Method for Detection of a Relapse Into a Depression or Mania State Based on Activity Data and/or Data Obtained by Questioning the Patient

Abstract

The invention relates to a method for detection of a relapse into a depression or mania state of a patient from a remission state wherein motor activity data is recorded using a wearable device worn by the patient and is received as input data by an evaluating unit and/or mood data is acquired by obtaining a questionnaire which has been completed by the patient, the questions of the questionnaire relating to the mania state, to the depression state and the questionnaire including at least one control question for checking the awareness and/or the ability to focus of the patient, the questions being designed such that they can be answered by multiple choice, and wherein the answers of the patient are input as input data into the evaluating unit, the input data is analyzed by the evaluating unit, wherein the condition of the patient is classified as remission, mania or depression by means of machine learning, and wherein a relapse is detected if the patient is classified as mania or depression.

Further aspects of the invention relate to an evaluating system for detection of a relapse into a depression or mania state of a patient in a remission state based on motor activity data and/or mood data.

Claims

1. Method for detection of a relapse into a depression or mania state of a patient from a remission state wherein i) motor activity data is recorded using a wearable device worn by the patient and is received as input data by an evaluating unit and/or ii) mood data is acquired by obtaining a questionnaire which has been completed by the patient, the questions of the questionnaire relating to the mania state, to the depression state and the questionnaire including at least one control question for checking the awareness and/or the ability to focus of the patient, the questions being designed such that they can be answered by multiple choice, and wherein the answers of the patient are input as input data into the evaluating unit, the input data is analyzed by the evaluating unit, wherein the condition of the patient is classified as remission, mania or depression by means of machine learning, and wherein a relapse is detected if the patient is classified as mania or depression.

2. Method according to claim 1, wherein motor activity data according to variant i) is analyzed using logistic regression method utilizing a time series of motor activity data and/or wherein mood data according to variant ii) is analyzed using logistic mixed effects models utilizing a time series of obtained mood data.

3. Method according to claim 1, wherein the machine learning model used according to variant i) comprises binary classification models for classification of remission/mania, depression/mania and remission/depression and for classification into one of the three classes remission, mania and depression a) a score is computed for each of the three classes based on probabilities obtained from the binary classification models, or b) one of the three classes is selected based on majority voting where two of the three classification models indicate the same class.

4. Method according to claim 2, wherein features are extracted from the time series of motor activity data and the extracted features are used in the logistic regression model, the features being selected from sleep features, in particular sleep duration, activity during sleep and fragmentation of sleep, activity distribution within the day, in particular the amount of activity in the active part of a day, overall activity level, fragmentation of activity within a day, the timing of activities throughout a day and combinations of at least two of said features.

5. Method according to claim 1, wherein in variant i) further data is collected in addition to motor activity data, wherein the further data is selected from daylight duration, moon phase, body temperature, pulse, blood pressure, data on skin galvanic response and combinations of at least two of said features.

6. Method according to claim 1, wherein the time series of motor activity data according to variant i) is pre-processed to extract epochs which represent motor activity data over a time window of individual days or weeks or represent the time of night-sleep.

7. Method according to claim 1, wherein the questions according to variant ii) are designed such that they can be answered using a scale from a given minimum value to a given maximum value, preferably being selected in the range of from −10 to 10, more preferably in the range of from 0 to 10 and in particular, the scale is from 0 to 4.

8. Method according to claim 7, wherein the analysis of mood data according to variant ii) comprises computing a probability for a depression and computing a probability for a mania.

9. Method according to claim 2, wherein computing of the probability for a depression comprises computing of a sum of the scales assigned the depression related questions and non-specific questions for a time period consisting of mood data acquired for the current week and the previous week and including said sum in the fixed part of the linear mixed effects model.

10. Method according to claim 2, wherein computing of the probability for a mania comprises computing of a sum of the scales assigned the mania related questions for a time period consisting of mood data acquired for the current week and including said sum in the fixed part of the linear mixed effects model.

