Evaluating the Role of Machine Learning in Migraine Detection and Classification ^†

Abstract

Migraine is a common neurological illness that has a major influence on the quality of life; yet, precise categorization and prediction remain difficult because of its complicated symptoms and multiple triggers. This work investigates the use of advanced machine learning (ML) algorithms to improve migraine diagnosis and prediction, drawing on a large dataset that includes clinical, lifestyle, and environmental aspects. Various machine learning models, such as ensemble methods, deep learning, and hybrid approaches, are tested to see how well they discriminate migraine from other headache conditions and predict migraine episodes. Feature selection approaches are used to identify the most important predictors, which improve model interpretability and performance. Experimental results show that the proposed machine learning framework outperforms established diagnostic methods in terms of classification accuracy, sensitivity, and specificity. The study demonstrates how ML-driven solutions may be used to manage migraines in a tailored way, helping medical practitioners with early diagnosis and intervention techniques. My suggested framework, NeuroVote(ensemble model), offers a remarkable 99.99% classification accuracy for migraines. Future studies will concentrate on optimizing models for clinical deployment and incorporating real-time data from wearable technology.

1. Introduction

Millions of individuals worldwide suffer from migraines, a complex neurological disorder. It is characterized by recurrent episodes of moderate to severe headaches, often with light and sound sensitivity, nausea, and vomiting. Migraines, as opposed to regular headaches, can be incapacitating and chronic, significantly impairing daily functioning and productivity. Although the exact cause of migraines is still unclear, a combination of genetic, environmental, and neurological factors are believed to be at play. It can be difficult to correctly diagnose and categorize migraines because of the wide range of symptoms and triggers. By automating the diagnostic process, detecting migraine subtypes, and forecasting possible triggers based on patient data, recent developments in machine learning (ML) provide encouraging options. In order to provide a data-driven strategy for better understanding and managing this neurological condition, this study intends to investigate the potential of machine learning (ML) in enhancing migraine classification and detection. Over 1 billion individuals worldwide suffer from migraines, making it one of the most prevalent neurological conditions. The World Health Organization (WHO) reports that migraines are the sixth most incapacitating and third most common ailment in the world. Due to hormonal factors, migraines are three times more common in women than in men, indicating that women are far more impacted than males. Even though migraines are common, they are still not well understood or adequately treated, particularly in underdeveloped nations with limited access to professional medical care. Migraines can have a significant financial impact, costing billions of dollars a year in lost productivity at work and medical expenses. Effective management of migraines depends on early diagnosis and accurate categorization, and machine learning models present a viable way to close the diagnostic and treatment gap. Numerous variables, including genetic predisposition, lifestyle choices, and environmental conditions, can cause migraines. Given that family history is a powerful risk factor and that research indicates a hereditary component to migraine susceptibility, genetics is important. Variations in estrogen levels during menstruation, pregnancy, or menopause can cause attacks, especially in women. Hormonal changes also play a role. Another important consideration is dietary practices, since migraines have been related to the use of processed foods, alcohol, caffeine, and artificial sweeteners. It is also recognized that inadequate or extreme sleep disturbances raise the risk of migraines. Environmental elements that can also serve as triggers include stress, bright lights, loud noises, and abrupt changes in the weather. A variety of symptoms, varying in strength and duration, are prevalent in migraines. A common symptom is a headache that throbs or pulses, usually on one side of the head. In addition, many individuals have increased sensitivity to light (photophobia) and sound (phonophobia), as well as nausea and vomiting that might get worse with movement. Aura, or visual disturbances, can also happen, such as zigzag lines, flashing lights, or momentary blindness. Additional symptoms like exhaustion, lightheadedness, and cognitive impairments—often called “brain fog”—can have a big influence on day-to-day functioning. Effective migraine management requires an understanding of these triggers and symptoms, and machine learning techniques present a promising way to do so by analyzing big datasets to find hidden patterns and provide early warnings, which will ultimately improve diagnosis and treatment. Since migraines are a complicated neurological disorder with many different manifestations, proper diagnosis and classification are essential to successful therapy. Although headaches, nausea, and light and sound sensitivity are common symptoms of all migraines, each person’s migraine can differ greatly in intensity, duration, and concomitant symptoms. While some migraines come on quickly with no warning symptoms, others have warning signals like aura or visual disturbances. While some are brought on by certain physiological or environmental circumstances, others are more chronic and incapacitating. To create focused management methods, patients and medical practitioners must have a thorough understanding of the various forms of migraines. Technological developments, especially in machine learning, have made it feasible to examine patient data to more precisely categorize migraines and spot trends that could indicate when they will occur.

