Adaptive Neurological Testing Methods

Abstract

A method, computer-implemented method, and device for assessing the neurological condition of a subject, all adaptively, comprising the steps of: asking one or more seed questions; obtaining one or more answers to the one or more seed questions; and calculating a score value of a latent trait of the subject from the answers to the one or more seed questions, and an adaptive testing loop comprising the steps of: (a) selecting one or more further questions from a set of questions based on the score value; (b) asking the subject one or more further questions; (c) updating the score value based on the answers to the one or more further questions; and (d) determining whether a test completion criterion has been met, wherein steps (a) through (d) are repeated in order until a test completion criterion has been met, and providing a result of the test based on the score value indicative of the neurological condition of the subject.

Description

[Technical field]

[0001] The present invention relates to an adaptive method for testing the mental capabilities of a subject. [Background technology]

[0002] Diagnosis of Alzheimer's disease and monitoring the mental abilities of subjects once diagnosed is an active and multifaceted area of ​​research. Alzheimer's disease is a neurological disorder in which excess proteins build up in the brain, impairing neuronal function and ultimately causing cell death. The disease is characterized by a continuous progression, but at different rates in different individuals. While being able to identify early stages of Alzheimer's disease would be useful, an obstacle preventing early diagnosis of Alzheimer's disease is that significant brain damage has already occurred before the disease is detectable by traditional methods.

[0003] Opinions differ about the best way to diagnose and stratify patients with Alzheimer's disease. Popular approaches include genotyping, imaging techniques such as MRI or PET scanning, and measuring patient decline using industry-standard neuropsychological assessments.

[0004] US Patent Application Publication No. US2008 / 0118899 relates to identifying individuals within a population who are at particular risk of suffering from disorders associated with neurocognitive degeneration, such as Alzheimer's disease. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] US Patent Application Publication No. US2008 / 0118899 Summary of the Invention

[0006] The inventors have provided a novel neuropsychological assessment that is believed to be more robust and sensitive than previous assessments. This assessment utilizes Item Response Theory (IRT), also known as Latent Trait Theory, in which questions or items in a comprehensive test relate to underlying latent traits. The central concept of IRT is the association between the way participants respond to a question and the latent trait that the question represents.

[0007] In some aspects, the present invention relates to an adaptive method of testing a neurological condition of a subject, where the questions used in the testing are selected based on the scores obtained for one or more latent traits of the subject.

[0008] The positive aspects of using neuropsychological tests for the assessment of dementia are significant. The tests are non-invasive, relatively inexpensive, and can be administered extensively and repeatedly as needed. However, repeated testing with the same questions can lead to learning effects. Each test focuses on a single domain of symptoms, and therefore multiple tests are required to adequately quantify the severity of dementia, and each test must be completed in its entirety as missing items invalidate the test, leading to lengthy testing times. Thus, there is the possibility of false proxies of latent traits due to training or fatigue. Using each of these tests as a proxy for the severity of a patient's dementia does not account for the fact that a typical battery of neuropsychological tests for dementia patients examines many different domains of a patient's abilities. It is possible that each of these domains does not fail in the same pattern for each patient, but rather deteriorates independently of each other. Current scoring of the tests has been shown to be an inaccurate way of gaining insight into a patient's abilities because it does not take into account the varying difficulty of the questions within the tests.

[0009] Accordingly, in a first aspect, the present invention provides a method for adaptively testing a neurological condition of a subject, the method comprising the steps of: asking one or more seed questions; obtaining one or more answers to the one or more seed questions; and calculating a score value of a latent trait of the subject from the answers to the one or more seed questions, the method comprising an adaptive testing loop comprising the steps of: (a) selecting one or more further questions from a set of questions based on the score value; (b) asking the subject the one or more further questions; (c) updating the score value based on the answers to the one or more further questions; and (d) determining whether a test completion criterion is met, wherein steps (a) to (d) are repeated in sequence until a test completion criterion is met, and providing a result of the test based on the score value indicative of the neurological condition of the subject.

[0010] The method of the first aspect can maximize the information that can be extracted from questions or items in industry-standard neuropsychological tests by considering them separately from the test that they are derived from and evaluating their difficulty with respect to multiple areas of patient performance.The method can be applied to evaluate the difficulty of items and the performance of patients with respect to multiple aspects of patient decline.This provides a more sensitive approach to scoring patients, since it allows items to be a measure of more than one underlying latent trait, and explicitly accounts for the fact that decline is not uniform across patient groups.It also takes into account the different difficulty of items with respect to the multiple underlying latent traits that contribute to patient decline.

[0011] The steps may be interleaved with one or more sub-steps and / or may not necessarily be performed directly one after the other, in some cases the steps are performed directly one after the other. In this specification, an item may refer to a question or exercise from one or more neuropsychological testing methods. The neurological condition to be tested may be a determination of whether the subject has dementia or Alzheimer's disease.

[0012] Optional features of the present invention are described herein. These may be applied alone or in any combination with any aspect of the present invention.

[0013] The method of the invention may be a method for assessing, testing or classifying the neurological condition of a subject for any purpose, for example, the score value or other result of the test may be used to classify the psychological or disease state of the subject according to predetermined criteria.

[0014] The subject may be any human subject. In one embodiment, the subject may be one suspected of suffering from a neurocognitive disease or disorder, such as a neurodegenerative or vascular disease described herein.

[0015] In one embodiment, the method is for diagnosing or prognosing a cognitive impairment, e.g., a neurocognitive disorder, in a subject.

[0016] Here, the underlying underlying trait, eg, one or more of the cognitive trait, the dementia trait and the depression trait, arises due to neurodegeneration resulting from the disease.

[0017] In one embodiment, the disease is mild to moderate Alzheimer's disease.

[0018] In one embodiment, the disease is mild cognitive impairment.

[0019] In one embodiment, the disease is dementia - for example, vascular dementia.

[0020] In one embodiment, the method is aimed at determining the risk of neurocognitive disorder in a subject. In addition, optionally, the risk may be calculated using further factors, such as age, lifestyle factors, and other measured physical or psychological criteria. The risk may be classified as "high" or "low", or may be expressed as a grade or spectrum.

[0021] In one embodiment, the method is for monitoring the progression of a neurocognitive disease in a subject diagnosed with the disease. For example, the method may be for monitoring neurodegeneration or other cognitive or related decline in a subject.

