User behavior recommendations for improving sleep

Abstract

A method is provided of generating behavior recommendations for a user, and communicating these to a user by means of a linguistic message, and wherein the recommended behavior or properties of the linguistic message are configured based on a measure of circadian inconsistency for the user. The measure of circadian inconsistency is derived by comparing an expected circadian curve of the user (e.g. an average curve derived from historical data for the user) with an empirical circadian curve for a given day. A deviation between the two provides an indication of the circadian inconsistency for the given day, and this is used to inform content, timing, wording, or other properties of the behavior recommendations.

Description

CROSS-REFERENCE TO PRIOR APPLICATIONS

This application claims the benefit of U.S. Provisional Application No. 62/990,110, filed on 16 Mar. 2020. This application is hereby incorporated by reference herein.

FIELD OF THE INVENTION

The present invention relates to a method for configuring behavior recommendations to a user for improving sleep.

BACKGROUND OF THE INVENTION

It is known that sleep can be improved through adoption of certain behaviors for example related to type and frequency of exercise, dietary habits, meditation, light exposure, sleep schedule, bedroom environment and others. More generally, behaviors that promote consistent, uninterrupted sleep can be referred to as sleep hygiene behaviors.

There is an association between level of adherence to healthy sleep behaviors and sleep improvements. There is evidence that that even modest improvements in adherence could result in improved outcomes. See for example the paper: E. E. Matthews, J. T. Arnedt, M. S. McCarthy, L. J. Cuddihy, and M. S. Aloia, “Adherence to cognitive behavioral therapy for insomnia: A systematic review,” Sleep Medicine Reviews, vol. 17, no. 6. W.B. Saunders, pp. 453-464, 1 Dec. 2013.

A common approach in behavior coaching platforms is to provide generic recommendations (one size fits all) or recommendations which are only partially personalized (e.g. based on demographic factors and/or based on past adherence patterns).

An improved approach to providing behavior-change recommendations would be of value, in particular one which is better personalized, and one which can improve adherence in a personalized way.

SUMMARY OF THE INVENTION

Adherence to sleep related habits is dependent on multiple internal and external factors. External factors include for example the environment, unforeseen distractions and events, and changing schedules. Internal factors include mood, perception to the habit change, and circadian rhythm. Embodiments of the present invention are based on customizing properties of behavior change recommendations based on the user's circadian rhythm.

The invention is defined by the claims.

According to examples in accordance with an aspect of the invention, there is provided a computer-implemented method comprising:

determining a predicted or expected circadian pattern for a user, based on a reference dataset comprising historical data for the user;

determining an actual (empirical) circadian pattern for day d based on patient sensor data gathered for a given day d;

determining a measure of circadian inconsistency for day d based on a deviation between the expected and actual circadian patterns for day d;

determining a linguistic message for delivery to the user using a sensory output means, the linguistic message comprising a recommended user behavior for improving sleep, wherein the recommended user behavior and/or one or more properties of the linguistic message are configured based on the circadian inconsistency; and

generating a control signal for controlling a sensory output means to generate a sensory output representative of the linguistic message.

Embodiments are based on customizing properties of behavior recommendations based on a circadian inconsistency of the user.

The expected circadian pattern may be an expected pattern for day d.

Circadian inconsistencies (e.g. misalignment or irregularity) are known to adversely affect cognitive performance and the ability to adhere to behavior recommendations (see e.g. S. L. Chellappa, C. J. Morris, and F. A. J. L. Scheer, “Daily circadian misalignment impairs human cognitive performance task-dependently,” Sci. Rep., vol. 8, no. 1, pp. 1-11, December 2018., and see also K. P. Wright, C. A. Lowry, and M. K. LeBourgeois, “Circadian and wakefulness-sleep modulation of cognition in humans,” Front. Mol. Neurosci., vol. 5, no. April, pp. 1-12, 2012).

Circadian inconsistency can be measured based on deviation between an expected (e.g. average or baseline) circadian pattern or curve for a user and an actual circadian pattern or curve.

Embodiments of the present invention proposes to evaluate circadian inconsistency, and to modify the content or delivery of behavior change recommendations based on the circadian inconsistency, to improve adherence and to better personalize the recommendations to the user. This improves outcomes for the user in terms of improved sleep.

By way of one example, some embodiments may customize the timing at which a recommendation is provided, and/or customize the recommended time for the user to enact the recommended behavior. Finding the ideal time to engage with the user and suggesting the ideal time to enact the recommended behavior, not only improves adherence but also increases likelihood of positive outcomes.

The expected circadian pattern for the user for day d may be an average circadian pattern for the user, determined based on historical user circadian pattern data, and/or based on historical monitoring data, e.g. activity and/or sleep monitoring data.

The expected circadian pattern may represent a baseline circadian pattern for the user, derived from historical information related to circadian pattern. This may comprise records of circadian pattern for the user over one or more days preceding day d, and/or based on historical patient sensor data over multiple 24 hour data blocks over one or more days preceding day d.

The actual circadian pattern may be determined in real time based on real time sensor data. The circadian pattern may be a curve with values which vary as a function of time, and the value of the circadian pattern at each moment in time may be determined in real time. The measure of circadian inconsistency may be determined in real time based on the real-time determined circadian pattern for day d and based on the expected circadian pattern for day d. For example, the actual circadian curve may be determined through continuous analysis (e.g. integration) of available sensor data. For example, the input sensor data may be analyzed at the smallest available time resolution and each computed circadian curve value compared in real time with the value on the expected (baseline) circadian curve from the one or more previous days.

The day d for which the actual circadian pattern is determined may be a continuously moving (e.g. 24 hour) time-window, so that at each instantaneous point in time, the relevant day d corresponds to the immediately preceding 24 hours. The benefit of this is that the complete 24 hour circadian pattern, both the expected pattern and actual pattern, for day d is always available. However this is not essential, the actual circadian pattern and circadian inconsistency may in some examples be derived in real time through the course of a given day d. In further examples, the day d may correspond to a completed 24 hour period up to a pre-determined prior time, for instance a wake time of a user.

The method may comprise determining a message delivery or provision time, this corresponding to a time at which the linguistic message is to be communicated to the user by means of the sensory output means, and wherein the message delivery time is determined based at least in part on the measure of circadian inconsistency.

By way of example, the method may be configured so as to avoid delivering the recommendation during time periods in which the circadian inconsistency of the user is high. Evidence shows that adherence may tend to be lower during such times.

In some examples, the measure of circadian inconsistency may comprise a measure of said deviation between the expected and actual circadian patterns as a function of time over day d. In other words, it comprises a time series of values of the deviation between the two circadian patterns spanning day d, each indicative of deviation between the expected and actual circadian curves at a certain time in day d. In other words, it comprises a pattern or function of circadian inconsistency or deviation as a function of time over day d.

In further examples, the measure of circadian inconsistency may comprise a measure of said deviation between the expected and actual circadian patterns at a specific time point in day d. Where day d is a moving time window corresponding to the immediately preceding 24 hours, this specific time point may be the time point at the (terminal) end of day d.