1.1. Types of Migraine

The following section explores the various types of migraines, their unique characteristics, and how they impact individuals differently.

1.1.1. Migraine Without Aura (Common Migraine)

The majority of migraineurs experience this type of migraine, which is the most common. Without any warning symptoms (aura) beforehand, it is characterized by a strong, pulsating headache that lasts for 4 to 72 h. Commonly reported symptoms include light and sound sensitivity, nausea, and vomiting.

1.1.2. Migraine with Aura (Classic Migraine)

This form of migraine attack is preceded by neurological problems known as aura. Aura can manifest as visual distortions, blind spots, flashing lights, or sensory alterations like tingling or numbness in the hands or face. These symptoms normally last 10 to 60 min, followed by a headache phase.

1.1.3. Sporadic Hemiplegic Migraine

A rare form of migraine known as sporadic hemiplegic migraine (SHM) is characterized by temporary paralysis on one side of the body in addition to typical migraine symptoms such as intense headache, nausea, and photosensitivity. Unlike familial hemiplegic migraines, SHM has no family history. It is caused by genetic abnormalities that alter brain ion channels. The diagnosis is based on symptoms and genetic tests. Standard migraine drugs are used for treatment, with more severe cases requiring specialized interventions. Early detection and treatment are critical for symptom relief and quality of life.

1.1.4. Familial Hemiplegic Migraine

Familial Hemiplegic Migraine (FHM) is an uncommon hereditary migraine type that causes transitory paralysis or weakness on one side of the body, comparable to stroke symptoms. It is associated with speech problems, confusion, visual abnormalities, and severe headaches. FHM is caused by genetic abnormalities that disrupt brain ion channels and occur in families. While symptoms are reversible, timely diagnosis and treatment are critical for managing episodes and improving quality of life. Anticonvulsants and migraine medicines are possible treatment choices.

2. Literature Review

A study by Khan et al. [1] focuses on predicting and classifying various migraine kinds utilizing state-of-the-art machine learning techniques including support vector machines (SVM), K-nearest neighbors (KNN), random forests (RF), decision trees (DST), and deep neural networks (DNN). Seven distinct migraine kinds were identified by training the suggested models with data augmentations. The findings show that with data augmentation, DNN, SVM, KNN, DST, and RF achieved accuracy of 99.66%, 94.60%, 97.10%, 88.20%, and 98.50%, respectively, highlighting AI’s revolutionary potential in migraine diagnosis. Torrente et al. [2] describe the many and most current clinical uses of artificial intelligence for migraine. AI-based models may help properly classify different patient types and can increase diagnosis accuracy, especially for non-headache professionals. Additionally, reactions to certain therapeutic tactics when used alone can be predicted using machine learning techniques. At 6, 9, and 12 months after beginning anti-CGRP(Calcitonin Gene-Related Peptide) medication, L-based models showed an F1 score range of 0.70–0.97 and an AUC(area under the curve) range of 0.87–0.98, with a response rate ranging from 50% to 75%. T. Choudhary [3] et al. provides a novel approach to migraine diagnosis by integrating PSO and genetic algorithms to find traits, then classifying them using Random Forest. The accuracy of the learned approach is 99.63%. From the beginning till 22 April 2024, a systematic search was undertaken by Aramruang et al. [4] using the databases of IEEE, ACM (Association for Computing Machinery) Digital Library, Medline/PubMed, Embase, and Scopus. The prediction model’s risk of bias evaluation tool was used to evaluate the risk of bias. The sample sizes varied from 17 to 1419 individuals, and the models included up to 53 predictors. When applying Logistic Regression and SVM to validate existing ratings, the AUROCs ranged from 0.84 to 0.85. Mitrovic’ et al. [5] show that advanced machine learning algorithms may perform significantly better than traditional statistical correlation techniques in estimating the average MACS(Mean Average Cortical Surface) from cerebral cortex structural MRI(Magnetic Resonance Imaging) data. The accuracy of the ML framework was 82%. Martinelli et al. [6] identified clinical traits that can forecast therapeutic response using machine learning (ML) approaches. We gathered clinical and demographic information from patients with high-frequency episodic migraine (HFEM) or chronic migraine (CM) who had received BoNT-A therapy at our clinic during the previous five years. RF has the greatest accuracy of 85.71%. Stubberud A et al. [7] teach headache researchers and clinicians the fundamentals of machine learning (ML) and artificial intelligence (AI). When using different machine learning techniques, the accuracy ranges from 60 to 90 percent. Saif et al. [8] investigated the prevalence of migraines among Bangladeshi university students, utilizing machine learning to predict their occurrence based on triggering events, and raising awareness to assist migraine sufferers with their everyday activities is the objective. Among male participants, the Logistic Regression approach had the highest accuracy of 78.1%. Among female participants, the Stacked Classifier and Random Forest Classifier reached the highest accuracy of 85.3%. Zhu B et al. [9] showed that somatosensory-evoked potentials are an important and trustworthy signal in migraine categorization, according to the proposed model. With an accuracy of 88.0%, XGB is the most accurate of the handmade feature techniques. Seach et al. [10] created an AI model that uses clinical, demographic, and medication data to forecast the risk of stroke in migraine sufferers. With an AUC-PR of 0.67, deep causal learning models outperform conventional machine learning techniques.