11. Method according to claim 7, wherein a patient is classified as remission if both the probability for a depression and the probability for a mania are below given depression and mania thresholds, respectively, and, if the patient is not classified as remission, the patient is classified as depression if the probability for a depression is larger than the probability for a mania and is classified as mania otherwise if both the probability for a depression and probability for a mania are above or equal their respective thresholds and is otherwise classified as mania if the probability for a mania is above or equal the mania threshold and as depression if the probability for a depression is above or equal the depression threshold.

12. Method according to claim 1, wherein the design of the questions of the questionnaire of variant ii) is validated by using principal component analysis and a test dataset which includes ground truth.

13. Method according to claim 1, wherein the machine learning model(s) used for evaluation of motor activity data according to variant i) is trained and/or verified against clinical classification scales.

14. Method according to claim 1, wherein the clinical scale for validating the questionnaire according to variant i) and/or for verifying the model(s) according to variant ii) is the Clinical Global Impression scale (CGI) to characterize the severity of a bipolar disorder, the MADRS scale, Bipolar Depression Rating Scale (BDRS) in case of depression and the YMRS scale, Observer-Rated Scale for Mania (IRSM) or the Bech-Rafaelsen Mania Rating Scale (MAS) in case of mania.

15. Evaluating system for detection of a relapse into a depression or mania state of a patient in a remission state based on motor activity data and/or mood data, characterized in that the evaluating system comprises an evaluating unit which is configured to analyze motor activity data and/or mood data according to a method of claim 1.

Description

EXAMPLE DATASETS

[0056] For validation and training of the machine learning models used in variants i) and ii), a study was performed which included up to 400 patients with bipolar disorder (BD) and 50 healthy controls. The participation time frame was 18 months for patients with BD and 3 months for healthy controls. The study population included men and women between 18 and 60 years old in standard clinical treatment for BD (ICD-10 diagnosis F31.). The experimental group was distributed into 3 groups: Core, Peripheral 1 (P1) and Peripheral 2 (P2).

[0057] The inclusion criteria to the study include: diagnosis and current treatment of Bipolar disorder (F31. ICD-10) (confirmed in the CORE group by an institutional psychiatrist). The age of the patients included in the study was between 18 and 60.

[0058] All study participants were instructed to wear a wearable device for monitoring activity all of the time and to complete a self-reported questionnaire according to table 1 on a weekly basis. Accordingly, the data acquired by this study is suitable for both variants i) and ii) of the present invention.

[0059] Patients in the Core group underwent an initial medical investigation at National Institution of Mental Health Czech Republic (NIMH-CZ) upon enrollment and their diagnosis of BD was either confirmed or they were moved to the P1 or P2 group. The patients were assessed by the Montgomery-Asberg Depression Rating Scale MADRS and the Young Mania Rating Scale YMRS every month. The clinical scales were used as objective measures of depressive and manic states. Information about clinical episodes, hospitalizations and work incapacities were collected retrospectively every 6 months from the caregiving psychiatrists in all patient groups. The rating team executing the clinical scales consisted of psychologists, who went through training sessions and inter-rater reliability evaluations during the study.

[0060] Dataset for Variant i)

[0061] Records of cases with hospitalizations were excluded from the dataset, as the activity is restricted in hospitals.

[0062] Only records having less than a defined threshold of missing activity data within a 14-day were included in the dataset. In the present example, the threshold was set to allow 10% of missing values in a 14-day window which allows for 1.4 days of missing activity data.

[0063] When features were extracted from the records based on activity data within a certain timespan defined by a day or by a window, preprocessing was performed in case of missing values. In particular, data processing was performed to fill in missing values.

[0064] In the case of features based on activity data within a given day, handling of missing data was done differently than for features based on a window. For daily based features, as these were only individual days missing, the missing values were estimated as a linear interpolation of surrounding values.

[0065] In case of week-window based features, the first seven values of these features are using unspecified data from before the annotated episode. Therefore any missing value was replaced by the average of values from the second week of the episode. In case of two-week-window based features, all, except the last value, of these features are using days from unspecified data prior to an annotated episode all missing values are replaced by the last day value of the episode.