[0022] The method may optionally include treating a subject diagnosed with a neurocognitive disease.

[0023] Thus, a subject may be one receiving medication or other treatment for a neurocognitive disorder, in which case the methods may be used to monitor the effectiveness of the treatment, e.g., in relation to improved or reduced decline, e.g., compared to a placebo or other control.

[0024] In one embodiment, the method is a method for recruiting subjects for a clinical trial, and the method is used to classify subjects according to whether they meet the study recruitment criteria or not.

[0025] The score value may be a latent trait performance score. The latent trait performance score may indicate the subject's performance in one or more latent traits.

[0026] Each of the further questions may have an associated difficulty rating and an indication of relevance to at least one of the one or more latent traits of the subject. Selection of the one or more further questions may be further based on one or both of the associated difficulty rating and the indication of relevance to the one or more latent traits of the subject.

[0027] The method may calculate a score value for each of a plurality of latent traits of the subject from the answer(s) to one or more seed questions, and selecting the further question(s) may be further based on at least one of these score values. In some examples, the method calculates a score value for each of three latent traits of the subject: a cognition latent trait, a dementia latent trait, and a depression latent trait. The method may include an initial step of determining a weighting value for each of the plurality of latent traits of the subject. The weighting value may be binary and indicate whether or not a particular latent trait of the subject should be investigated. Selecting one or more further questions may be further based on the weighting value for at least one of the plurality of latent traits of the subject.

[0028] The one or more latent traits of the subject may include one or more of a cognitive trait, a dementia trait, and a depression trait.

[0029] A subject's score for the latent trait or each of the latent traits may be calculated using an expectation maximization algorithm and / or a generalized linear model. The generalized linear model may have the following formula:

[0030]

number

[0031] Here, Y i,q is the response by subject i at a particular visit, q is the question, α is the intercept, and β q , γ q and δ q are the difficulty scores for questions related to cognitive, dementia and depression traits, respectively, and cog i , dem i and dep i are the subject's ability scores related to the cognitive, dementia and depression traits, respectively.

[0032] Seed questions and / or further questions may be based on the Alzheimer's Disease Assessment Scale-cognitive test (WG Rosen, RC Mohs, and KL Davis, A new rating scale for Alzheimer's disease, American Journal of Psychiatry, Vol. 141, 1984), the Alzheimer's Disease Cooperative Study - Activities of Daily Living test (D. Galasko, D. Bennett, M. Sano, C. Ernesto, R. Thomas, M. Grundman, and S. Ferris, An inventory to assess activitiesof daily living for clinical trials in Alzheimer's disease. Alzheimer's Disease and Associated Disorders, Vol. 11(Suppl. 2), 1997), the Neuropsychiatric Inventory test (JL Cummings, M. Mega, K. Gray, S. Rosenberg-Thompson, DA Carusi and J. Gorbein, The Neuropsychiatric Inventory: Comprehensive assessment of psychopathology in dementia, Neurology, vol. 44, 1994), the Montgomery-Asberg Depression Rating Scale (S.A. Montgomery and M. Asberg, A new depression scale designed to be sensitive to change. British Journal of Psychiatry, vol. 134, 1979) and the Mini-Mental State Examination Test (M.F. Folstein, S.E. Folstein, P.R.McHugh, Mini-Mental State: A practical method for grading the cognitive state of patients for the clinician. Journal of Psychiatric Research, Vol. 12, 1975). As will be appreciated, essentially any question from any validated neuropsychological assessment (suitably characterized, as discussed below) or equivalent to such a currently validated neuropsychological assessment may be utilized in the adaptive testing methods herein.

[0033] The seed question and / or further questions may be used to calculate and update a score value for each of a plurality of latent traits for the subject.

[0034] Selecting the one or more further questions may be further based on an information content for each question in the database of questions. The information content for each question in the database of questions may be calculated for the subject based on the score values ​​of the subject's latent traits, and the one or more questions having the greatest information content may be selected from the database of questions.

[0035] In a second aspect, the present invention provides a computer-implemented method for adaptively testing a neurological condition of a subject, the method comprising the steps of: displaying one or more seed questions on a display; receiving answers to the one or more seed questions via an input device; calculating by a processor a score value of a latent trait of the subject from the received answers to the one or more seed questions; and entering an adaptive testing loop comprising the steps of: (a) selecting one or more further questions from a database of questions based on the score value; (b) displaying the one or more further questions on a display; (c) updating the score value based on the received answers to the one or more further questions; and (d) determining whether a test completion criterion has been met, wherein the computer-implemented method repeats steps (a) to (d) in sequence until a test completion criterion is met and provides a result of the test based on the score value indicative of the neurological condition of the subject.

[0036] The steps may be interleaved with one or more sub-steps and / or may not necessarily be performed directly one after the other. In some cases, the steps are performed directly one after the other. In this specification, an item may refer to a question or exercise from one or more neuropsychological testing methods. The neurological condition to be tested may be a determination of whether the subject has dementia or Alzheimer's disease.

[0037] Optional features of the present invention are described herein. These may be applied singly or in any combination with any aspect of the present invention.

[0038] The score value may be a latent trait performance score. The latent trait performance score may indicate the subject's performance in one or more latent traits.

[0039] Each of the further questions may have an associated difficulty rating and an indication of relevance to at least one of the one or more latent traits of the subject. Selection of the one or more further questions may be further based on one or both of the associated difficulty rating and the indication of relevance to the one or more latent traits of the subject.

[0040] The method may calculate a score value for each of a plurality of latent traits of the subject from the answer(s) to one or more seed questions, and selecting the further question(s) may be further based on at least one of these score values. In some examples, the method calculates a score value for each of three latent traits of the subject: a cognition latent trait, a dementia latent trait, and a depression latent trait. The method may include an initial step of determining a weighting value for each of the plurality of latent traits of the subject. The weighting value may be binary and indicate whether or not a particular latent trait of the subject should be investigated. Selecting one or more further questions may be further based on the weighting value for at least one of the plurality of latent traits of the subject.

[0041] The one or more latent traits of the subject may include one or more of a cognitive trait, a dementia trait, and a depression trait.