3. Methodology

The purpose of this study is to investigate how well-sophisticated machine learning methods can identify and categorize migraines, especially when applied to datasets with polynomial feature correlations. Using RapidMiner(9.0) as a simulation environment, the suggested architecture starts with feature engineering, data preprocessing, and model selection in an organized manner. To find the best-performing classifiers, RapidMiner’s built-in tools will make it easier to do preliminary testing using algorithms like Random Forest, neural networks, gradient boosting (XGBoost), and support vector machines (SVM) with a polynomial kernel (MLP). The study will go to Google Colab for real-world implementation after assessing model correctness and interpretability in RapidMiner. There, it will adjust hyperparameters and make use of Python-based machine learning tools including sci-kit-learn, TensorFlow, and XGBoost. This method will allow for a thorough assessment of the computing efficiency, generalization capacity, and model performance. The final findings will shed light on the best machine learning methods for classifying migraines, guaranteeing maximum accuracy while skillfully managing polynomial class distributions.

3.1. Support Vector Machine

The SVM with a polynomial kernel is a powerful machine learning technique for non-linear classification tasks. It uses a polynomial function to transfer input data into a higher-dimensional space, allowing intricate correlations between features to be captured. Key parameters are polynomial degree (d), influence range (γ), and regularization (C). A lower degree keeps the model simple, whereas a greater degree catches complex patterns but may result in overfitting. This kernel performs exceptionally well on datasets containing polynomial feature interactions, giving it an excellent choice for migraine classification.

Table 1. Specified parameters for SVM model.

Sr	Parameters	Specified Values
1	SVM Type	C-SVC
2	Kernel Type	Polynomial
3	Degree (d)	2 or 3
4	$Gamma\left(\gamma \right)$	Auto or 1.0
5	C (Regularization)	1.0 or (0.1–100)
6	Coef()	0.0
7	Epsilon($\epsilon$)	0.001
8	Scaling	Yes (Z-score Normalization)

presents the specified parameters used for configuring the SVM model.

3.2. Random Forest

By averaging predictions across trees, the Random Forest ensemble of many decision trees decreases overfitting and increases accuracy, making it very useful for polynomial class characteristics. Compared to a single decision tree, it is more effective at capturing non-linear interactions. The stability of the model is improved by important parameters, including bootstrap sampling, max features for feature selection, and the number of trees (n_estimators), as listed in

Table 2. Specified parameters and their assigned values used for configuring the tree-based model during implementation.

Sr	Parameters	Specified Values
1	Criteria	Gini or Entrophy
2	Max depth	auto
3	Min Samples Split 2	2
4	Min Samples Leaf	1
5	Max Features	4.80
6	Number of trees	100
7	Bootstrap	True

. When polynomial feature interactions are present, Random Forest is a good option for migraine classification because of its robustness..

3.3. Decision Tree

The Decision Tree is a rule-based machine learning technique that divides data into branches based on feature conditions, making it ideal for datasets with polynomial correlations because it captures feature interactions hierarchically [11]. However, a single decision tree might overfit complex polynomial class properties, necessitating pruning, and depth control for generalization [12]. Criterion (Gini or Entropy) and maximum depth are important parameters for refining the model for structured data.

Table 3. Defined parameters and their default or general values considered in the tree-based model configuration.

Sr	Parameters	Specified Values
1	Criteria	Gini or Entrophy
2	Max depth	auto
3	Min Samples Split 2	1
4	Min Samples Leaf	1
5	Max Features	all
6	Number of trees	N/A
7	Bootstrap	N/A

shows the specified parameters applied in the tree-based model.