[0066] In the following, some features which may be used are described in more detail as an example. It is to be understood that the proposed method may also be applied to different features or a different selection of the described features.

[0067] The features used may be divided by their origin:

[0068] The first group are parametric features which are obtained by fitting the cosine function of a given period (24 hours) on the raw activity data in a week or two weeks long window. The function is fitted using the least squares method. Obtained features are the offset of the fitted cosine function, and amplitude of the fitted cosine, and acrophase which is representing the phase shift of the function, and daily activity rhythm which is the amplitude normalized by mesor.

[0069] The second group is represented by nonparametric features (no specific function is used to represent the data). These features are estimated either based on individual days or an average obtained by averaging 7 or 14 consecutive days. These parameters are average activity in the 10 most active hours of the day (M10), average activity of the least active 5 hours of the day (L5), relative amplitude obtained from formula (M10-L5)/(M10+L5), ant the midtimes of both L5 and M10 windows (M10/L5 time). To describe stability of the rhythm within a day or between the consecutive days the intradaily variability (IV) and interdaily stability (IS) features were estimated. These were calculated based on a known formula from 20 to 60 minute aggregated actigraphic data segments. The formular is published in Van Someren E J, Swaab D F, Colenda C C, Cohen W, McCall W V, Rosenquist P B. “Bright light therapy: improved sensitivity to its effects on rest-activity rhythms in Alzheimer patients by application of nonparametric methods.” Chronobiol Int. 1999; 16(4):505-518. http://www.ncbi.nlm.nih.gov/pubmed/10442243. Other features focus on levels of activity, the histogram of activities without sleep is divided individually into four levels (low, sedentary, moderate, high) and percentage of each level within a calendar day is calculated. Then activities are estimated in different parts of day midnight to 6 AM, 6 AM to noon and so on (QA1-QA4), special focus is given the time after wakeup and before sleep onset (the average activity in a two hour window).

[0070] The third group are sleep describing features. These are duration of the main sleep of the day, as well as all sleeps in a given day (18 o'clock previous day till 18 o'clock actual day). Based on the main daily sleep the sleep quality is detected using number of immobile minutes (minutes where almost no movement is detectable <20 mg {after filtering}), the midtime of the main sleep, Wake after sleep onset (number of minutes that are not detected as sleep during the main sleep of a day). To include part of the year and approximate night illumination features of daylight duration and moon phase are also included in the models.

[0071] In a final dataset used for variant i), there were 56 depression episodes, 25 mania episodes, and 172 remission episodes, which were analyzed. Multiple episodes of one type may be found for some patients. As the models are trained solely on patients with episodes in both tested classes, there were 10 patients with both depression and mania, 15 patients with both remission and mania and 28 patients with remission and depression.

[0072] Dataset for Variant ii)

[0073] Matching of the clinical scales and the self-report questionnaires was performed based on the date of the event of the filling-in of the questionnaire and the date of the structured interview for the assessment of the clinical scales. If the difference between these dates was lower than a threshold value, the clinical scale and the self-report were considered matching. In the present example, the threshold was set to 3 days for questionnaires relating to the present situation (zero delay). For questionnaires relating to the previous week, the threshold was set to 10 days and for questionnaires relating to the two weeks before the clinical scale, the threshold was set to 17 days. All the matchings were unique, none of the clinical scales was matched to more than one of the self-reports and vice versa for each delay. In order to exploit the dependencies in the consecutive questionnaires, the data from more than one week preceding the clinical scale were considered.

[0074] Only patients that had both relapse and remission according to the clinical scales were included in the obtained dataset. A relapse according to the clinical scales is defined as ΣMADRS >=15 in case of depression and ΣYMRS >=15 in case of mania. Only complete records were included in the dataset.

[0075] Concerning depression according to the MADRS scale, the dataset includes records of 67 patients with 1264 paired observations for the concurrent clinical scales and self-report, 1159 observations from the previous week and 11180 two weeks before. From the 1264 observations there are 251 relapses and 1013 relapses.

