Machine Learning Models for Classification of Migraine Headaches

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Mar 7 2024

In a paper published in the journal Scientific Reports, researchers tackled the challenges of accurately diagnosing migraine headaches using machine learning (ML) techniques. By training state-of-the-art ML algorithms on publicly available datasets, the study demonstrated the effectiveness of models such as support vector machine (SVM), k-nearest neighbors (KNN), random forest (RF), decision tree (DST), and deep neural networks (DNN) in classifying seven different types of migraines.

With data augmentation, the DNN achieved the highest accuracy of 99.66%, highlighting the transformative potential of artificial intelligence (AI) in improving migraine diagnosis and healthcare outcomes.

Related Work

Past research has highlighted the prevalence of migraine headaches, a common neurological condition affecting a significant portion of the population. Despite not posing a severe threat to life, migraines can substantially impair work performance, quality of life, and overall well-being. The International Classification of Headache Disorders (ICHD) categorizes headaches into primary and secondary types, with migraines being one of the primary headache disorders.

Various types of migraines include those with aura, those without aura, and chronic migraines. Symptoms such as vomiting, nausea, sensitivity to light and noise, and throbbing headaches characterize these migraines. Identifying migraine triggers, such as stress, caffeine, and hormonal changes, is crucial for managing the condition. However, traditional diagnostic methods like MRI, PET, and CT scans are costly and require specialized expertise, posing challenges, particularly in developing countries.

Migraine Classification: ML Approach

Processing data for ML or deep learning (DL) methods is time-consuming and resource-intensive. Therefore, a robust machine learning network is necessary to collect relevant insights effectively. Developing such a framework presents challenges, and researchers achieve personalization by actively adjusting the parameters of ML classifiers. This study utilized various machine learning algorithms, including SVM, KNN, DT, RF, and DNN, with modified parameters tailored to the input dataset to enhance classification accuracy.

Data pre-processing is essential to improving the performance of machine learning models. It involves eliminating noisy and inconsistent data, identifying errors, and converting data into numerical variables. Additionally, data augmentation techniques, such as the synthetic minority oversampling technique (SMOTE), were employed to increase the volume of the dataset and address imbalances. SMOTE generates synthetic examples of minority class instances to create a more balanced dataset, thus enhancing model training.

The dataset used in this study consisted of medical records from 400 individuals diagnosed with various migraine-related diseases. Expert medical personnel collected the data, focusing solely on clinical presentation while ensuring patient anonymity. After data augmentation, the dataset size increased to 1447 records. Training and testing datasets were then separated, with 1157 records allocated for training and the remaining for testing. The initial dataset exhibited significant imbalances, rectified through data augmentation, resulting in a balanced distribution across migraine classes.

Following pre-processing and data augmentation, various machine learning classifiers, including SVM, KNN, DT, RF, and DNN models, were applied to classify migraine cases. The deep neural network architecture consisted of input, two hidden layers, and classification layers designed explicitly for migraine classification tasks. The study aimed to accurately classify different migraine types based on patient data through these classification models.

Migraine Dataset Analysis Summary

The experiments section details a comprehensive set of experiments conducted on the dataset to validate the effectiveness of data augmentations. Additionally, researchers present accuracy and loss graphs for the DNN model with data augmentation, visually representing the model's performance. The study utilized a corpus of 400 clinical records of patients with various pathologies related to migraines.

Following data augmentation techniques, the dataset size increased to 1447 patient records. These records were collected by skilled medical staff at the Hospital Materno Infantil de Soledad, Colombia, ensuring anonymity and focusing solely on clinical presentation. Researchers conducted the experiments before and after data augmentation, employing various ML classifiers such as KNN, SVM, RF, DST, and DNN models.

The dataset encompassed variables related to patient symptoms and diagnoses, with identifiable data excluded from the study. Symptoms experienced during headaches, such as age, family background, dizziness, vomiting, and others, were considered when selecting variables alongside those related to migraine diagnosis. The classification accuracy of the DNN model was validated against annotations by domain experts, with the highest accuracy achieved by the DNN model with data augmentation, reaching 99.6%.

The classification report presents the performance of all utilized algorithms with and without data augmentation, highlighting the performance improvement achieved through augmentation techniques. Furthermore, researchers compare classification results from previous studies with those obtained from the proposed model with data augmentation. The study also applied various preprocessing methods and machine learning models using the Pandas library for data cleaning and outlier removal. Despite the proliferation of automatic classification techniques based on machine learning in recent years, the deep neural network model outperformed traditional classifiers, achieving an accuracy score of 99.66% on the migraine dataset.

Conclusion

To sum up, considering the prevalence of migraine and the scarcity of medical resources, ML models offered a promising solution for classification and prediction. This study used various performance metrics to implement and evaluate four traditional machine learning algorithms (SVM, KNN, RF, and DST). Following preprocessing, researchers trained these algorithms on a publicly available corpus to classify migraine into seven basic types. Results indicated that the DNN outperformed other traditional models significantly.