[0042] A subject's score for the latent trait or each of the latent traits may be calculated using an expectation maximization algorithm and / or a generalized linear model. The generalized linear model may have the following formula:

[0043]

number

[0044] Here, Y i,q is the response by subject i at a particular visit, q is the question, α is the intercept, and β q , γ q and δ qare the difficulty scores for questions related to cognitive, dementia and depression traits, respectively, and cog i , dem i and dep i are the subject's ability scores related to the cognitive, dementia and depression traits, respectively.

[0045] The seed questions and / or further questions may be selected from any one or more of the Alzheimer's Disease Assessment Scale-Cognitive Function (ADAS-cog), Alzheimer's Disease Cooperative Study-Activities of Daily Living (ADL), Neuropsychiatric Inventory (NPI), Montgomery-Asberg Depression Rating Scale (MADRS), and Mini-Mental State Examination (MMSE). As will be appreciated, essentially any question from any (suitably characterized, as discussed below) valid neuropsychological assessment may be utilized in the adaptive testing methods herein.

[0046] The seed question and / or further questions may be used to calculate and update a score value for each of a plurality of latent traits for the subject.

[0047] Selecting the one or more further questions may be further based on an information content for each question in the database of questions. The information content for each question in the database of questions may be calculated for the subject based on the score values ​​of the subject's latent traits, and the one or more questions having the greatest information content may be selected from the database of questions.

[0048] In a third aspect, the present invention provides a device for performing adaptive testing of a psychological state of a subject, the device comprising: a processor, a memory, a display and an input device, said memory containing processor executable instructions for carrying out the method of either the first or second aspect.

[0049] In some embodiments, the device may be a portable and / or handheld device. For example, the device may be a tablet PC, a smartphone, etc.

[0050] Further aspects of the invention provide the following: a computer program comprising code which, when executed on a computer, causes the computer to perform the method of the first or second aspect; a computer readable medium having recorded thereon a computer program comprising code which, when executed on a computer, causes the computer to perform the method of the first or second aspect; and a computer system programmed to perform the method of the first or second aspect. [Brief description of the drawings]

[0051] Embodiments of the invention will now be described, by way of example, with reference to the accompanying drawings, in which: [Figure 1] FIG. 1 is a diagram showing that a conventional neuropsychological test is considered a single entity, but is nevertheless made up of individual questions or items that can be asked outside of the conventional testing regime to obtain information about the subject's abilities. [Diagram 2] FIG. 1 is a diagram showing that the individual items that make up a standard neuropsychological test each have a range of difficulty levels. [Diagram 3] Graphs illustrating that disease symptoms and progression are not uniform across a subject's decline, with some becoming functionally impaired and others cognitively impaired before the subject's decline trajectory reaches a more severe disease stage. [Figure 4] FIG. 13 is a graph showing item difficulty scores β and γ associated with cognition and latent dementia traits, respectively, for MMSE, ADAS-cog, ADL, NPI and MADRS test items through 26 iterations of the expectation maximization algorithm discussed below. [Diagram 5]FIG. 13 shows the item difficulty scores γ and δ associated with dementia and depression latent traits, respectively, for the MMSE, ADAS-cog, ADL, NPI and MADRS test items through 26 iterations of the expectation maximization algorithm discussed below. [Figure 6] FIG. 13 is a graph showing item difficulty scores β and γ associated with cognition and latent dementia traits, respectively, for the MMSE, ADAS-cog, ADL, NPI and MADRS test items at the convergence of the expectation maximization algorithm discussed below. [Figure 7] FIG. 13 is a graph showing item difficulty scores γ and δ associated with dementia and depression latent traits, respectively, for the MMSE, ADAS-cog, ADL, NPI and MADRS test items at the convergence of the expectation maximization algorithm discussed below. [Figure 8] 13A-13D are graphs showing item difficulty scores β and γ associated with cognition and latent dementia traits, respectively, for items from the MMSE, ADAS-cog, ADL, NPI and MADRS tests at convergence of the expectation-maximization algorithm discussed below, which uses a reduced set of seed questions (6.8% of the full set of questions) to estimate initial conditions. [Figure 9] FIG. 13 shows the item difficulty score β for cognitive latent traits compared to the proportion of subjects responding correctly to questions for MMSE, ADAS-cog, ADL, NPI and MADRS test items at convergence. [Figure 10] FIG. 13 is a graph showing the item difficulty score γ for latent dementia traits compared to the proportion of subjects responding correctly to questions for MMSE, ADAS-cog, ADL, NPI and MADRS test items at convergence. [Figure 11] FIG. 13 shows the item difficulty score δ for the depression latent trait compared to the proportion of subjects who answered the questions correctly for the MMSE, ADAS-cog, ADL, NPI and MADRS test items at convergence. [Figure 12] 1 shows a flow diagram of the administration of adaptive neuropsychological testing. [Figure 13] 1 shows a flow diagram of the further implementation of adaptive neuropsychological testing. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0052] Detailed Description and Further Optional Features Aspects and embodiments of the present invention will now be discussed with reference to the accompanying drawings. Further aspects and embodiments will be apparent to those skilled in the art. All documents mentioned in this text are incorporated herein by reference.

[0053] Item response theory To quantify the severity of dementia in patients, there is a series of different neuropsychological tests, which can be aimed at diagnosing or monitoring the disease. These tests are delivered in the form of individual items of varying difficulty that can be answered by patients or caregivers to give an understanding of the severity of the patient's condition. This is as illustrated in Figures 1 and 2. Each item in these tests targets a specific clinical feature of dementia and a particular deficit related to the clinical feature being tested. The current scoring method of these tests typically includes the sum of question scores or some variation thereof, which means that the information obtained by individual items is not maximized.

[0054] The improvement provided by the present invention over current scoring methods is Item Response Theory (IRT). IRT allows both the individual items of a test and the patient to be scored according to their difficulty and the patient according to their ability, with respect to some underlying latent trait, in this case a specific clinical feature of dementia-related decline. The central concept of IRT is thereby the relationship between the way participants respond to a question and the latent trait that the question represents.