3.4. XGBoost

Extreme Gradient Boosting, or XGBoost, is a potent machine learning method that excels in classification problems. Using L1 and L2 regularization, it manages overfitting, missing values, and big datasets [13]. By identifying intricate, non-linear patterns, XGBoost helps differentiate between several migraine kinds in migraine classification with polynomial labels. It is useful for precise diagnosis and medical research because of its feature importance analysis, which assists in identifying important clinical indicators.

Table 4. Hyperparameters of the XGBoost model with specified ranges and evaluation metrics used in this study.

Sr	Parameters	Specified Values
1	n_estimator	100–1000
2	Learning rate	0.01–0.3
3	Max depth	3–10
4	Min Child Weight	1–10
5	Subsample	0.5–1.0
6	Colsample by tree	0.5–1.0
7	Gamma	0–5
8	Reg_alpha	0–1
9	Reg_lambda	0–10
10	Objektif	“binary:logistic”, “multi:softmax”
11	Eval_metric	“logloss”, “mlogloss”, “AUC-ROC”

presents the specified parameters for the tree-based model.

3.5. Vote

To categorize migraines and their varieties, we have used a voting ensemble model. This ensemble enhances overall forecasting performance by combining the advantages of multiple distinct models. In particular, the three fundamental learners that make up my voting ensemble are a Decision Tree, a Random Forest, and a Bagging model. Notably, the Bagging model adds another layer of ensemble learning by using Decision Trees as its foundation learners. A majority vote among the three models determines the final prediction. Compared to employing any one model alone, this method possibly achieves superior accuracy and generalization by utilizing the diversity of the models, from the robust, aggregated predictions of the Random Forest and the Bagging ensemble to the single, interpretable Decision Tree.

3.6. Framework

Based on a Kaggle dataset, we present a migraine classification approach that includes missing value imputation and data preparation using SMOTE for imbalance handling. Optimize Selection (Brute Force) is used to pick features, while PCA is utilized to increase accuracy. For evaluation, a variety of classifiers are used, including SVM, XGBoost, Random Forest, and Decision Tree. A variety of tactics are employed to improve model performance, which raises classification reliability and accuracy. Furthermore, we have put out a system called NeuroVote, which shows notable gains in migraine categorization ability. The methods employed in this study are thoroughly discussed. Figure 1 shows the overall framework diagram of the methodology.

3.6.1. Data Collection

The dataset used in this work was gathered from public repository and includes 24 characteristics and 401 occurrences, giving it a current and extensive resource for migraine classification. It features polynomial labels for the many types of migraines, including “Sporadic hemiplegic migraine,” “Basilar-type aura,” “Typical aura with migraine,” “Typical aura without migraine,” and “Familial hemiplegic migraine.” The dataset provides useful information by collecting essential traits that enable accurate classification and analysis. It is an essential tool for advanced research, diagnosis, and treatment techniques because of its organized labeling and variety of features, which make it possible to identify significant trends in migraine occurrences.

3.6.2. Replace Missing Values

To ensure data integrity and completeness, missing values in the dataset were filled up using the corresponding feature mean. By avoiding information loss and maintaining the overall data distribution, this method eventually raises the classification models’ accuracy and dependability.

Table 5. Clinical attributes categorized into different groups used for feature representation in the study.

Attribute 1	Attribute 2	Attribute 3	Attribute 4
Age	Duration	Frequency	Location
Character	Nausea	Intensity	Vomit
Phonophobia	Photopobia	Visual	Sensory
Dysphasia	Dysarthria	Vertigo	Timnitus
Hyperacusis	Diplopia	Defect	Ataxia
Conscience	Paresthesia	DFP	Type

lists the specified parameters used for configuring the tree-based model.

3.6.3. Split Data

Twenty percent of the data is utilized for testing and 80 percent is used for training in our suggested framework. This ensures that the model learns well while being evaluated on unknown data, improving generalization and accuracy.

3.6.4. SMOTE

The label classes had a class imbalance problem that might affect model performance by giving preference to the majority classes. I created synthetic samples for the minority classes using SMOTE (Synthetic Minority Over-sampling Technique) to balance the dataset. Predictions made using this method are more trustworthy and equitable since it increases classification accuracy, guarantees improved model learning, and avoids bias toward dominant classes.

