| Sl.No. |
Author Name, Year, Place of Study |
Nature of Study |
Number of participants (N), Software/Application Used |
Key elements/Outcome/ Learning Points |
Limitations in the Study |
Reference |
| 1 |
Almadhoun et al [1], March 2021, Gaza, Palestine. Engineering Department |
Subjects not used; Designed an expert system |
Delphi programming language and clips |
Diagnoses 11 headache problems |
11 questions to answer, and each question has multiple sub-questions describing many symptoms. Overlapping of symptoms can cause errors in diagnosis |
[1] |
| Does not require any training before using this expert system |
Inability to diagnose other headache types not listed in the system |
| Not checked the accuracy, specificity, or sensitivity of the expert system. |
| 2 |
Kwon et al [2], 2020 Seoul, South Korea |
Retrospective study |
N=2162. Divided into 2 cohorts: Training=1286, Test=876. |
75 screening questions were used in details |
Other less prevalent although significant primary headaches and secondary headache (other than the causes of thunderclap headaches) were excluded because of a long list of heterogenous diseases causing them. |
[2] |
| Stacked classifier model used with4 layers of binary XGBoost classifiers for differentiating: Migraine, (tension-type headache) TTH, (trigeminal autonomic cephalalgia) TAC, Epicranial headache, and thunderclap headache. |
Stacked XGBoost classifier result: Accuracy: 81% |
Used only 3 clinical features in each stack to draw insight about headache types and the clinical symptoms used here are different from (the International Classification of Headache Disorders) ICHD-3 criteria. |
| LASSO (least absolute shrinkage and selection operator) used for each stacked classifier layer. LASSO compared to SVM-RFE (support vector machine recursive feature elimination) and mRMR (minimum-redundancy maximum-relevancy). |
Sensitivity for: Migraine=88%, TTH=69%, TAC=53%, Epicranial=51%, Thunderclap=51%. |
The clinical course cannot be understood from these 75 questionnaires, so accurate diagnosis is difficult. |
| The selected features used XGBoost classifier which was compared to k-NN (k-nearest neighbor), SVM (support vector machine), and random forest. |
Specificity for: Migraine=95%, TTH=55%, TAC=46%, Epicranial=48%, Thunderclap=51%. |
Data derived from a single center. |
| Performance report in migraine classification was excellent, rest was inferior. This study can be used as pre-screening. |
Conventional machine learning utilized here. No use of deep learning. |
| 3 |
Krawczyk et al [3], 2012. Wroclaw, Poland Mixed department: Technology, Technical Sciences & Medicine |
Questionnaire filled by subjects where headache patients also included; ML algorithms developed and tested; a Prospective study |
N=579 |
Best results noticed in with accuracy %: Random Forest=79.97±3.13, |
Included only migraine, TTH, and loosely defined other headache types which included all remaining headaches be it primary or secondary. |
[3] |
| Age: 20- 65 years |
Bagging=78.24±2.98, |
| Used algorithms: Naive Bayes (a probabilistic classifier), C4.5 (based on ‘Top-Down Induction of Decision Tree’ (TDIDT), Support vector Machine (SVM), Bagging (or bootstrap aggregating), Boosting, Random Forest. |
Boosting 76.68±2.43 |
| Used filter selection algorithms: Consistency measure filter, Relief, Genetic algorithm wrapper |
| 4 |
Julian et al [6], 2019. Study conducted in a hospital, Emergency Department. Aim: Detection of probable secondary headache |
Retrospective study |
N=7972. Primary headache=7098. Secondary headache=874. |
Probable secondary headache: Sensitivity=89%, Specificity=73%, Negative predictive value (NPV)=98.2%. |
Limited to emergency setup |
[6] |
| Records were processed using: Latent Semantic Analysis (LSA). Support Vector Machine (SVM) model used for training. Used Python program. |
Optimized the time in the emergency |
Emergent primary headaches need exploration. |
| 5 |
Messina et al. [7], April 2020. Mila, Italy. Neurosciences Department |
Opinion |
An opinion about Machine Learning in Headache |
– |
– |
[7] |
| 6 |
Celik et al [8], 2009. |
Retrospective collecting records |
Artificial immune system (computational artificial intelligence) |
They were working on a headache classification project and were creating a database from the neurology department in a private hospital |
Claims that would publish results after completion of the project. |
[8] |
| |
Details not known. |
| 7 |