[0055] Item response theory is based on three important assumptions. The first, known as unidimensionality, is that all items in a given test are measures of exactly one latent trait. If more than one latent trait is assumed, multidimensional IRT is possible but computationally complex due to the increased number of parameters to fit from the data. In this case, the dimensionality of the data is similar to the number of latent traits measured by a given item. The number of latent traits is typically estimated using some form of factor analysis, such as principal components analysis. In general, factor analysis assumes a normal distribution, which is somewhat rare to obtain when the data collected are dichotomous in nature, and therefore factor analysis will typically fail in this special case. The second assumption is that each item in a test is locally independent, i.e., a correct or incorrect answer on one item should not automatically lead to the same answer to another question. Items in a test should only relate to each other through the latent traits they measure; if the items are measures of more than one latent trait (i.e., multidimensional), correlations may be valid across multiple latent traits. The third assumption is that subjects or participants answer questions in a way that indicates the latent trait being measured. More specifically, the probability that participants will answer a question correctly is a function of their latent ability for a given trait. Thus, the higher a participant's latent ability for a particular trait, the higher the probability that they will answer the question correctly. Typically, latent traits cannot be measured directly. To indirectly assess a participant's ability with respect to this latent trait, items that are indirect measures must be used. This theory may be more generally applied to measure many different kinds of latent traits, such as cognition and depression. Typically, the relationship between participants' latent trait abilities and the probability that they will answer an item correctly is modeled using a standard logistic function.

[0056] In previous applications of item response theory to neuropsychological test data collected from subjects with dementia, the test data were assumed to be unidimensional, meaning that the items included in the analysis were assumed to measure a single latent trait, and this latent trait was assumed to be a proxy for dementia or a single aspect of dementia.

[0057] However, as shown in Figure 3, clinical findings demonstrate that individuals exhibit different impairments at different stages and that disease symptoms and progression are not uniform. For example, some patients initially exhibit cognitive impairment but are still able to perform activities of daily living, while the opposite pattern may be evident in other patients. Some patients exhibit neuropsychiatric symptoms that are absent in others. This leads to the suggestion that dementia may be measured in a multidimensional space rather than a single unidimensional scale.

[0058] At a basic level, the multidimensional aspects of Alzheimer's disease and dementia are already recognized within the clinical space. This is most clearly demonstrated by the requirement from the US Food and Drug Administration and the European Medicines Agency that any proposed treatment of the disease must demonstrate efficacy in cognition as a co-primary endpoint, and also present clinical significance, generally interpreted as improving functional ability. Traditionally, separate tests have been used to measure each feature of interest. Scales such as ADAS-Cog have been designed to measure cognition, while scales such as ADL have been designed to measure function. Scores in one test (and therefore also in one feature) cannot provide any information about the subject's performance in the second test / feature, meaning that the interpretation of the score of each test is limited to only the single feature that the test was designed to measure. This also results in the definition of the feature being limited by the test used to measure it. For example, MMSE and ADAS-Cog are both considered tests of cognition. However, due to its brief nature, the MMSE only provides a measure of general cognition, whereas the more specific ADAS-Cog provides a measure of AD-related cognition.

[0059] The use of IRT allowed a subtle but important shift to occur, moving the focus from the test used to measure the trait to the trait itself. This not only allows the definition of the trait to remain independent of the test used to measure it, but also allows the test itself to be informed by multiple different traits. The question then arises as to which trait or trait is of interest. The traditional dichotomy of cognition and function remains useful, with a slightly expanded definition of function. Cognitive traits primarily reflect certain aspects of behavior: the ability to remember special terms or copy special shapes. Functional traits have traditionally been interpreted as activities of daily living, leading to tests designed to measure abilities such as making a meal or moving independently. However, the presence of neuropsychiatric symptoms is also important in diagnosis and clinical practice. Thus, the functional traits referred to as dementia traits therefore extend not only to tasks of daily living but also to neuropsychiatric symptoms, more in line with FDA and EMA guidance on clinical relevance. The focus on traits also allows a confound not previously discussed to be considered: depression. In classical testing methods, the presence of depression may cause a subject to perform lower than his / her true score in cognitive or functional tests. As a result, the subject is perceived as impaired when in fact the low score may be due to, for example, apathy. As mentioned above, detecting the presence of depression in classical testing methodology requires additional testing to be performed, and even then, it may not be possible or very accurate to quantify the level of the impact of depression on cognitive performance. IRT allows depression to be identified as a separate latent trait without requiring maximal testing to be performed and the confounding effect of depression to be accounted for.

[0060] When multidimensional IRT is applied to data from neuropsychological tests that measure many different aspects of dementia, heterogeneous data sets can be combined into a single analysis that captures progression across multiple domains of the disease.

[0061] general mathematical methods The multidimensional IRT algorithm used in the embodiments of the present invention is realized in the form of an expectation maximization (EM) algorithm, which iteratively applies a generalized linear model (GLM) to score questions and subjects. Initial conditions are estimated using the sum of the subjects' scores on seed questions that are assessed by psychologists to be good measures of latent traits that underlie the subject's decline in dementia, namely cognition, impairment (dementia) and depression. GLMs are used to analyze the subjects' binary responses to individual test items, first using the subject's scores on each hypothesized latent trait as covariates to score each individual item. The individual item scores on each hypothesized latent trait are then used to rescore the subject. An EM algorithm is used to apply these GLMs in an iterative fashion to locally maximize the likelihood of the final test and subject scores.

[0062] Since the data to be analyzed herein are binary or binarized, a special case of the binomial distribution - the Bernoulli distribution - can be used, and therefore a GLM with a link function of logistic regression analysis is used (P. McCullagh and JA Nelder. Generalized Linear Models. Chapman and Hall., 1952, Chapter 4.3, p. 108). As discussed above, the logistic regression model is the basis for other IRT models (e.g., the Rasch model, the two- and three-parameter logistic model, illustrated in B. Wright. A History of Social Science Measurement. Educational Measurement: Issues and Practice, Vol. 16 (No. 4), 1997), since it limits the probability outcome between 0 and 1.