In some examples, the measure of circadian inconsistency may be determined in real time as mentioned above, and may comprise a continuously or recurrently varying instantaneous value of circadian inconsistency, representative of an instantaneous deviation between the expected and actual circadian curves at the instantaneous time point.

The delivery time of the message may be selected to be at a time of day in which the circadian inconsistency for day d is below a defined threshold. It has been found that at times of high inconsistency or deviation between expected and actual circadian patterns, a subject is less responsive to recommendations, and thus effect on improving sleep is lower.

The threshold may be a pre-defined threshold, or it may be set based on values of the measure of circadian inconsistency. In other words, it may be an absolute threshold, or a relative threshold, normalized to the values of the circadian inconsistency over day d.

As mentioned above, the actual circadian pattern may be determined in real time based on real-time sensor data and the comparison with the expected circadian pattern performed in real time. In this way, the circadian inconsistency can be determined in real time, so that a period of high or low inconsistency can be identified in real time.

In some examples, the recommended behavior may include a recommended commencement time by the user of the behavior, and wherein the recommended commencement time is determined based at least in part on the determined measure of circadian inconsistency.

By way of example, a user may be advised to perform a certain action or behavior at a time of high circadian inconsistency, so as to adjust the circadian pattern at that time to better conform to the expected (e.g. average) circadian pattern.

The reference dataset referred to above (based upon which the expected circadian pattern is determined) may comprise historical sleep and/or activity data for the user. It preferably comprises historical sleep and/or activity data for a plurality of days prior to day d. This data may be determined based on acquired sensor data for the user over the plurality of days. This sensor data may include physiological parameter data and/or movement data.

In (less preferred) alternative cases, the reference data may comprise generic circadian pattern data (i.e. not personalized to the user).

The expected circadian pattern may be determined based on fitting parameters of a pre-determined circadian curve equation, the parameters being fitted based on the historical sleep and/or activity data for the patient over one or more days. It is preferably fitted based on historical data for a plurality of days.

A single expected circadian curve may be determined, based on the combined (e.g. averaged) data for the plurality of days, this representing an average circadian curve for the user, and this single curve used as the expected circadian pattern for day d. Alternatively, a separate circadian curve may be determined for each day of the week, if sufficient historical data exists, and the circadian curve corresponding to the same day of the week as day d is used as the expected circadian pattern for day d.

The sensor data (based upon which the actual circadian pattern is determined) may in some examples comprise physiological parameter data, and/or movement data for the patient.

A physiological parameter sensor signal or a movement sensor signal may in some examples be used as a proxy signal, wherein this proxy signal over day d is used as the actual circadian pattern (actual circadian curve). In other words, the actual circadian pattern for day d is determined using a proxy signal, wherein the proxy signal is based on the sensor data.

In some examples, the measure of circadian inconsistency may comprise an index of circadian inconsistency for day d. This may be determined based on deriving a correlation (e.g. a correlation coefficient) between the expected and actual circadian patterns. The index comprises for example a single value. The index may be assigned the value 1−r, where r corresponds to the correlation coefficient between the expected and actual circadian patterns over for example a 24 hour period.

The method may comprise monitoring the index of circadian inconsistency over a plurality of days and determining a measure of adherence based on changes in the index of circadian inconsistency.

Adherence may be assumed to be correlated with reduction in circadian inconsistency, i.e. the user is more likely to adhere to the recommended behavior changes when the circadian inconsistency falls. The recommended behavior and/or the properties of the message may in some examples be configured based on the measure of adherence.

The recommended behavior change (i.e. the content of the message) may in some examples be determined based on the circadian inconsistency, and based on use of a lookup table. The lookup table may contain pre-determined recommended behaviors associated with different levels of circadian inconsistency, and/or different patterns of circadian inconsistency as a function of time over a given day. The table can be queried based on the determined measure of circadian inconsistency.

In some examples, the method may comprise configuring one or more properties of the linguistic message based on the circadian inconsistency, and wherein the properties include a wording of the linguistic message. The wording is variable separately from the behavior change being communicated (i.e. separately from the content of the message).

For example, the optimal tone, mood and complexity of the recommendation message can for example be adjusted depending upon the level of circadian inconsistency.

The configuring of the wording of the message, for communicating a particular behavior change, may in some cases be performed by selecting from a lookup table one of a plurality of pre-determined linguistic messages conveying said behavior change and each with a different wording.

In other words, the look up table contains a plurality of curated messages, each corresponding to a (range of) values of the circadian inconsistency.

In further examples, the configuring of the wording of the message may be performed using a machine learning algorithm. This may be a machine learning algorithm in the domain of natural language generation.

Examples in accordance with a further aspect of the invention provide a computer program product comprising computer program code, the computer program code being configured, when executed on a processor, to cause the processor to perform a method in accordance with any example or embodiment outlined above or described below, or in accordance with any claim of this application.

Examples in accordance with a further aspect of the invention provide a processing arrangement comprising:

an input/output; and

one or more processors adapted to:

• • determine a predicted or expected circadian pattern for a user based on a reference dataset for the user accessed using the input/output;
  • determine an actual or empirical circadian pattern for a given day d based on user sensor data gathered for day d, the sensor data received at the input/output;
  • determine a measure of circadian inconsistency for day d based on a deviation between the expected circadian pattern and actual circadian pattern for day d;
  • determine a linguistic message for delivery to the user using a sensory output means, the message comprising a recommended user behavior, and wherein the recommended behavior and/or one or more properties of the message are configured based on the circadian inconsistency; and
  • generate a control signal for controlling a sensory output means to generate a sensory output representative of the linguistic message.

Examples in accordance with a further aspect of the invention provide a system comprising: a sensory output means for generating a sensory output for conveying a linguistic message to a user; and a processing arrangement in accordance with any example or embodiment outlined above or described below, or in accordance with any claim of this application.

These and other aspects of the invention will be apparent from and elucidated with reference to the embodiment(s) described hereinafter.

BRIEF DESCRIPTION OF THE DRAWINGS

For a better understanding of the invention, and to show more clearly how it may be carried into effect, reference will now be made, by way of example only, to the accompanying drawings, in which:

FIG. 1 outlines steps of an example method in accordance with one or more embodiments;

FIG. 2 shows a processing workflow according to one or more embodiments;

FIG. 3 shows an example of an expected circadian pattern, derived using a model circadian curve equation;

FIG. 4 shows examples of empirical circadian curves, derived using user sensor signals as proxy signals of circadian pattern;

FIG. 5 shows a set of example empirical circadian curves;

FIG. 6 illustrates one example means of deriving an index of circadian inconsistency;

FIG. 7 illustrates determining an example period of circadian inconsistency; and

FIG. 8 illustrates example circadian resiliency levels.

DETAILED DESCRIPTION OF THE EMBODIMENTS

The invention will be described with reference to the Figures.

