processor, the method comprising:

receiving initial values for a plurality of model variables of a mathematical model, one or more of the model values comprising variables which specify or estimate probability distributions;

receiving a sleep history input indicative of the subject'"'"'s asleep and awake status between a first time and a second time;

identifying, by the processor, one or more transition time points within the sleep history, each transition time point corresponding to one of;

the subject'"'"'s transition from awake to asleep status and the subject'"'"'s transition from asleep to awake status;

dividing, by the processor, a time between the first time and the second time into a plurality of time segments, each time segment extending between;

a corresponding time segment start time which is one of;

the first time and one of the one or more transition time points;

a corresponding time segment end time which is one of;

the one or more transition time points and the second time;

each time segment also associated with a sleep status value as indicated by the sleep history;

initializing, by the processor, the model variables at the first time to be the received initial values;

for each time segment, starting at a first time segment whose time segment start time corresponds to the first time through to a last time segment whose time segment end time corresponds to the second time;

using the model, by the processor, to estimate values of the model variables at the time segment end time, using the model comprising;

configuring the model based at least in part on the sleep status value of the time segment, and

basing the estimated values of the model variables at the time segment end time on the values of the model variables at the time segment start time and a duration of the time segment;

for time segments other than the last time segment, setting, by the processor, the estimated values of the model variables at the time segment end time to be the values for the model variables at the time segment start time of a next time segment; and

estimating, by the processor, alertness values of the subject at the second time based at least in part on applying the model to the values of the model variables at the second time.