Automatic migraine classification via feature selection committee and machine learning techniques over imaging and questionnaire data

DOI: https://doi.org/10.1186/s12911- 017- 0434- 4

Abstract-Summary Feature selection methods are commonly used to identify subsets of relevant fea- tures to facilitate the construction of models for classification, yet little is known about how feature selection methods perform in diffusion tensor images (DTIs).

Feature selection and machine learning classification methods were tested for the purpose of automating diagnosis of migraines using both DTIs and questionnaire answers related to emotion and cognition  – factors that influence of pain perceptions.

The DTI images and test results were then introduced into feature selection algo- rithms (Gradient Tree Boosting, L1-based, Random Forest and Univariate) to reduce features of the first dataset and classification algorithms (SVM (Support Vector Machine), Boosting (Adaboost) and Naive Bayes) to perform a classification of migraine group.

We implement a committee method to improve the classification accuracy based

on feature selection algorithms.

When classifying the migraine group, the greatest improvements in accuracy

were made using the proposed committee-based feature selection method.

The proposed feature selection committee method improved the performance of migraine diagnosis classifiers compared to individual feature selection methods, producing a robust system that achieved over 90% accuracy in all classifiers.

Background As has been stated, there is an internal classification for the feature selection method, although to obtain quantitative data to establish whether the selection made using the method is correct, we use an external classification with the following classifi- ers: SVM (Support Vector Machine) [45, 46], Boosting (Adaboost) [47] and Naive Bayes [48].

The features used in this study to select them and subsequent classification of subjects to obtain via different questionnaires, which focus on items related to migraine pathologies and magnetic resonance images with diffusion tensor.

To undertake the classification, we obtain values taken from DTI (Diffusion Tensor Imaging) images in addition to selecting specific questionnaires about migraine [49].

The main goal of this study is to automate the diagnosis of patients with migraine through the classification of features obtained from DTI images and psychological tests applied to patients.

466

3 Diagnosis

Methods Once the algorithms complete the classification, the fit method is then applied to adjust the model to the classification, followed by the transform method, which enables to confine the entry reduce feature set to the most important ones.

The algorithm apply fit transform method to this classification, which first adjusts the classification model and then reduces the entry reduce feature set according to the most important ones in terms of this classification.

The selection of “best K” method for this case study from among the univariate feature selections available, whereby the user may select the exact number of fea- tures with which they wish to subsequently carry out classification by disregarding the others that form part of the initial set of data.

To check whether the feature selection has been effective, a classification carries

out with the new set of data obtained via each feature selection method.

Results Feature selection will then be carried out using each of the 4 methods, obtaining a new reduce dataset for each of them.

We obtaine the worst classification using the univariate feature selection method and subsequent classification using Naive Bayes with a 60% accuracy (precision 53%, recall 87% and F1score 66%), while the best classification is the gradient tree boosting - again using the Naive Bayes classification and with 98% accuracy (preci- sion 96%, recall 100% and F1score 98%).

The gradient tree boosting method improved the classification percentage by 8%

with the features selected as being the most important.

We perform a method committee obtaining the most important features from each feature selection method, i.e. the features most frequently found in the result- ing dataset will be the ones selected to carry out a new classification.

Discussion Features were selected in this study to carry out a better classification when differ- entiating between subject who suffer from sporadic and chronic migraine patholo- gies and medication overuse.

From an initial full set comprising all the subjects and all the features belonging both to psychological questionnaires, data regarding days with pain, amount of painkillers and values obtained from DTI images, we obtained 90% classification accuracy between the three groups subject to study in the case of the SVM classifier, 93% in the case of Boosting and 67% in the case of Naive Bayes.

There is no study in which machine learning techniques have been applied to classify the migraine pathology, although they applied in other studies to other pathologies, e.g. in the study conducted by Dyrba, in which the used machine learn- ing techniques  – specifically the SVM classifier for classification of subjects via DTI images belonging to Alzheimer’s disease [50].

Conclusion The classification ratio was improved by 28% in the case of the Naive Bayes clas- sification due to this feature selection.

3.1 Classification

467

The method can classify patients due to a small set of features from specific questionnaires related to emotion and cognition, combined with features obtained from DTI images resulting from prior selection of the most important features.

Acknowledgement A machine generated summary based on the work of Garcia-Chimeno, Yolanda; Garcia-Zapirain, Begonya; Gomez-Beldarrain, Marian; Fernandez-Ruanova, Begonya; Garcia-Monco, Juan Carlos 2017  in BMC Medical Informatics and Decision Making.

Machine learning-based automated classification of headache disorders using patient-reported questionnaires