It should be understood that the detailed description and specific examples, while indicating exemplary embodiments of the apparatus, systems and methods, are intended for purposes of illustration only and are not intended to limit the scope of the invention. These and other features, aspects, and advantages of the apparatus, systems and methods of the present invention will become better understood from the following description, appended claims, and accompanying drawings. It should be understood that the Figures are merely schematic and are not drawn to scale. It should also be understood that the same reference numerals are used throughout the Figures to indicate the same or similar parts.

The invention provides a method of generating behavior recommendations for a user, and communicating these to a user by means of a linguistic message, and wherein the recommended behavior change or properties of the linguistic message are configured based on a measure of circadian inconsistency for the user. The measure of circadian inconsistency is derived by comparing an expected circadian curve of the user (e.g. an average curve derived from historical or prior data for the user) with an empirical circadian curve for a given day. A deviation between the two provides an indication of the circadian inconsistency for the given day, and this is used to inform content, timing, wording, or other properties of the behavior recommendations.

Circadian rhythm determines sleep patterns, so a behavior change recommendation that aims at improving sleep may benefit from taking into account the user's circadian rhythm.

Embodiments of this invention are based in particular on adjusting (the content, wording and/or the timing of) a behavior change recommendation based on circadian inconsistency of the user. Low circadian alignment (i.e. low circadian consistency) can affect adherence to healthy sleep behaviors.

The circadian inconsistency may in some examples be continuously or recurrently derived in real-time based on real-time sensor data and reference circadian pattern data for the user, and the linguistic message also determined in real time.

FIG. 1 outlines steps of an example computer-implemented method 10 according to one or more embodiments. The method comprises determining 12 an expected circadian pattern for a user based on a reference dataset. The method further comprises determining 14 an actual circadian pattern for a given day d based on user sensor data gathered for day d. The method further comprises determining 16 a measure of circadian inconsistency for day d based on a deviation between the expected and actual circadian patterns for day d. The method further comprises determining 18 a linguistic message for delivery to the user using a sensory output means, the linguistic message comprising a recommended user behavior for improving sleep. One or both of the recommended user behavior 20 and/or properties of the linguistic message 22 are configured based on the circadian inconsistency. The method further comprises generating 24 a control signal for controlling a sensory output means to generate a sensory output representative of the linguistic message.

The method may be implemented by a processing arrangement comprising one or more processors adapted to perform the steps as set out above, and an input/output for accessing the reference data and receiving the patient monitoring data. A further aspect of the invention provides such a processing arrangement. Another aspect of the invention may provide a system comprising a processing arrangement and a sensory output device adapted to generate a sensory output for conveying the generated linguistic message to the user. This may for example comprise a display unit to generate a visual presentation of the linguistic message, or an acoustic output device, such as a speaker, to provide an audible acoustic presentation of the linguistic message.

FIG. 2 presents a schematic representation of an example processing workflow according to one or more embodiments, elements of which will be discussed below. These processing steps may be performed by a single processing component or collectively by a plurality of processing components, one or more of which may be comprised by a system provided by the invention.

In order to facilitate the determining of the actual (empirical) circadian cycle, sensor data 42 may be acquired relating to activity and/or sleep of the user. The user may carry one or more sensors on their body as they go about their daily life, and the sensors collect data as a function of time over multiple days. The data may be collected continuously or at regular intervals 24 hours a day. The one or more sensors may include one or more physiological parameter sensors, such as a PPG sensor (which can measure for example heart rate and blood oxygen saturation), a temperature sensor, a respiratory sensor, an ECG sensor, a blood pressure sensor. The one or more sensors may additionally or alternatively include activity sensors such as an accelerometer and/or gyroscope. The sensors may be operatively coupled to a local controller, for example a controller carried by the user, which controller co-ordinates data collection and sensor settings. The local controller might be facilitated by a mobile computing device, such as a smartphone. The real-time acquired sensor data may be locally cached or stored in a local memory, for example a memory of the local controller. Alternatively, it may be directly communicated to a remote datastore, for instance a cloud-based datastore.

Using the sensor data collected for a given day, d, an actual, empirical circadian cycle 48 can be determined by a processing arrangement according to an embodiment of the invention. In some particular examples, the day d may be a moving 24 hour time window. It may be a continuously moving 24 time window, so that it corresponds at all times to a 24 time window immediately preceding the current time. This will be described in more detail to follow.

The acquired sensor data can also be further processed to derive sleep and/or activity data 44 for the user. The sleep data can be derived from the sensor data acquired during the night time (or otherwise during sleep periods of the user) and the activity data derived from the sensor data acquired during the day time (or otherwise during the awake periods of the user). For example, patterns of movement data over the night time and/or heart rate can be used to track sleep phases, e.g. deep sleep and REM sleep phases. Movement data and heart rate data over the day time can be used to track exercise patterns of the user for example. The derived activity and/or sleep data can stored in a user-specific dataset 46 which over time builds to form a dataset of historical sleep and/or activity data for the user. This historical dataset can be stored in a datastore which is comprised as part of the system of the invention, or it may be stored remotely, for instance in a cloud-based data store.

The historical activity and/or sleep data 46 may be used in some embodiments to derive the expected circadian pattern 50 for the user. This will be described in greater detail to follow.

Based on comparing the expected and actual circadian patterns, a circadian inconsistency 52 can be determined. This can then be fed to a behavior recommendation module 54 which comprises one or more algorithms configured to generate a recommended behavior change 56 for the user based on the circadian inconsistency and optionally further based on one or more additional inputs such as properties of the actual circadian cycle itself or the sensor data for the user for the day d. Properties of the linguistic message used to communicate the recommendation are also configured by the behavior recommendation module, based at least in part on the circadian inconsistency 52.

An example approach to determining the expected circadian pattern will now be outlined.

In this example, the expected circadian pattern comprises a circadian phase as a function of time (circadian curve), and is determined based on use of historical sensor data 42 for the user or based on historical sleep and/or activity data 46 for the user. The historical data may correspond to data acquired for the user over at least one day, and more preferably over a plurality of days, for example at least five days. The historical sensor data can be used to fit a standard circadian curve equation, and the resulting circadian curve equation can be used as an expected circadian curve for the user. It effectively represents a baseline, or average, circadian curve for the user, based on their historical activity, sleep, and/or other sensor parameter data over the historical data period.

In more detail, the expected (baseline) circadian pattern for the user can be derived using the so-called two-process model for sleep/wake regulation. Details of this model can be found for example in the paper: P. Ackermann and A. A. Borbély, “Mathematical models of sleep regulation.,” Front. Biosci. a J. virtual Libr., vol. 8, no. 13, pp. s683-s693, 2003.

According to this model, sleep and wake periods alternate throughout a 24-hour cycle and are regulated by two major factors, which can be represented by a homeostasis, S, function and a circadian, C, function. The homeostasis, S, function effectively represents ‘sleep need’, and thus declines during sleep, and rises during wakeful periods. The circadian, C, function represents the changing circadian phase, and can represent any of the various biological processes which follow a cyclical oscillatory pattern with a cycle period of approximately 24 hours.