[0063] Given a set of parameters θ, the expectation-maximization (EM) algorithm (disclosed, for example, in A. Dempster, N. Laird, and D. Rubin. Maximum likelihood from incomplete data via the EM algorithum. Journal of the Royal Statistical Society, Vol. 39, 1977) provides an iterative procedure for arriving at a locally maximized likelihood L(θ), a quantity for which it is not always possible to directly compute a maximum. At each iteration of the EM algorithm, an update for the unknown parameters θ is computed, where L(θ t+1 ) is L(θ t ) The expectation (E) step of the algorithm is calculated based on the most recent conditional parameter estimates θ of the system being studied, whether these are known parameters or observed data. t Given the model, we calculate the expected value of the likelihood based on the observed aspects of the statistical model. The maximization (M) step of the algorithm then optimizes the model using the updated parameters θ t+1 To compute , we maximize the likelihood with respect to the parameters estimated in the E-step. The EM algorithm then continues to iterate these two steps until the estimated parameters converge to some tolerance.

[0064] The EM algorithm can take the form of repeatedly applying a GLM to the data to find maximum likelihood estimates of the hidden parameters. This allows two dependent hidden parameters to be estimated simultaneously. Initial estimates are made for a first set of hidden parameters, and these initial estimates are used within the explanatory variables of the GLM to estimate a second set of hidden parameters. This estimated second set of hidden parameters can then be used as explanatory variables to estimate the first set of hidden parameters, reducing the dependency on the initial estimates. This process is then repeated until both estimates of the first and second sets of hidden parameters independently converge to some tolerance. If A and B are the first and second hidden sets of parameters, respectively, this follows the following algorithm: 1.A (0) is initialized to some reasonable estimate. 2. Next, estimate A included in GLM and explanatory variables (t) Using B (t) is estimated. 3. GLM and the previous estimate B included in the explanatory variables (t) Using A (t+1) is also estimated. 4.A (t) and B. (t) Steps 2 and 3 are repeated until the estimates for x, y, and z, respectively, converge.

[0065] Multidimensional IRT methods are known per se and estimate scores for individual items in multiple domains. However, the nature of neuropsychological data collected from patients is highly variable depending on the patient's "on the day" condition. These natural variations increase the complexity of data analysis. The primary concern in analyzing the data is to ensure that any algorithm implemented reaches a mathematically meaningful result; that is, that a global optimum is found in the likelihood landscape of the multidimensional space, and not just a local optimum. This means that traditional methods of multidimensional item response theory analysis may result in an inability to converge to a global optimum or a complete failure to converge, as is often the case with the particular data set exemplified herein.

[0066] Therefore, a modified form of item response theory is provided that allows initial conditions to be pre-specified to help guide the algorithm to a global optimum solution. This is accomplished by using a generalized linear model seeded with pre-specified initial conditions and contained in an expectation-maximization algorithm to score items and subjects from the training data set until convergence, as described above. The use of a priori knowledge of the question set in the form of initial conditions allows the algorithm to converge to a result that can be externally verified and is robust to changes in the initial conditions.

[0067] The model used to analyze the data is expressed as follows:

[0068]

number

[0069] where i is the subject at a particular visit and q is the question. Thus, α is the intercept and β q , γ q and δ qare the cognitive latent trait scores, the dementia latent trait scores, and the depression latent trait scores for each question, respectively. i , dem i and dep i are the cognition latent-trait score, dementia latent-trait score, and depression latent-trait score, respectively, for each subject at that particular visit(s).

[0070] To eliminate the dependency of question scores on initial question initialization, the expectation maximization algorithm discussed above was used, which works iteratively to ensure that the final results of item difficulty for each latent trait are data-driven.

[0071] The algorithm was constructed as follows: 1. For each subject, at each visit, the initial scores for the latent cognition, latent dementia, and latent depression traits (CoG i 0 ,dem i 0 and dep i 0 ) is calculated by counting the item responses associated with each category. These item responses are finally summed and normalized to be between 0 and 30. 2. Cog as a fixed covariate with α=0 i t ,dem i t and dep i t Use the to fit a generalized linear model to find β q t , γ q t and δ q t Request. a. Identify extreme outliers (those outside the interquartile range ±3 times the interquartile range) and set these extreme outliers to the max / min of the data within this range. b. Set the question score to 0 for a given latent trait if the standard error overlaps with 0. c. Take the absolute value of all the scores. 3.β as a fixed covariate with α=0 q t , γ q t and δ q t Use the following to fit a generalized linear model to find the cog i (t+1) ,dem i (t+1) and dep i (t+1) Request. a. Identify extreme outliers (those outside the interquartile range ±3 times the interquartile range) and set these extreme outliers to the max / min of the data within this range. b. Set the subject score to 0 for a given latent trait if the standard error overlaps 0. c. Take the absolute value of all subject scores. d. Normalize the scores to be between 0 and 30. 4. Repeat steps 2 and 3 until convergence.

[0072] As can be seen, certain steps of the algorithm discussed above may be omitted or modified while still resulting in an algorithm that can achieve the desired results. For example, the substeps of identifying extreme outliers, setting scores to zero, or using different normalization ranges in steps 2 and 3 may be omitted or their absolute values ​​may be modified. Similarly, the normalization of the initial patient scores in step 1 may also be omitted or modified. However, the general methods described in steps 1, 2, 3, and 4 must be applied.

[0073] In this specification, the algorithm is specifically described in the framework of application to binary data from subjects with dementia, and the specific data set analyzed in this specification is from subjects with mild Alzheimer's disease.However, this type of algorithm can be more generally applied across various forms of data, for example, non-binary data, by adapting GLM to fit different types of models.

[0074] Dataset and preprocessing The data used to create the embodiments of the present invention were collected from patients with diagnosed dementia, specifically patients diagnosed with Alzheimer's disease. However, the principles discussed herein are equally applicable to data collected from patients with, for example, behavioral variant frontotemporal dementia. The data are from clinical trials of drugs to inhibit the aggregation of tau protein in the brain. Dementia-like pathologies are associated with tangles of tau protein that inhibit cellular trafficking. Clinical trials are conducted in separate phases to test their safety and efficacy. The data analyzed herein are from phase 3 trials conducted to investigate the efficacy of drugs. Data were collected from various neuropsychological tests at various specific time points during the treatment of subjects.

[0075] By way of example, the tests analyzed herein include: 1. Mini-Mental State Examination (MMSE), cognitive assessment 2. Alzheimer's Disease Assessment Scale-Cognitive (ADAS-Cog), a Differential Assessment of Cognition 3. Alzheimer's Disease Collaborative Study - Activities of Daily Living (ADL), evaluation of the functionality of daily activities in patients with Alzheimer's disease 4. Neuropsychiatric Inventory (NPI), assessment of neuropsychiatric symptoms 5. Montgomery-Asberg Depression Rating Scale (MADRS), Depression Assessment It is.