[0076] Concerning mania according to the YMRS scale, the dataset includes records of 31 patients represented by 614 paired observations for the concurrent clinical scales and self-reports, 560 observations from the previous week and 563 observations two weeks before. From the 614 clinical scales there are only 74 relapses and 540 relapses.

[0077] Training Results Using the Example Dataset

[0078] Analysis of Activity Data According to Variant i)

[0079] An example Depression-Mania model is based on cosine of the amplitude, from non-parametric L5, L5 time, M10, RA, daily part of sedentary activity, IV, IS, Average activity during morning and evening (QA2, QA4), based on sleep the immobile minutes (percentage), sleep midtime. The model also uses a year season represented by daylight duration. The most important features are L5, amount of sedentary activity and IV.

[0080] An example Depression-Remission model is based on cosine acrophase, mesor and amplitude, based on non-parametric analysis IS, IV, M10, M10 time, L5, L5 time, before and after sleep average activities, daily part of moderate and low activities, and activity during morning hours (QA2), based on sleep the WASO, and sleep duration. The strongest predictors are acrophase, IS, IV, M10 and activity before sleep

[0081] An example Mania-Remission model uses from cosine analysis acrophase, and mesor, from non-parametric analysis L5 time, L5, M10 time, IV, IS, night and morning activity (QA1, QA2), DAR, RMSSD, and daily percentage of sedentary and low activity, based on sleep percentage of immobile minutes, main sleep duration, and daily (overall) sleep duration. The model also uses a year season represented by daylight duration. The most important are immobile minutes, L5 time, IV, and M10 time.

[0082] Two models of each type based on logistic regression were trained using the example dataset. A first model includes values computed using features based on the evaluation of motor activity data of a certain time window, for example a day. A second model is based on difference values of the respective features between two time periods (these are the same features as in the first model, only in this case it's differences from individual patients mean values). The results for the two models are given in table 2.

[0083] The features for the respective models are selected separately based on the prediction model. The features calculated for each day (end of window) are used in the model to distinguish between states and the final output (classification) is obtained based on the combined results of the models.

TABLE-US-00002 TABLE 2 Dep-man Dep-rem Man-rem Logistic Accuracy 0.61 Accuracy 0.61 Accuracy 0.61 Regression Sensitivity 0.65 Sensitivity 0.60 Sensitivity 0.49 Specificity 0.55 Specificity 0.52 Specificity 0.64 AUC: 0.71# AUC: 0.58# AUC: 0.70# Logistic Accuracy 0.66 Accuracy 0.57 Accuracy 0.60 Regression Sensitivity 0.70 Sensitivity 0.61 Sensitivity 0.44 (diff features) Specificity 0.64 Specificity 0.51 Specificity 0.71 AUC: 0.66# AUC: 0.56# AUC: 0.70#

[0084] In the table, AUC denotes the area under the receiver operating characteristic curve (ROC curve) and is an indication of the performance of the classification model. A value of 0 indicates 100% wrong classification and a value of 1 indicates a 100% correct classification.

[0085] Analysis of Mood Data According to Variant ii)

[0086] In the present example, the first logistic mixed effects model uses a sum of the scales assigned to the mania related questions for a time period consisting of mood data acquired for the current week in the fixed part of the model.

[0087] The second logistic mixed effects model uses a sum of the scales assigned to the depression related questions and non-specific questions for a time period consisting of mood data acquired for the current week and the previous week in the fixed part of the model.

[0088] Both the first logistic mixed effects model and the second logistic mixed effects model preferably use random intercepts in the random effects part.

[0089] For training of the first and second logistic mixed effects models, the dataset was split into training and testing data. The proportions were 70% of the data for training and 30% for testing. 999 sets of training and testing data were randomly chosen from the dataset. The models were trained on the training sets; the training sets predictions and true values of the scale based relapses were used to construct a receiver operating characteristic (ROC), which was used to estimate the threshold probability maintaining 0.9 specificity to reduce the number of false alarms. The model prediction was modified to maintain specificity 0.9 by altering the threshold probability according to the ROC curve from training data prediction.

