A Machine Learning Approach to Design an Efficient Selective Screening of Mild Cognitive Impairment

The purpose of this video is to describe how to design a cost-effective screening of mild cognitive impairment, which involves more difficulties than expected to perform low cognitive skills mainly related to memory and language. We want to stress the importance of this kind of condition affecting the elderly, and the versatility of this procedure. This procedure could be easily adapted to other diseases, converting a costly screening into a feasible one.

Search Cochrane Systematic Reviews for terms of the condition to screen. For instance, in our study, cognitive impairment or dementia"joined with risk factors. Search on PubMed Terms presents some evidence of the relationship with cognitive deterioration or dementia.

Select the variables with more evidence of association with cognitive deterioration or dementia to elaborate a questionnaire. A thorough search including terms as cognitive impairment"and risk factors"was conducted by using PubMed and the Cochrane Systematic Reviews. Before starting the study with the purpose of compiling the greatest number of factors that appear in the scientific literature.

As possible characteristics related to cognitive impairment, specifically in the Cochrane Database, it was heard for all articles with the term cognitive impairment"or dementia, and in PubMed articles with the terms of sex, education level, cognitive activities, physical activity, diabetes, hypertension, cholesterol, depression, obesity, smoking, alcohol, sleep, diet, and economic conditions. In combination with cognitive impairment"or dementia, too. Hence, a questionnaire was elaborated with the variables at the bibliographic review that present some evidence of the relationship with cognitive deterioration or dementia.

A questionnaire is designed with information of the selected associated variables. For our performance study, the chosen variables are demographic lifestyle and chronic pathologies in addition to the presence or absence of depression, which is frequently associated with cognitive deterioration. The consumption of drugs was collected following the methodology.

The Anatomical Therapeutic Chemical Code was used to classify the drugs. The screening test selected could be administered by staff of primary health such as pharmacists. Basic characteristics necessary to perform the test were taking into account.

In particular, for Spanish elderly population with large amount of illiterates who lived in the Spanish Civil War, our proposal was using Short Portable Mental State Questionnaire of favor and Mini-Mental State Examination Both are widely used in memory clinic, since one of them requires literacy. People recruited were non-institutionalized patients age 65 years or older who went regularly to the pharmacy and wished to participate in the present study. People who had any difficulty to perform evaluation test or people who are in treatment for dementia are excluded.

Study participants are considered cognitive impaired when at least one of the following criteria is met. Data scores in Short Portable Mental State Questionnaire are four or more points in the case of illiterate participants and three or more points for other subjects. A less than or equal to 24 points in the corrected Mini-Mental State Examination test.

Cognitive-impaired participants are referred to a medical specialist, as a neurologist, for their clinical diagnosis. Pharmacist researchers are trained in basic knowledge about cognitive impairment and in the management of screening tools. Our cross-sectional study to detect cognitive impairment and potential society-derived factors in population age 65 years and older is design.

Estimated sample size for a prevalence of cognitive impairment is 541 people, and with an increase of 10%due to losses, are 600 people. Communication letters have been designed between the healthcare network informing about the project. The definitive diagnosis through a specific test is reserved for specialized care, following the protocol shown in the diagram.

This is a proprietary step before applying machine-learning techniques, transforming data according to the p-test of the algorithms to be applied. The algorithms to generate decision trees are to change of variability or correlation across variables. So the focus is on categorizing variables.

For instance, generating variables to classify whether or not the patient is taking a drug according to second and third level of the Anatomical Therapeutic Chemical Classification Code, depending on the pharmacotherapy follow-up sheet. A logistic regression analysis was performed for each variable to evaluate if it is significant enough to be included in the data set to generate the decision tree. In order to get the max amount accuracy in in possible cognitive decline, several machine-learning techniques have been assemble.

And finally, we have develop a model with an 80%of our core and a tree model based on our recursive petition algorithm to develop a decision tree to get the most significant variables in a screening. The machine-learning algorithm applied to the training data set, which consists in a 80%of the whole data set. The remaining part is used to estimate the accuracy of the model.

Data set is expected to be in balance, and downsampling is one of the techniques to face this problem. For screening, we're very interested in reducing the number of false negatives as much as possible. This can be achieved by means of a upper PA selection of the loss matrix.

The optimal parameter of the algorithm was selected with a cross-validation. Our cross-sectional study was conducted with 728 non-institutionalized participants older than 65. One-hundred and 27 participants score positive on mild cognitive impairment tests.

Participants classified as positive were referred to clinical diagnosis. After performing the research study to estimate the percentage of users with mild cognitive impairment, a new variate logistics regression is performed with all the variables with the purpose of selecting variables. For some of the more significant variable, a 99%confidence interval of the odds ratio is displayed in this error bar graph.

It is a common intergraphical representation for the confidence interval of the odds ratio, where low scale is used for odds ratio. All these variables whose p-value is larger than 0.01 are selected to generate a wide box model based on decision tree, while many other variables with a higher p-value were not selected to generate the model. For instance, for these plain variables, the 99%confidence interval of the odds ratio is included in the value one.

And hence, the p-value is higher than 0.01. After this pre-processing, we have split the data set into training data test and the test data set. The decision tree is generated in the training data set as input, which consists in 583 individuals, and validated with the test set with 145 users.

The performance of the algorithms has been assessed by means of the area under the ROC curve in the test set. After using car Library in R, for each user, the resulting tree assigned a probability and a recommendation of whether or not the user should take the mild cognitive impairment test. They are depending on the final node in the tree.

The value at the bottom of the box is the percentage of individuals with these characteristics in the training set. The warmer is the color of the box, the more likely to be positive in mild cognitive impairment tests. Pay attention that the top node corresponds to the question of the absence of memory complaint.

Here, a positive answer hails to the left branch, being the following question the user's sex, whereas a negative answer implies to go to the right branch and to ask about user's sleeping time per day. To evaluate the full casting capability of the decision tree, the ROC curve is displayed. It area under ROC curve is.0.763.

The recommendation of the tree about taking the mild cognitive impairment test achieved a sensitivity. of 0.76 and a specificity of 0.7, represented with a blue point in the figure. As a result, a short interview to select users in risk of mild cognitive impairment with the tool of a decision tree produces significantly the number of users taking mild cognitive impairment tests, which is quite time consuming for the tester.

This reduction can be estimated in the test set, interpreting the confusion matrix of the observed and predicted classes. Indeed, 55 out of 145 users in the test set are selected by the decision tree, reducing a 62%of users taking tests, whereas most of the users being positive in mild cognitive impairment are selected, namely 19 out of 25. As a conclusion of this study, given a mild cognitive impairment screening whose prevalence is low, 17%such as the considered for the research study, it is possible to design a set of adequate selection criteria by means of machine-learning techniques, increasing the percentage of positive in mild cognitive impairment up to more than 30%among selected users.

Consequently, these tools have asked to be more efficient in screening with a substantial cost reduction. Data-driven models with benefit from understanding what is the most important information in order to construct a reduced model. The construction of a decision tree give us an insight on which variables we should put a focus, in order to discriminate in a cost-effective way people who are or not recommended taking the mild cognitive impairment tests.

the protocol design is time consuming. Because of that, some other tests may be considered for future detection of mild cognitive impairment in just a few minutes. Moreover, we have decide to serve our screening study at the age of 50 instead of 65 as a preventive task in order to increase the effectiveness in the chain of mild cognitive impairment.

Pharmacists are one of the most accessibles and regularly visited healthcare professionals, and they can play a vital role in early detection of mild cognitive impairment.