[0076] Each test includes multiple items that together aim to provide insight into the severity of a patient's dementia. A scoring system is discussed herein that can describe each of these tests in the same space, i.e., mathematically describe the independent tests and the items within those tests in the same n-dimensional space according to the latent traits that they measure.

[0077] The data collected from the specific neuropsychological tests investigated for the purposes of discussion herein may be in the form of dichotomous or polytomous data. For standardization purposes, all data is processed or cut to ensure that it is binary and that a "1" response indicates a positive response and a "0" response indicates a negative response. To ensure that all data is binary, a divide-by-cut method is employed. Following the example of the ADL test:

[0078] "Regarding eating: Which best describes the subject's usual performance over the past 4 weeks? 3- Eating without physical assistance and using a knife 2- I used a fork or spoon to eat, but not a knife 1- I used my fingers to eat 0 - Subject usually or always had someone else feed them

[0079] The possible responses to this item are 0, 1, 2, and 3. To make the item dichotomous, we split it into multiple items, each with a binary type of response.

[0080] "On eating: Do subjects score higher than 0? On eating: Do subjects score higher than 1? On eating: Do subjects score higher than 2?”

[0081] This method was applied to all multi-choice questions to ensure that all data collected were dichotomous.

[0082] The second point is to ensure that all positive responses are coded as "1" and all negative responses as "0". As the neuropsychological tests of interest are all relevant to the diagnosis of dementia, negative responses are indicative of the disease, and for more general application this can be adapted to broadly code "1" as positive and "0" as negative. Responses to all items were oriented to reflect this. Of course, it can be seen that the reverse situation could also apply, i.e. negative responses are coded "1" and positive responses are coded "0". In this case, the higher the score, the more likely the subject is to have dementia.

[0083] result The method discussed herein for scoring items according to underlying latent traits was accomplished using the expectation-maximization algorithm and generalized linear models described in the General Mathematical Methods section.

[0084] Running this algorithm involves calculating the item scores β q , γ q and δ q Figure 4 shows the item difficulty scores, β, associated with cognition and latent dementia traits, respectively, for items from the MMSE, ADAS-cog, ADL, NPI, and MADRS tests, after 26 iterations of the expectation maximization algorithm described above. q and γ q Figure 5 shows the item scores γ for the latent traits of dementia and depression for the above test. q and δ q In early iterations, the difficulty score for each question is q , γ q and δ qis subject to some variation, which occurs as a result of the uncontrolled nature of the algorithm. i ,dem i and dep i Apart from the initial input of approximate scores for β, the algorithm is data-driven. Variation in item scores between iterations is expected, justifying the use of an expectation-maximization algorithm. This is because the item scores β in later iterations q , γ q and δ q is the initial subject score calculated from the seed questions, cog i 0 ,dem i 0 and dep i 0 However, in later iterations, the change between iterations is minimal, and item scores β q , γ q and δ q Convergence of is achieved. As discussed above, the expectation maximization algorithm converges to the maximum likelihood estimates of the parameters. Since convergence is achieved at this point, it can be said that the expectation maximization algorithm discussed above has yielded locally likelihood-maximized item scores per question.

[0085] In Figure 6, the final item score β q and γ q are shown. The MMSE and ADAS-cog tests vary primarily along the beta or cognitive latent trait axis; the ADL tests vary along both the cognitive (beta) and gamma or dementia latent trait axes, and the NPI varies along the dementia (gamma) latent trait. Figure 7 shows the final item score gamma q and δ q , indicating that the MADRS is broadly a measure of depression trait, δ only, whereas the NPI appears to be a measure of both depression (δ) and dementia (γ) latent traits.

[0086] Robustness Analysis initial conditions The results shown in Figures 6 and 7 are i 0 ,dem i 0 and dep i 0 This is achieved when the items used to calculate the latent trait score are chosen by the psychologist as good measures of the underlying trait. The subject scores for these questions are then summed and normalized to give the algorithm an initial first estimate of the subject's latent trait score. Figures 6 and 7 show the results when 68% of the scored items are used to calculate the a priori subject score in this manner. The robustness of the algorithm was examined by randomly reducing the number of these initial questions, first to 34% and then to 6.8% of the total set of items used. Including fewer seed items reduces the initial information used to score the items. Using a large number of items can be inherently problematic because, although there can never be too many seed questions because the algorithm is completely non-prescriptive, a large proportion of items that can be initialized are not always available. However, initializing with too few items can result in an algorithm that is numerically unstable and therefore cannot accurately score further items according to the latent trait.

[0087] Initial conditions are devised by grouping together questions that are likely to be answered in the same way, and devising a simple score from these items for each subject at each of their visits. This means that individual items are not assigned to latent traits from which they cannot deviate, but are used to contribute to a broader view of what each latent trait should look like. If too few seed items are used, it means that the algorithm will not be given a sufficient concept of the latent traits, which will prevent it from producing the same results as if more items had been included a priori.

[0088] Initializing 34% of the complete set of items and calculating question scores using the same methodology resulted in a median percent change in individual item scores of 2.9% in the cognitive domain, 2.2% in the dementia domain, and 1.8% in the depression domain. Repeating this procedure using 6.8% of the initial seed questions resulted in a median percent change of 6.7% in the cognitive domain, 7.7% in the dementia domain, and 6.9% in the depression domain. Despite the large reduction in the number of initial questions used to calculate the initial patient score estimates, the percent change that occurs in the question scores is relatively small. These small percent changes do not affect the interpretation of the items because they are related to the underlying latent traits, as demonstrated in Figures 6 and 8, which show little difference between the initial initialization of 68% (Figure 6) and the initial initialization of 6.8% (Figure 8) of the complete set of questions for the estimation of initial conditions. Taken together, the above provides evidence that the algorithm is robust to different initial conditions.

[0089] Ordering item difficulty Individual items are scored in a similar manner to their difficulty level for cognitive, dementia and depression latent traits. The items with higher scores for a latent trait are less difficult for that latent trait. Figures 9, 10 and 11 show the difficulty scores of each item in the MMSE, ADAS-cog, ADL, NPI and MADRS tests - according to cognitive, dementia and depression latent traits - and the percentage of subjects who answered the questions correctly.