[0090] In the case of the first logistic mixed effects model for assessing mania, the validation results for the trained model are summarized in Table 3 below:

TABLE-US-00003 TABLE 3 Set Accuracy AUC Sensitivity Specificity training 0.886 0.916 0.760 0.903 testing 0.875 0.871 0.711 0.897

[0091] In the case of the second logistic mixed effects model for assessing depression, the validation results for the trained model are summarized in Table 4 below

TABLE-US-00004 TABLE 4 Set Accuracy AUC Sensitivity Specificity training 0.883 0.931 0.807 0.902 testing 0.849 0.878 0.717 0.880

[0092] The presented results show the predictive ability of the self-report questionnaire in scale-based relapse prediction. The area under the ROC curve (AUC) is 0.871 in the case of the first model (mania) and is 0.878 in the case of the second model (depression) on 999 randomly chosen testing data sets. The value of AUC is less sensitive to the class imbalance than the accuracy. The class imbalance is rather high in both depression and mania (only 19.8% events represent relapses for depression and 12.0% for mania) and using only accuracy, it would be possible to obtain better results just by assigning all the events as remission (appr. 0.80 and 0.88 for depression and mania respectively for the case of assigning all events as remission; in comparison to the actual observed 0.85 and 0.88 for depression and mania respectively). The relatively high values of the sensitivity in both cases indicate that the model is actually correctly predicting the relapses.

[0093] Preferably, the design of the questions of the questionnaire of variant ii) is validated by using principal component analysis and a test dataset which includes ground truth.

[0094] Principal component analysis is a method for analyzing the relation of data and for feature extraction. The method involves an orthogonal linear transformation that transforms the data to a new coordinate system defined by the eigenvectors. The transformation is performed such that the greatest variance of a scalar projection of the data comes to lie on the first coordinate. The second largest variant comes to lie on the second coordinate and so on.

[0095] The mania related questions being assigned to the same principal component would indicate that the questions are related. Likewise, the depression related questions being assigned to the same principal component would indicate that the respective questions are related. Still further, if questions relating to mania as well as questions relating to depression would be assigned to the same principal component this would indicate a low selectivity. Accordingly, the two major principal components should only include questions relating to mania or relating to depression, respectively.

[0096] Tables 5 and 6 present the results of the PCA analysis performed on the example dataset. There are two dominant components (self report pc1 50.3%, pc2 22.1%). The first principal component pc1 relates to the depression-related questions (pc1 high values of loadings for questions 1-4) and to the non-specific questions (pc1 high values of loadings for questions 9 and 10). The second principal component pc2 summarizes the mania-related questions (pc2 high values of loadings for questions 5-8), the non-specific questions (pc2 high values of loadings for questions 9 and 10), but their loading coefficients are smaller than those of the main mania-related questions. All the remaining principal components pc3 to pc10 explain less than 7% variability each.

[0097] The results of the principal component analysis indicate that the questions presented in table 1 are suitable as they relate to depression and mania, respectively, as intended.

TABLE-US-00005 TABLE 5 PC1 PC2 PC3 PC4 PC5 PC6 PC7 PC8 PC9 PC10 Q 1 0.428 0.281 −0.272 −0.348 −0.44 −0.57 0.129 Q 2 0.43 0.304 −0.14 −0.282 0.716 −0.311 Q 3 0.42 −0.118 0.202 0.721 0.427 0.165 0.178 Q 4 0.396 0.237 −0.34 −0.286 0.212 −0.109 0.719 −0.103 Q 5 0.502 0.271 0.364 −0.358 −0.157 −0.376 −0.164 −0.468 Q 6 0.492 0.225 −0.164 −0.103 0.284 0.761 Q 7 0.425 0.156 0.116 0.853 −0.183 Q 8 0.432 −0.367 0.727 0.15 −0.323 −0.139 Q 9 0.366 0.259 −0.608 −0.597 0.225 0.134  Q 10 0.399 0.208 −0.46 −0.216 0.639 −0.333 −0.116