The homeostatic process S(t) may be defined for each time t by two exponential functions: one for wakefulness (Equation 1) and one for sleep (Equation 2 Equation):

$\begin{array}{cc}\frac{dS\left(t\right)}{dt}=\left(U-S\left(t\right)\right)/{\tau }_w& \mathrm{Equation}1\\ \frac{dS\left(t\right)}{dt}=-\left(S\left(t\right)-L\right)/{\tau }_s& \mathrm{Equation}2\end{array}$

where U represents an upper asymptote, L represents a lower asymptote, τ_w denotes the time constant of the increasing saturating exponential function during wakefulness, and τ_s denotes the time constant of the decreasing exponential function during sleep. In the classical two-process model, both U and L are constant for all t. The circadian process C(t) at time t may be modeled by a five-harmonic sinusoidal equation:

$\begin{array}{cc}C\left(t\right)=\sum _{i=1}^5{c}_i\sin \left[i\frac{2\pi }{24}\left(t+\phi \right)\right]& \mathrm{Equation}3\end{array}$

where c_i, i=1, . . . , 5, represent the amplitude of the five harmonics of the circadian oscillator and φ denotes the circadian phase. The amplitudes c_i are determined from the historical dataset for the patient. By way of one non-limiting illustrative example, in one computation of the circadian process, C(t), the amplitude values were computed as c₁=0.97, c₂=0.22, c₃=0.07, c₄=0.03, and c₅=0.001.

In the simplest approach, the five parameters U, L, τ_w, τ_s, and φ are estimated using the historical sleep and activity data for the user. In other words these parameters of the set of equations defining the model are fitted based on historical sleep and activity data for the user. Since there are five parameters, the sleep and activity data for at least five days should be used. In practice, the data from at least 10 days is more preferably used, to thereby solve an overdetermined system of equations to estimate U, L, τ_w, τ_s, and φ.

In further embodiments, sleep EEG data may be used to estimate the homeostatic process S(t) according to the following model equation:

$\begin{array}{cc}\frac{dS\left(t\right)}{dt}=-\gamma \mathrm{SWA}\left(t\right)& \mathrm{Equation}4\end{array}$

where SWA(t) (EEG power in the 0.5 to 4 Hz) is the slow-wave activity during NREM sleep. The speed of homeostatic sleep dissipation is directly proportional to slow-wave activity. The estimation of the parameter y is described in the paper: G. Garcia-Molina et al., “Automatic characterization of sleep need dissipation using a single hidden layer neural network,” in 2017 25th European Signal Processing Conference (EUSIPCO), 2017, pp. 1305-1308.

The circadian pattern itself may be defined solely by the parameter C(t). As discussed above, C represents the circadian curve model, C_i (i=1, . . . , 5) corresponds to the circadian amplitude and φ corresponds to the circadian phase. The expected (i.e. baseline) circadian curve can be built based on fitting the model equation to data from multiple 24 hour data blocks for the user, e.g. an average of multiple 24 hour data blocks.

The resulting set of fitted equations can be used to derive the user's average circadian phase and amplitude over a 24 hour period. This average is taken as a reference and can be used as the baseline or expected circadian rhythm (t) for a day.

An example of an average circadian curve for the user, to be used as the expected circadian pattern for day d, is shown in FIG. 3.

As an alternative to the above, the expected or predicted circadian curve for day d can be derived from a generic circadian curve stored in a datastore.

The method further comprises determining an actual (empirical) circadian pattern of the subject. This can be derived using sensor data 42 acquired for the subject for day d, or based on activity and/or sleep data for the subject for day d, the activity and/or sleep data derived using the sensor data.

The procedure to estimate the empirical circadian pattern depends upon the type of empirical signal (e.g. physiological or behavioral) that is used to derive it.

Without loss of generality, in one preferred embodiment, one or more cardiac signals are recorded. The cardiac signals may be facilitated by one or more photoplethysmography (PPG) sensor signals. Using a PPG sensor signal, correlates of heart rate variability (HRV) can be obtained.

HRV is a measure of the amount of variability in time intervals between adjacent heartbeats (RR intervals) within a pre-defined temporal window. The window duration may for example be at least one-minute. Typically, HRV calculations are performed on NN intervals, meaning RR intervals from which artifacts have been removed. Identification of RR and NN intervals is a routine procedure in the art and the skilled person will immediately recognize how it is performed.

A popular time-domain metric to estimate HRV is the standard deviation of the NN-interval duration, referred to as SDNN. More details on the derivation and use of SDNN can be found in the paper: F. Shaffer and J. P. Ginsberg, “An Overview of Heart Rate Variability Metrics and Norms,” Front. Public Heal., vol. 5, no. September, pp. 1-17, 2017.

SDNN values can be estimated for 24-hour periods to estimate the actual (empirical) circadian curve for said 24 hour periods. This is shown schematically in FIG. 4 in which SDNN values (normalized with respect to the day-average) for four different days are plotted as a function of time of the day (in 24-hour format). The plotted circadian curves correspond to two pairs of consecutive days, a first 62 and second 64 weekday and a first 66 and second 68 weekend day. In the particular illustrated example, the circadian curve for day two 68 during the weekend shows a slight misalignment (inconsistency) compared with the circadian curves for the other days. This type of deviation from the average, can be detected using methods outlined in this disclosure.

In some embodiments, the actual or empirical circadian curve can be derived using other types of signals. For instance, FIG. 5 illustrates example patient sensor signals as a function of time-of-day over a 24-hour period. In order, from top to bottom, the illustrated signals correspond to light exposure, skin temperature, and activity (tracked using an accelerometer). In some examples, at least one of these sensor signals may be taken directly as the actual circadian pattern of the user. In other words, the sensor signal can be used as a proxy signal for the circadian curve of the user. In some embodiments, multiple such sensor signals may be combined, e.g. averaged, and this combination or average used as the actual (empirical) circadian curve for a given day.

As mentioned above, in some examples, the day d for which the actual circadian pattern is derived may correspond to the 24 hour period immediately preceding a current time point. In other words, it may be a continuously moving 24 hour time window, so that it corresponds at all times to a 24 time window immediately preceding the current time. This represents only one option however. In further examples, the day d may be the last full day (midnight to midnight) preceding the present day.

Based on the derived baseline or expected circadian pattern for day d and based on the actual circadian pattern for a given day d, a measure of circadian inconsistency is derived. The measure of circadian inconsistency is based on a deviation between the expected and actual circadian patterns for day d. A deviation between the expected and actual circadian patterns may mean a difference between the expected and actual circadian curves as a function of time (i.e. one circadian curve or signal subtracted from the other), or the difference between the curves at a specific one or more time points of day d.

In some examples, the measure of circadian inconsistency comprises a measure of said deviation between the expected and actual circadian patterns as a function of time over day d. It is thus an inconsistency curve or function or signal, as a function of time.

In accordance with one or more further embodiments, the method may comprise deriving a single-value index of circadian inconsistency. This may be determined based on deriving a correlation coefficient between the expected and actual circadian patterns. It may correspond to a difference between the expected and actual curves for a specific time point. In either case, this index may in some example be used as the measure of circadian inconsistency.