[0090] Figure 9 shows item difficulty scores for β, the cognitive latent trait, and the percentage of subjects who answered the questions correctly. Questions that a high percentage of subjects answered correctly are estimated to be of low difficulty, and vice versa. Figure 9 shows that there is a strong relationship between item scores and the percentage of subjects who answered questions correctly in the cognitive latent trait. This indicates that the items are well ordered by difficulty for the individual latent traits. The same results are present in Figures 10 and 11, which represent item difficulty for γ, the dementia latent trait, and δ, the depression latent trait, showing that the algorithm is able to correctly order items according to difficulty in multiple dimensions.

[0091] 3D model justification Taken together, the results presented provide evidence that the MMSE, ADAS-cog, ADL, NPI and MADRS are answered independently of each other; that is, three independent latent traits govern the way subjects answer the items of these tests. The independent dimensions are named cognition, dementia and depression latent traits. These latent traits can be most clearly identified when seed questions are used to drive the algorithm with some a priori knowledge of the latent traits to be discovered and an expectation maximization algorithm is used, but they are also present in item response theory without prior knowledge. The use of a three-dimensional model to represent the data is tested by using a likelihood ratio test to evaluate whether the three-dimensional model fits the data statistically significantly better than a reduced two-dimensional model. Sequential elimination of latent traits and comparison of the resulting two-dimensional model to the three-dimensional model resulted in 80% of subjects benefiting from the three-dimensional model when the depression dimension was removed. When the dementia latent feature was removed from the model, 60% of subjects benefited from the 3D model, and finally, when the cognitive latent feature was removed, 79% of subjects benefited from the 3D model. Taken together, these results suggest that at least 60% of subjects benefited from fitting a 3D model and justify the use of at least three latent features to characterize the progression of mild Alzheimer's disease.

[0092] The methods discussed above demonstrate that by characterizing multiple dementia test items, it is possible to probe specific latent traits by selecting items whose responses are (i) indicative of a particular latent trait; and (ii) of appropriate difficulty given prior responses.

[0093] conclusion In conclusion, what is presented here is evidence that the individual items of different neuropsychological tests can exist in the same multidimensional space. This space mathematically characterizes how well the individual items measure the underlying latent traits in subjects with mild Alzheimer's disease. The general algorithm is the application of expectation maximization to item response theory; it is realized here in the form of a generalized linear model.

[0094] Computer adaptive testing algorithms implemented in neuropsychological testing 12 shows a flow diagram of the administration of an adaptive neuropsychological test using the principles discussed above. The test begins at step 101, followed by the asking of one or more seed questions at step 102.

[0095] After the subject has answered one or more seed questions, a score for the latent trait being measured is calculated in step 103. To calculate the score, an item value lookup function 101 is used to access storage 109 or question ability or difficulty values. Details of this calculation are discussed below.

[0096] The score is then used (at step 104) to select one or more questions to be administered. Again, the question / item ability or difficulty value is accessed via item value lookup function 101. Details of the selection of one or more subsequent questions are discussed below.

[0097] After selecting one or more further questions, they are asked in step 105 and the results are recorded. These results are used to update the scores for the latent trait in step 106. A decision is then made as to whether a test completion criterion has been met (see decision box 107). The test completion criterion could be, for example, a minimum confidence value. Alternatively, the test completion criterion could be triggered manually by an administrator of the test or automatically when a predetermined number of questions have been asked.

[0098] If the test completion criteria has not been met, then the loop returns to step 104 and repeats the sequence discussed above. This loop from steps 104 to 105, then 106, and back to 104 defines an adaptive test loop.

[0099] If the test completion criteria is met, the loop moves to step 108 and the test ends. A confidence value is provided in the test result. The test result may be, for example, a dementia assessment score or a score value for the latent trait being measured.

[0100] Figure 13 shows a flow diagram of the further implementation of adaptive neuropsychological testing. The difference between the flow diagram in Figure 12 and the flow diagram in Figure 13 is that in Figure 13 the test is carried out on N latent traits / dimensions.

[0101] Thus, after the test begins in step 101, the N-dimension weight preferences for each dimension may be selected directly or may be predefined by the test. The weight preferences may in some instances be binary weightings, i.e., 0 indicates that the latent trait should not be measured and 1 indicates that it should be measured. Alternatively, the weight preferences may indicate to the administrator a preference for which latent traits should be measured. For example, a weight preference of 0.5 for dem and 1 for cog may indicate that the test administrator requires score values ​​for the cog latent trait and (if possible) would like score values ​​for the dem latent trait. The N-dimension weightings are sent via link 212 to module 210, discussed below.

[0102] After the N-dimension weight preferences are selected, testing begins at step 203 where one or more seed questions are asked. Then, at step 204, the N-dimension scores are calculated. To calculate the N-dimension scores, an item-by-item N-dimension value lookup function 213 operates and interacts with storage 211, which contains the N-dimension question ability / difficulty values.

[0103] Having calculated the initial values ​​for each of the N dimensions, the test uses these values ​​(at step 205) to select one or more questions. Here, the test utilizes the same functions 213 used to explore the values ​​of the N dimensions for each question, but the selection is modified or guided based on the N-dimension weightings previously set by module 210. For example, if a weight preference of 0 is set for dem, then questions related to dementia latent traits will not be selected.

[0104] Once the questions have been selected, they are asked and the answers recorded in step 206. The answers to these questions are used to update the N-dimensional scores in step 207. This update step also uses function 213 to retrieve the N-dimensional question ability / difficulty value.

[0105] After updating the scores, the test determines whether the test criteria have been met (see decision box 208). The test completion criteria may be, for example, a minimum confidence value. Alternatively, the test completion criteria may be manually triggered by an administrator of the test, or automatically triggered when a predetermined number of questions have been asked.

[0106] If the test completion criteria have not been met, the test returns to step 205 and enters an adaptive test loop.

[0107] If the test completion criteria are met, the test moves to step 209 where the test is terminated and N-dimensional scores and confidence values ​​are provided.