3.6.5. Principal Component Analysis

We used Principal Component Analysis (PCA) to decrease dimensionality while maintaining crucial information to increase accuracy. PCA assists in removing noise, lowering computing cost, and improving model performance by converting the dataset into a lower-dimensional space. Better accuracy and efficiency result from ensuring that only the most pertinent features are used in classification.

3.6.6. Optimize Selection

To find the most essential attributes for enhancing model performance, I used Optimize Selection (Brute Force). Several feature combinations were methodically investigated using this procedure, and the ones that produced the best accuracy were chosen. Despite being computationally demanding, it reduced overfitting and improved efficiency by removing duplicate and unnecessary features. The precision and dependability of the model were thus greatly increased.

3.6.7. Principal Component Analysis

We used a migraine dataset (2024) containing 401 cases and 24 characteristics to create a machine learning model for migraine classification. We used Optimize Selection (Brute Force) for feature selection, corrected class imbalance using SMOTE, and substituted the mean for missing values to guarantee data quality. PCA was also utilized to increase accuracy and decrease dimensionality. For training and testing, the dataset was divided into an 80:20 ratio. We used a variety of strategies to improve model performance and built several classifiers, such as SVM, XGBoost, Decision Tree, and Random Forest. With the greatest accuracy of 99.99% among all classifiers, the proposed framework NeuroVote proved to be successful in classifying migraines. Higher accuracy, improved generalization, and more trustworthy classification findings were all influenced by these actions. Figure 2 shows the accuracy comparison of classifiers (SVM, Decision Tree, Random Forest, XGBoost, and Voting Ensemble) in terms of Accuracy, Error, Recall, and Precision, based on experimental results from this study.

4. Result

The migraine classification model demonstrated high accuracy and reliability by effectively differentiating between different types of migraines. Among the classifiers that were employed, XGBoost fared better than SVM, Decision Tree, and Random Forest, with an accuracy of 98.09%. Optimize Selection (Brute Force) and PCA reduced dimensionality and selected the most relevant features to increase accuracy, whereas SMOTE managed class imbalance well and ensured better model generalization. A balanced method for training and evaluating the model was offered by the 80:20 data split. With XGBoost’s efficiency and accuracy, SVM’s capacity to handle high-dimensional data, Random Forest’s ability to reduce overfitting, Decision Tree’s interpretability, and Time Series Analysis’s ability to capture temporal patterns in migraine occurrences, each classifier was chosen based on its unique benefits. These findings demonstrate how well machine learning methods classify migraines and stress the value of feature selection and data preparation in enhancing model performance.

Table 6. Summary of classifier performance in related studies.

Cite	Year	Author	Classifier	Accuracy
[1]	2024	Khan L	DNN	99.66%
[2]	2024	Torrente A	L-Based Model	97%
[3]	2024	Choudary T	RF	99.63%
[4]	2024	Aramruang	SVM	0.84%
[5]	2023	Mitrovic’	ML Model	82%
[6]	2023	Martinelli D	RF	85.71%
[7]	2024	Stubberud A	ML Model	60–90%
[8]	2023	Saif Z	RF	85.3%
[9]	2019	Zhu B	XGB	88%
[10]	2024	Seah N	DL	0.67%

summarizes the performance of classifiers reported in related studies, while

Table 7. Performance metrics of various classifiers.

Sr	Classifier	Accuracy	Error	Recall	Precision
1	SVM	92.06%	7.94%	87.48%	91.82%
2	Decision Tree (DT)	74.79%	25.21%	30.18%	26.04%
3	Random Forest (RF)	91.01%	8.99%	83.78%	80.49%
4	XGBoost	86.77%	13.23%	72.01%	63.13%
5	Voting Ensemble (NeuroVote)	99.99%	0.01%	99.97%	99.00%

presents the performance metrics of the various classifiers evaluated in this work.

The comparison of accuracies achieved by classifiers is given here.

5. Conclusions

Using cutting-edge methods, this study effectively created a machine learning framework for migraine categorization, achieving 99.99% accuracy with Vote. My suggested system, NeuroVote, has shown remarkable effectiveness in migraine classification. Model performance was greatly enhanced when feature selection, PCA, and SMOTE were combined. Accuracy and interpretability were guaranteed by the classifier selection, which made the model useful for managing complicated and unbalanced medical data. By adding deep learning models, investigating real-time patient data, and integrating cloud-based deployment for accessibility, this study can be expanded in the future. To improve medical insights, future research can concentrate on finding more characteristics associated with migraines and improving the interpretability of the model.