In one simple example, the index of circadian inconsistency I_c may be determined as I_c=1−r, where r is a correlation coefficient between the expected and actual circadian curves.

In a further simple example, the index of circadian inconsistency I_c may be determined as the RMS of the difference between the expected and actual circadian curves for day d, i.e. (t)−C_d(t) where (t) is the expected circadian curve and C_d(t) is the actual circadian curve for day d.

This is illustrated schematically in FIG. 6 which shows an example expected circadian pattern (curve) 74 and an example actual (empirical) circadian pattern 72. The expected curve 74 may for example be obtained from the two-process model described above which, for this illustrated case resulted in a phase parameter, ϕ=6 hours. The empirical circadian curve 74 is in this case derived from an activity sensor signal for the user (measured with an accelerometer) for the day d.

A correlation coefficient, r, between the two is derived 76. In order to account for the difference in units between the actual 72 and expected 74 curves, the former may be normalized as fraction of the day-average. In other words, all values of the sensor signal used for the empirical circadian curve for day d are divided by the average signal value for day d. In the illustrated example, a correlation coefficient of −0.86 is determined.

In a simple embodiment, an index of circadian inconsistency, I_c, may be derived as equal to 1 minus the absolute value of the correlation, i.e. I_c=1−|r|. In the example illustrated in FIG. 6, circadian inconsistency is low, with a value of 0.14.

Once a measure of circadian inconsistency is determined, the method comprises determining a linguistic message for delivery to the user using a sensory output means, the linguistic message comprising a recommended user behavior for improving sleep, and wherein the recommended user behavior and/or one or more properties of the linguistic message are configured based on the circadian inconsistency.

Thus, two primary dimensions of the behavior recommendation that are proposed to be modified are: the activity or behavior change being recommended; and/or properties of the delivery of the sleep habit recommendation, i.e. properties of the message and its delivery to the user

The modification of the behavior change recommendation may be based on one or both of two aspects of the circadian inconsistency: value(s) of the circadian inconsistency index and/or timing of periods of circadian inconsistency above or below a certain threshold.

There will now be outlined a number of options for configuring the recommended behavior change and/or properties of the message.

First are outlined a number of options for configuring the recommended behavior change.

In a simplest example, the recommended behavior change is selected from a look-up table based on the detected circadian inconsistency and optionally one or more further parameters.

The lookup table may be populated in advance by a relevant clinical expert. The different recommendations may be associated in the table with different relevant factors. These may include, but are not limited to, a sleep abnormality being experienced by a user, properties related to sleep hygiene of a user, and other factors which may be determined using a questionnaire filled out by a user. Therefore, the specific recommendation may be selected based on the user's feedback (e.g. response to the questionnaire) and/or based on sensor data for the user. Hence, the recommendation may not be based on the measure or circadian inconsistency alone.

By way of illustration, if the user is experiencing difficulty falling asleep, some example behavior recommendations which may be generated are as follows:

• • Going to bed and waking up at the same time every day. This helps to shift the circadian rhythm to be in synchrony with the user's bedtime.
  • Avoiding stimulants such as caffeine, nicotine, strenuous exercise etc. before bedtime. This makes it easier to wind down and fall asleep.
  • Limiting light exposure and screen time before bedtime. This makes it easier to wind down and fall asleep.
  • Increasing light exposure upon waking—makes the user feel fresher and shifts circadian rhythm to the desired timing.
  • Performing wind-down activities such as meditation, paced breathing, journaling. These can be performed if there is circadian inconsistency close to the user's bedtime, meaning that the user will have difficulty falling asleep.

In addition to the action or behavior to be adopted, the recommendation may further comprise a recommended commencement time by the user of the behavior, and wherein the recommended commencement time is determined based at least in part on the determined measure of circadian inconsistency.

An optimal time for performing a recommended action varies depending upon the recommendation. For example, if the recommendation is to practice meditation (to wind-down and relax), a recommended commencement time may be a time of higher circadian inconsistency and closer to the user's desired bedtime (as configured in advance by the user), so that the meditation would help the user wind-down to sleep.

By way of a further example, if the recommendation is to exercise 4-6 hours before bedtime (to regulate body temperature), since this is a more engaging activity for the user, a recommended commencement time may be a time of lower calculated circadian inconsistency, while remaining within the range of 4-6 hours before the user's desired bedtime. A time period of 4-6 hours is the time required for the body temperature to lower after exercise, thus this timing has the effect that the user feels sleepy around the desired bedtime.

As discussed above, the circadian inconsistency may at least in some embodiments be determined in real time. Thus, at any given moment in time, a circadian inconsistency for that time point may be determined. As a result, a recommended commencement time for a behavior can also be determined in real time based on the real-time circadian inconsistency. For instance, taking the example presented above in which the recommended commencement time is a time of lower calculated circadian inconsistency, while remaining within the range of 4-6 hours before the user's desired bedtime, the circadian inconsistency can be monitored in real time within the time window of 4-6 hours before the bedtime, and when circadian inconsistency falls below a threshold, the behavior is recommended to be started at that time.

More generally, determining a recommended commencement time for a behavior may be based on use of a lookup table populated in advance by a clinician. The lookup table may list different possible commencement times for different recommended behaviors, each associated with a different measured circadian inconsistency level at a certain time, or different circadian inconsistency patterns. By querying the table with the measured circadian inconsistency and the behavior being recommended, a recommended time for commencement may be obtained.

Thus, circadian inconsistency can be considered as one field of the aforementioned lookup table. For instance, if there is sufficient historic data available for the patient, the system may determine an average circadian inconsistency for the user over a certain time period (e.g. the period of day d), and the recommended commencement time may be set to a time when the actual circadian inconsistency is within a pre-determined threshold of the average value (as long as it's also in the desired window of recommendation). In further examples, the recommended commencement time may be set to a time when the measured circadian inconsistency is less than a pre-determined threshold value. Through recommending the action be performed at the right time, it can increase the likelihood that the user performs the action and that it is targeted most directly at a time where the circadian cycle is in need to adjustment.

Additionally or alternatively to configuring the content of the linguistic message (i.e. the recommended behavior) the method may comprise configuring properties of the linguistic message, i.e. properties of the delivery of the message.

In accordance with one or more embodiments, the method comprises configuring a wording of the linguistic message based at least in part on the determined circadian inconsistency. For example, the recommendation communicated to the user may be modified to have the optimal tone of voice, mood and complexity.

In a simple embodiment, this may be done based on use of a lookup table.

For example, the configuring of the wording of the linguistic message, for communicating a particular behavior change, may be performed by selecting from a lookup table one of a plurality of pre-determined linguistic messages conveying said behavior change and each with a different wording. The message wordings in the lookup table may be generated in advance, for instance manually (i.e. human-generated).

A selected version of the recommendation text is selected from the table of curated messages based on the value(s) of the circadian inconsistency.

In accordance with an alternative set of examples, the configuring of the wording of the message may be performed using a machine learning algorithm, and based on the determined circadian inconsistency. In other words, the wording of the message is configured automatically using machine-learning-based natural language generation.