TABLE-US-00006 TABLE 6 PC1 PC2 PC3 PC4 PC5 PC6 PC7 PC8 PC9 PC10 Standard deviation 1.416 0.938 0.529 0.422 0.401 0.348 0.333 0.326 0.274 0.269 Proportion of variance 0.503 0.221 0.07 0.045 0.04 0.03 0.028 0.027 0.019 0.018 Cumulative proportion 0.503 0.723 0.794 0.838 0.878 0.909 0.936 0.963 0.982 1 of variance

[0098] For further verification of the selection of the questions so that the obtained results correspond to an established clinical scale, mood data according to variant ii) is preferably analyzed using a linear mixed effects model utilizing a time series of obtained mood data.

[0099] Mixed linear effects models may be used in binary classification problems and comprise fixed and random effects. In a first mixed effects linear model describing the probability for the patient being in the state of remission or mania, the mania related questions are included as a linear combination. In a second mixed effects linear model describing the probability for the patient being in the state of remission or depression, the depression related questions are included as a linear combination.

[0100] The first model uses the scales assigned to the mania-related questions as a fixed effect and includes random slopes and random intercepts for the sum of mania-related questions for each patient.

[0101] Likewise, the second model uses the scales assigned to the depression-related questions as fixed effect and includes random slopes and random intercepts for the sum of depression-related questions for each patient.

[0102] Preferably, the machine learning model(s) used for evaluation of activity data according to variant i) is trained and/or verified against clinical classification scales.

[0103] Suitable clinical scales for validating the questionnaire according to variant i) and/or for verifying the model(s) according to variant ii) may, for example, be selected from the Clinical Global Impression scale (CGI) to characterize the severity of a bipolar disorder, the MADRS scale, Bipolar Depression Rating Scale (BDRS) in case of depression and the YMRS scale, Observer-Rated Scale for Mania (IRSM) or the Bech-Rafaelsen Mania Rating Scale (MAS) in case of mania. The clinical scale may be selected independently for each of the variants i) and ii) and for each of the model(s) used in any of the variants i) and ii).

[0104] The invention further relates to a method for treating a patient having bipolar disorder wherein i) activity data is recorded using a wearable device worn by the patient and is input into an evaluating system and/or ii) mood data is obtained by obtaining a questionnaire which has been completed by the patient, the questions of the questionnaire relating to the mania state, to the depression state and further include at least one control question for checking the awareness and/or the ability to focus of the patient, the questions being designed such that they can be answered by multiple choice, and the answers of the patient are input into the evaluating system, the evaluating system analyzes the input data by means of machine learning, wherein the condition of the patient is classified as remission, mania or depression, and wherein a relapse is detected if the patient is classified as mania or depression. Further, in case a relapse is detected, therapeutic intervention may be performed.

[0105] In case a therapeutic intervention is needed, the physician may, after he/she may have discussed the possible consequences with the patient, take the appropriate measures depending on the disease. This may include the administration or the adjustment of the medication needed for treating the patient's disease, other measures like psychotherapy, light therapy, rTMS (Repetitive Transcranial Magnetic Stimulation) or physical therapy as well as surgery. Administration of medication may include antipsychotics and mood stabilizers e.g. based on lithium/valproate like lurasidone, olanzapine and quetiapine.

[0106] It is a further object of the invention to provide an evaluating system for detection of a relapse into a depression or mania state of a patient in a remission state based on activity data and/or mood data. The evaluating system comprises an evaluating unit which is configured to analyze activity data and/or mood data according to any one of the methods described herein.

[0107] Preferably, the system comprises at least one wearable device for obtaining activity data which is configured to periodically transmit recorded time series of activity data to the evaluating unit. The wearable device is preferably configured as a wristband or wristwatch.

[0108] The system may further comprise an interface device configured to receive activity data and/or mood data from other devices, in particular user devices such as smartphones or tablets.

[0109] The invention is not limited to the exemplary embodiments described herein and the aspects highlighted therein. Rather, within the range indicated by the claims, a large number of variations are possible.