[0108] Calculation of subject scores Given the subject's answers to the questions asked, the maximum likelihood estimates are used to calculate the subject's score and corresponding confidence interval. The likelihood estimates for a subject with a given score are calculated, and the ability score of the subject with the maximum likelihood is selected. The likelihood estimates for the questions, r teeth,

[0109]

number

[0110] or vice versa if the subject's answer is incorrect.

[0111]

number

[0112] where q and e are both variables related to the difficulty of a given question, and x is the ability of the subject. The likelihood for each ability x of the subject is

[0113]

number

[0114] Lx is calculated as ≈ 0.9889 and summed over all questions answered for each x. The subject's performance x with the highest Lx is the maximum likelihood estimate of the subject's performance. Confidence intervals are calculated based on one standard deviation of the maximum likelihood estimate of the subject's score, which corresponds to a 68% confidence interval. The 68% chi-square test has a threshold of 0.9889, and 2 x log-likelihood is mathematically similar to the chi-square test. Thus, the confidence interval is

[0115]

number

[0116] The range of x that fits this range in Lx is the confidence interval.

[0117] The above method can be applied simultaneously to multiple dimensions to calculate a subject's ability score for multiple latent traits. Such calculations can be performed on a specific trait, q d and e d where d is the latent trait / dimension being measured and, correspondingly, the subject's ability x d is specific to the latent trait / dimension being measured.

[0118] Calculating the next question To select the next question that best fits the subject's ability range, we use the information content of each question at a given ability score x, I qx but,

[0119]

number

[0120] It is calculated as:

[0121] An iterative procedure then begins to select one or more next questions. This is accomplished by calculating the information content of the most recent proposed set of next questions. The information content of this set of next questions is calculated by multiplying the information content I by 1 over all questions in the set of proposed next questions (one is added each iteration) for each patient's ability within the appropriate range. qx The appropriate range is calculated to be between the lower confidence interval −2, lowerCl−2, and the upper confidence interval +2, upperCl+2.

[0122]

number

[0123] where n is the number of questions in the set of proposed next questions.

[0124] The information content for each possible eligible question for the proposed next question set is then calculated by summing the information content for each question over all eligible subject capabilities x, and weighting it by the information content of the most recent proposed next question set at each x.

[0125]

number

[0126] Next, the largest amount of information I qis added to the set of proposed next questions and the iterative process begins again. The process continues until either a maximum set of next questions has been selected (e.g., n=1, or n=5), the list of eligible questions is empty, or the maximum information content of any question is zero.

[0127] As described above, the method generates a question score q specific to the latent trait. d and e d , and the subject's ability score x d The same approach can be applied to multiple latent traits by using EXAMPLES

[0128] Subjects answer example seed questions in this manner (questions are coded for ease of use):

[0129]

number

[0130] While the present invention has been described in conjunction with the exemplary embodiments above, many equivalent modifications and variations will be apparent to those skilled in the art given this disclosure. Accordingly, the exemplary embodiments of the present invention described above are considered to be illustrative and not limiting. Various modifications to the described embodiments may be made without departing from the spirit and scope of the invention.

Claims

1. 1. A computer-implemented method for adaptively testing a neurological condition of a subject, comprising: a display displaying one or more seed questions; Mediating an input device to receive answers to the one or more seed questions; a processor calculating a score value of a latent trait of the subject from the received answers to the one or more seed questions; An adaptive inspection loop comprising: (a) the processor selecting one or more further questions from a database of questions based on the score value; (b) the display displaying the one or more further questions; (c) the processor updating the score value based on received answers to the one or more further questions; (d) the processor determining whether a test completion criterion has been met; entering an adaptive inspection loop, and repeating steps (a)-(d) in sequence until a test completion criterion is met, and providing a result of the test based on the score value indicative of the neurological status of the subject, each of the additional questions having an associated difficulty rating and an indication of relevance to at least one of the one or more latent traits of the subject; and wherein the processor's selecting the one or more additional questions is further based on one or both of the associated difficulty ratings and an indication of relevance to one or more latent traits of the subject. calculating a score value for each of a plurality of latent traits of the subject from the answer(s) to the one or more seed questions; and wherein the processor selects the further question based on at least one of these score values; A computer-implemented method comprising an initial step of the processor determining a weighting value for each of a plurality of latent traits of the subject.

2. The computer-implemented method of claim 1 , wherein the score value is a latent trait ability score.

3. The computer-implemented method of claim 1 or 2, wherein the processor's selection of the one or more further questions is further based on a weighting value for at least one of a plurality of latent traits of the subject.

4. The computer-implemented method of claim 1 , wherein the latent traits of the one or more subjects include at least one of a cognitive trait, a dementia trait, and a depression trait.

5. 5. The computer-implemented method of claim 1, wherein the score for the or each of the latent traits of the subject is calculated using an expectation maximization algorithm and / or a generalized linear model.

6. The generalized linear model has the following formula: [0010] where Yi,q is the response by subject i at a particular visit, q is the question, α is the intercept, βq, γq and δq are difficulty scores for the questions related to the cognitive, dementia and depression traits, respectively, and cogi, demi and depi are the subject's ability scores related to the cognitive, dementia and depression traits, respectively.

7. 7. The computer-implemented method of claim 1, wherein the seed questions and / or the further questions are selected from any one or more of the following: Alzheimer's Disease Assessment Scale-Cognitive Function Test, Alzheimer's Disease Cooperative Study-Activities of Daily Living Test, Neuropsychiatric Inventory Test, Montgomery-Asberg Depression Scale Test, and Mini-Mental State Examination.

8. 8. The computer-implemented method of claim 1, wherein the seed question and / or the further question are used to calculate and update a score value for each of a plurality of latent traits of the subject.

9. The processor's selection of one or more further questions is further based on an information amount for each question in the database of questions, the information amount Iqx for each question at a given ability score x being calculated according to the following formula: [0025] The computer-implemented method of any one of claims 1 to 8, wherein q and e are both variables related to the difficulty of each question.

10. 10. The computer-implemented method of claim 9, wherein the processor calculates the information content for each question in the database of questions for the subject based on score values ​​of the subject's latent traits and selects one or more questions from the database of questions having the greatest information content.

11. 1. A device for performing adaptive testing of a psychological state of a subject, comprising: A processor, a memory, a display and an input device, A device, the memory containing processor-executable instructions for carrying out the method of any one of claims 1 to 10.

Images