This preferred embodiment aims at modifying the wording of the recommendation provided for changing behavior, based on the calculated circadian inconsistency.

Machine learning architectures in the domain of natural language generation are known in the art and can be utilized to paraphrase the wording of a recommendation based on various factors determined in advance. By way of example, the wording of the recommendation communicated to the user can be modified to have an optimal tone of voice, mood and complexity, depending upon the circadian inconsistency.

For example, reference is made to the following paper: L. Logeswaran, H. Lee, and S. Bengio, “Content preserving text generation with attribute controls,” in 32nd Conference on Neural Information Processing Systems, 2018, pp. 2-11. This outlines an approach to modifying textual attributes of sentences based on various attributes, such as mood. The optimal tone, mood and complexity of the recommendation message can be set for different ranges of circadian inconsistency values.

By way of brief summary, the building and application of a machine learning model for modifying the wording of the linguistic message according to circadian inconsistency may be performed as follows.

The machine learning model is built in advance. The computer implemented method may include an additional preliminary phase of building the model, or the method of the invention may make use of a model which has already been built.

The model is trained based on a training dataset comprising records of previously administered messages to different users, and the messages having defined sets of linguistic properties or attributes, and each previous message tagged with a measure of circadian inconsistency of the user to whom it was administered, and a measure of adherence of the user to the message (i.e. success of the message in changing the user's behavior). The measure of adherence may be binary, i.e. the user followed the recommendation or did not follow the recommendation. A machine learning algorithm may be trained to output a linguistic attribute set for a message which maximizes predicted adherence, based on an input indicative of the measure of circadian inconsistency, and preferably one or more further characteristics of the user, and preferably the behavior change being recommended.

For example, the message attribute set may comprise a vector of {tone, mood, complexity} and wherein each element of the vector is assigned a particular value (e.g. {soft, neutral, simple}). These values may be determined manually for each data element in the training dataset, and the data elements tagged accordingly.

Once the machine learning algorithm has been built, during execution of the computer implemented method 10, the model can be used to determine the linguistic properties of the message based on provision, as an input to the model, of at least the circadian inconsistency, and preferably one or more further characteristics of the user, and preferably the behavior change being recommended. The model may generate the linguistic message to be conveyed, with wording configured according to the identified linguistic attribute set. Alternatively, a different model, or a lookup table, may be used to determine the specific wording for the message, based on the recommended behavior change and the determined linguistic attribute set.

Suitable machine-learning algorithms for being employed in the present invention will be apparent to the skilled person. Examples of suitable machine-learning algorithms include decision tree algorithms and artificial neural networks. Other machine-learning algorithms such as logistic regression, support vector machines or Naïve Bayesian models are suitable alternatives.

The structure of an artificial neural network (or, simply, neural network) is inspired by the human brain. Neural networks are comprised of layers, each layer comprising a plurality of neurons. Each neuron comprises a mathematical operation. In particular, each neuron may comprise a different weighted combination of a single type of transformation (e.g. the same type of transformation, sigmoid etc. but with different weightings). In the process of processing input data, the mathematical operation of each neuron is performed on the input data to produce a numerical output, and the outputs of each layer in the neural network are fed into the next layer sequentially. The final layer provides the output.

Methods of training a machine-learning algorithm are well known. Typically, such methods comprise obtaining a training dataset, comprising training input data entries and corresponding training output data entries. An initialized machine-learning algorithm is applied to each input data entry to generate predicted output data entries. An error between the predicted output data entries and corresponding training output data entries is used to modify the machine-learning algorithm. This process can be repeated until the error converges, and the predicted output data entries are sufficiently similar (e.g. ±1%) to the training output data entries. This is commonly known as a supervised learning technique.

For example, where the machine-learning algorithm is formed from a neural network, (weightings of) the mathematical operation of each neuron may be modified until the error converges. Known methods of modifying a neural network include gradient descent, backpropagation algorithms and so on.

To make the machine learning algorithm more accurate in providing an optimum message wording for a given user, the prior dataset employed to train the model may be subjected to clustering according various attributes or characteristics of the users to whom each data element corresponds.

By way of explanation, one approach to deriving a machine learning algorithm using clustering will now be outlined. In this example, a prior dataset of user data for a population of users is utilized.

In at least one embodiment, the data is clustered according to user attributes. In this case, a first step is to create user clusters. In a preferred embodiment, two-levels of clustering may be performed. At a first level, a k-medians clustering algorithm can be applied to cluster users as S₁, S₂, . . . , Sn base clusters based on a set of non-modifiable attributes where S_i is the ith cluster and n is a pre-determined number of clusters. It is possible to determine n via prior knowledge of the user population. The non-modifiable attributes may include attributes such as age, gender, ethnicity and genetics. In some embodiments, at least a subset of the non-modifiable attributes may be automatically detected from the dataset. For example, some personality traits can be determined from sensory data, for example data indicative of patterns of eye movements (which can be measured using cameras or frontal electrodes). See for example the paper: Shlomo Berkovsky et al. 2019. Detecting Personality Traits Using Eye-Tracking Data. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems (CHI '19). The personality traits may be mapped along five dimensions according to the big-five framework (extraversion, agreeableness, conscientiousness, emotional stability, and openness to experience).

For the second level of clustering, a Gaussian Mixture Model (GMM) may be applied for further clustering within each base cluster, based on a set of modifiable attributes of the users of the population. The resulting clusters may be represented as C_i1, C_i2, . . . , C_im, where i is the ith base level cluster and m is the number of modifiable clusters within the ith base level clusters.

The modifiable user attributes may include for example one or more of: motivation, social support, emotional status, and health literacy. These attributes may be determined in advance for each user based on user responses to questionnaire questions. For example, in advance of the assessment of circadian inconsistency, the user may be provided the questionnaire, and their answers collected. The answers may then be converted into a set of data values within one or more quantitative metrics corresponding to different user attributes. For instance, in the case of motivation, the motivation may be represented by a metric comprising three quantitative dimensions: valence (which quantifies the positive or negative attitude of the user in a scale from 1 to 3, expectation (which quantifies how often the user is likely to follow recommendation), and instrumentality (which quantifies whether enacting the behavior depends on the user, the user and others, or uniquely on others). Reference is made for example to the paper: L. Mihailescu et al, “The Quantification of The Motivational Level of the Performance Athletes”, Procedia—Social and Behavioral Sciences, Volume 84, 2013, Pages 29-33.

The GMM model offers a number of advantages over the k-median algorithm such as mixed membership with associated probability, and non-circular clusters (i.e. elliptic shape). It therefore provides greater flexibility for the second level of clustering and is well suited to cases where there is less prior knowledge of user clustering over modifiable factors. With a GMM model, there will be a probability P_ij associated with each C_ij, where i is from 1 to n and j is from 1 to m. Therefore, a user in S_i cluster has a P_ij probability of belonging to cluster C_ij. The same user may also be assigned to other second level clusters with corresponding probabilities. If a decision is made to assign a user to a second level cluster with maximum probability, that user will have one specific second level cluster membership only.

In accordance with a variation on the above, for the second level of clustering (i.e. modifiable attributes), the users may be clustered using a Hierarchical Agglomerative Clustering (HAC) algorithm. The algorithm begins by assigning one user per cluster, before pairs of clusters are merged as the algorithm moves up the hierarchy. A clustering stop condition is selected based on prior knowledge of the number of clusters, as derived from the base level clustering results. An advantage of this algorithm is that there is no requirement to specify the number of clusters at the starting point, and the algorithm is not sensitive to the choice of distance metric.

Clusters may be defined which segment the space of non-modifiable user attributes into disjoint clusters. Within each cluster, sub-clusters are defined that segment the space of modifiable factors. This further sub-clustering may be based on factors derived from the user sensor (monitoring) data discussed above. These further factors may be pre-determined factors or parameters which may be each extracted from the user sensor data based on pre-determined equations or algorithms. Within the context of the present invention, useful parameters to extract are those related to sleep/wake regulation. These may include the circadian curve C(t) and hemostasis curve, S(t) discussed above. These can be derived using the known models such as the two-process model discussed above.

A next step comprises user preference prediction. Historic adherence data is extracted from the particular cluster (or sub-cluster) to which the current user belongs. In particular, properties are identified of linguistic messages previously provided to users in the relevant cluster and which are associated with high adherence by the user. High adherence means that the user followed the recommended behavior change. For example, properties of the linguistic messages, such as tone, mood and complexity, which led to successful execution by the user of the recommended behavior are identified. The different linguistic messages within the relevant data cluster may be further stratified according to the level of circadian inconsistency of the user when they were administered. This allows properties of messages to be identified which had high adherence when administered with a similar level of circadian inconsistency as detected for the current user. Since the data analyzed is data for users lying within the same cluster as the current user, this makes it likely that the same linguistic message properties will be successful in changing the behavior of the current user.

In this way, the optimal linguistic message properties, such as tone, mood and complexity, are identified based on circadian inconsistency.

Once prior messages are identified in the data cluster which had high adherence, a process can be performed for extracting the relevant linguistic properties of the message and applying them to the behavior recommendation message to be delivered to the current user.

One approach to doing this is with a so-called ‘style transfer technique’, described in the paper: L. Logeswaran, H. Lee, and S. Bengio, “Content preserving text generation with attribute controls,” in 32nd Conference on Neural Information Processing Systems. The steps of this technique may be briefly summarized as follows:

A message which was successful in changing the user behavior is identified as discussed above. One or more linguistic properties of the message are detected (i.e. sentiment or style). The message may be pre-annotated with these properties, or the properties can be determined automatically, e.g. using a machine learning algorithm.

The style transfer technique described in detail in L. Logeswaran, H. Lee, and S. Bengio, “Content preserving text generation with attribute controls,” in 32nd Conference on Neural Information Processing Systems,) is then applied to modify the linguistic properties (sentiment/style) of the message to be communicated to the current user.

By way of further illustration, there will now be outlined some examples of modifying the linguistic properties of the linguistic message based on the circadian inconsistency.

A general example of changing the wording of a recommendation message based on circadian inconsistency index is presented in Table 1 below:

TABLE 1
Circadian	Message Attributes Set
inconsistency ranges	Tone	Mood	Complexity
0-0.25	casual	calm	complex
0.26-0.65	intense	hopeful	average
0.66-1.0	assertive	reflective	simple

The specific attribute set corresponding to different circadian inconsistency ranges or values can be determined using a machine learning model applied to previous collected data, as outlined in detail above. With the specific attribute set selected, the system can generate the desired recommendation in the corresponding linguistic style.

For instance, one example of a recommendation message presented in different styles is set out in Table 2 below.

TABLE 2
                     Recommendation Message:
Circadian            Avoid caffeine after 1:00 pm (default style
inconsistency ranges in lookup table)
0-0.25               If you can avoid caffeine after 1:00 pm, you
                     can fall asleep faster.
0.26-0.65            You need to avoid caffeine after 1:00 pm to
                     fall asleep faster.
0.66-1.0             You can fall asleep faster by avoiding
                     caffeine after 1:00 pm.

In accordance with one or more embodiments, the method may comprise determining a message delivery time corresponding to a time at which the message is to be communicated to the user by means of the sensory output means, and wherein the message delivery time is determined based at least in part on the measure of circadian inconsistency.

By way of example, the method may be configured so as to avoid delivering the recommendation message during time periods in which the circadian inconsistency of the user is high.

In these cases, the delivery time of the message may be selected to be at a time of day in which the circadian inconsistency for day d is below a defined threshold. It has been found that at times of high inconsistency or deviation between expected and actual circadian patterns, a subject is less responsive to recommendations, and thus effect on improving sleep is lower. Reference is made for example to the following paper which details this phenomenon: S. L. Chellappa, C. J. Morris, and F. A. J. L. Scheer, “Daily circadian misalignment impairs human cognitive performance task-dependently,” Sci. Rep., vol. 8, no. 1, pp. 1-11, December 2018.

The threshold may be a pre-defined threshold, or it may be set based on values of the measure of circadian inconsistency. In other words, it may be an absolute threshold, or a relative threshold, normalized to the values of the circadian inconsistency over day d.

As discussed above, the measure of circadian inconsistency may be determined in real time. Thus, in these examples times points where circadian inconsistency is below a defined threshold can be identified in real time.

For example, FIG. 7 illustrates the example set of circadian curves previously illustrated in FIG. 4. FIG. 2 shows that for the second weekend day 68, at around 12:00 (period within circle 82), the highest level of circadian misalignment can be observed. Thus, the method may be configured to avoid communicating behavioral recommendations during this period.

In accordance with one or more embodiments, the method may further comprise monitoring the index of circadian inconsistency over a plurality of days and determining a measure of adherence based on changes in the index of circadian inconsistency.

Adherence may be assumed to be correlated with reduction in circadian inconsistency, i.e. circadian inconsistency falls if the user adheres to the recommended behavior changes. The recommended behavior and/or the properties of the message may in some examples be configured based on the measure of adherence.

A user's adherence can optionally be characterized according to the circadian inconsistency, to derive a measure of circadian resilience. Users may be optionally grouped into phenotypes depending on their circadian resiliency level. This is illustrated schematically in FIG. 8 which shows three different circadian resilience phenotypes, each corresponding to a certain correlation between circadian inconsistency and adherence. A level of circadian resilience can be quantified as the correlation between adherence and circadian inconsistency.

In accordance with one or more embodiments, the method may comprise estimating a reason for low adherence to a behavior change recommendation. This may comprise selecting a most likely of two or more possible reasons.

For example, there may in general be two main categories of reasons for failure to adhere to behavior change recommendations: (a) Extrinsic: misalignment of circadian rhythm from desired sleep routine (indicated by high circadian inconsistency at the time at which the user was recommended to perform the behavior) or (b) intrinsic: lack of desire on the part of the user or conscious resistance to performing the recommended behavior (indicated by low circadian inconsistency at the time at which the user was recommended to perform the behavior).

Thus, the method may include determining whether a user's failure to adhere to a behavior is caused by extrinsic or intrinsic factors, based on detecting a level of circadian inconsistency as a time when the behavior was recommended to be performed.

This determination can in some examples be used as a factor to inform determination of a recommended sleep schedule.

For example: the reason why a user would not follow a recommended or target bedtime may be either be because (a) the user cannot fall asleep at the recommended time due to a circadian misalignment—indicating a sleep disorder such as Delayed Sleep Phase Syndrome (DSPD), or because (b) the user simply does not want to go to sleep at the recommended time.

Based on the reason for lack of adherence, either the recommended time to do the activity or the recommendation itself may be modified.

The specific modification may be performed based on a lookup table prepared in advance, or based on one or more algorithms. The one or more algorithms may include algorithms specific to the recommended behavior change which was not adhered to, and may determine the behavior recommendation modification based on the derived measured of circadian inconsistency for the user.

For example, in the case of DSPD, if it is detected that a user has not followed the behavior recommendation (e.g. go to bed at a defined time, such as 10 pm) and the circadian inconsistency is low at the time of the failure, the recommended bedtime and/or wake time may be altered (e.g. bedtime two hours later and/or wake time two hours later), or use of light therapy may be recommend to alter the circadian cycle.

In one set of embodiments, the one or more algorithms may include one or more machine learning algorithms, which may be trained in advance based on historical data for the specific user, or for a whole population of users. The data inputs to the algorithm may include user-provided answers to subjective questions, as well as user sensor data.

Embodiments of the invention described above employ a processing arrangement. The processing arrangement may in general comprise a single processor or a plurality of processors. It may be located in a single containing device, structure or unit, or it may be distributed between a plurality of different devices, structures or units. Reference therefore to the processing arrangement being adapted or configured to perform a particular step or task may correspond to that step or task being performed by any one or more of a plurality of processing components, either alone or in combination. The skilled person will understand how such a distributed processing arrangement can be implemented. The processing arrangement includes a communication module or input/output for receiving data and outputting data to further components.

The one or more processors of the processing arrangement can be implemented in numerous ways, with software and/or hardware, to perform the various functions required. A processor typically employs one or more microprocessors that may be programmed using software (e.g., microcode) to perform the required functions. The processor may be implemented as a combination of dedicated hardware to perform some functions and one or more programmed microprocessors and associated circuitry to perform other functions.

Examples of circuitry that may be employed in various embodiments of the present disclosure include, but are not limited to, conventional microprocessors, application specific integrated circuits (ASICs), and field-programmable gate arrays (FPGAs).

In various implementations, the processor may be associated with one or more storage media such as volatile and non-volatile computer memory such as RAM, PROM, EPROM, and EEPROM. The storage media may be encoded with one or more programs that, when executed on one or more processors and/or controllers, perform the required functions. Various storage media may be fixed within a processor or controller or may be transportable, such that the one or more programs stored thereon can be loaded into a processor.

Variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure and the appended claims. In the claims, the word “comprising” does not exclude other elements or steps, and the indefinite article “a” or “an” does not exclude a plurality.

A single processor or other unit may fulfill the functions of several items recited in the claims.

The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

A computer program may be stored/distributed on a suitable medium, such as an optical storage medium or a solid-state medium supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems.

If the term “adapted to” is used in the claims or description, it is noted the term “adapted to” is intended to be equivalent to the term “configured to”.

Any reference signs in the claims should not be construed as limiting the scope.

Claims

1. A computer-implemented method comprising:

determining an expected circadian pattern for a user based on a reference dataset comprising historical data for the user;

determining an actual circadian pattern for a given day d based on user sensor data gathered for day d;

determining a measure of circadian inconsistency for day d based on a deviation between the expected circadian pattern and actual circadian pattern for day d;

determining a linguistic message for delivery to the user using a sensory output means, the linguistic message comprising a recommended user behavior for improving sleep, and wherein the recommended user behavior and/or one or more properties of the linguistic message are configured based at least in part on the circadian inconsistency; and

generating a control signal for controlling a sensory output means to generate a sensory output representative of the linguistic message.

2. A method as claimed in claim 1, wherein the measure of circadian inconsistency comprises a measure of said deviation between the expected and actual circadian patterns as a function of time over day d, or at a particular time point in day d.

3. A method as claimed in claim 1, wherein the method comprises determining a message delivery time corresponding to a time at which the message is to be communicated to the user by means of the sensory output means, and wherein the message delivery time is determined based at least in part on the measure of circadian inconsistency.

4. A method as claimed in claim 3, wherein the delivery time of the message is selected to be at a time of day in which the circadian inconsistency for day d is below a defined threshold.

5. A method as claimed in claim 1, wherein the recommended behavior includes a recommended commencement time by the user of the behavior, and wherein the recommended commencement time is determined based at least in part on the determined measure of circadian inconsistency.

6. A method as claimed in claim 1, wherein the reference dataset comprises historical circadian pattern data, or historical sleep and/or activity data for the user.

7. A method as claimed in claim 6, wherein the expected circadian pattern is determined based on fitting parameters of a pre-determined circadian curve equation, the fitting being based on the historical sleep and/or activity data for the patient over one or more days.

8. A method as claimed in claim 1, wherein the sensor data comprises physiological parameter data, and/or movement data.

9. A method as claimed in claim 1, wherein

the measure of circadian inconsistency comprises an index of circadian inconsistency for day d, determined based on deriving a correlation between the expected and actual circadian patterns, and

optionally wherein the method comprises monitoring the index of circadian inconsistency over a plurality of days and determining a measure of adherence based on changes in the index of circadian inconsistency.

10. A method as claimed in claim 1, wherein the recommended behavior is determined based on the circadian inconsistency, and based on use of a lookup table.

11. A method as claimed in claim 1, wherein the method comprises configuring one or more properties of the linguistic message based on the circadian inconsistency, and wherein the properties include a wording of the linguistic message.

12. A method as claimed in claim 11, wherein the configuring of the wording of the linguistic message, for communicating a particular behavior change, is performed by selecting from a lookup table one of a plurality of pre-determined linguistic messages conveying said behavior change and each with a different wording.

13. A method as claimed in claim 11, wherein the configuring of the wording of the message is performed using a machine learning algorithm.

14. A computer program product comprising computer program code, the computer program code being configured, when executed on a processor, to cause the processor to perform a method in accordance with claim 1.

15. A processing arrangement comprising:

an input/output; and

one or more processors adapted to: determine an expected circadian pattern for a user based on a reference dataset for the user accessed using the input/output; determine an actual circadian pattern for a given day d based on patient sensor data gathered for day d, the sensor data received at the input/output; determine a measure of circadian inconsistency for day d based on a deviation between the expected and actual circadian patterns for day d; determine a linguistic message for delivery to the user using a sensory output means, the message comprising a recommended user behavior, wherein the recommended behavior and/or one or more properties of the linguistic message are configured based on the circadian inconsistency; and generate a control signal for controlling a sensory output means to generate a sensory output representative of the linguistic message.