Tezel et al. [9], |
Subjects not used. Designed and developed an AI system |
Clonal selection algorithm (an artificial immunity approach) |
Work on headache diagnosis. Included 250 different symptoms for the training set. 150 symptoms related to headaches. |
Classified into migraine headache, TTH, and set headache. |
[9] |
| Based on the clonal selection principle |
For Test set: Correctly classified symptom set: 96.74% |
| Inspired by biological immunology. |
Incorrectly classified symptom set: 3.26% |
| 8 |
Katsuki et al [10], 2020. Neurology, Neurosurgery Department. Aim: For automated primary headache diagnosis |
Retrospective investigated headache database and developed a DL system |
N=848 |
Accuracy: 0.7759 |
The sample size is small. |
[10] |
| Age: 40-74 years |
Categorized into: Migraine, TTH, TAC, and Other primary headache disorders. |
They did the study in a single hospital. |
| Used Deep learning framework-Prediction One. |
External validation not done |
| Utilized artificial neural network (ANN) with internal cross-validation. |
No separation between chronic and episodic frequent headaches of >=15 days per month to >15 days per month for migraine or TTH headache. |
| Also used Confusion matrix of model |
| Used Japanese language with onomatopoeia, therefore utilized Japanese natural language processing (NLP) |
| 9 |
Keight et al., [11]. Engineering, Medicine and Neurosurgery Department |
Retrospective headache dataset collection from two medical facilities |
N=836 |
Classified headache into Tension-Type Headache, Chronic Tension-Type Headache, Migraine with Aura, Migraine without Aura, Trigeminal Autonomic Cephalalgia. |
– |
[11] |
| Study was done in two medical centers in Turkey. |
Area under the curve (AUC): 0.985 |
| 9 machine learning classifiers used in a supervised learning setting |
Sensitivity: 1 Specificity: 0.966 |
| 10 |
Yin et al [12], 2015. China. Aim: To diagnose two headache types namely probable migraine and probable TTH |
This comprehensive study worked on 3 steps viz. data acquisition through clinical interviews, construction of a case library, and lastly development of a case-based reasoning system |
Clinical decision support systems (CDSSs) are based on case-based reasoning (CBR). |
Can be a diagnostic tool for the general practitioner. |
Inadequate case library due to complex headaches |
[12] |
| K-Nearest Neighbor (KNN) method implemented. |
Accuracy is very high in recognizing these two headaches. |
Needs multi-centric study and validation |
| N=676 cases |
Earlier CBR used: (1) CASEY: to diagnose heart complication |
| Probable migraine (PM) 56.95% Probable TTH (PTTH): 43.05% |
(2) Decision-based support system to diagnose (chronic obstructive pulmonary disease) COPD |
| Test set: N=222. PM: 76.1%, PTTH: 23.9% |
(3) Hybrid case -based reasoning approach to diagnose breast cancer and thyroid disease. |
| 11 |
Qawasmeh et al [13], 2020. Jordan |
Developed an ML-based system where its prediction accuracy checked by a web-based questionnaire’s answer |
N=614 patients records. Public hospital. Males=199; Female=415. Different age group. |
Hybrid model (clustering and classification): Integrated K-means clustering with Random Forest classifier |
Excluded migraine with aura from this study as its differential could be a stroke. |
[13] |
| High-performance headache prediction support system (HPSS) was employed based on a hybrid machine learning model. |
Migraine prediction accuracy=99.1% |
| Used 19 questions related to headache symptoms, according to ICHD-3 criteria. |
Overall accuracy=93% (random forest) |
| 26 classification algorithms were applied to 614 patients. |
HPSS claimed good positive feedback from patients, medical students, and doctors. It is an easy-to-use interface that saves time and effort. |
| 12 |
Woldeamanuel et al [14], 2021. Division of Headache & Facial Pain. Stanford, CA, USA |
A meta-analysis of 41 studies |
Total=41 studies. Median age 43 years, 77% women. The median sample size was 288. |
Used case-based reasoning, DL, classifier ensemble, ant-colony, artificial immune, random forest, white and black box combination, hybrid fuzzy expert system |
60% of the digital tools were based on ICHD criteria. |
[14] |
| 4 studies were based on a questionnaire |
10 studies (25%) compared multiple ML programs |
12% of tools were evaluated in non-clinical centers |
| Phone interviews in 2 studies |
Diagnostic accuracy=89%, sensitivity=87%, specificity=90% |
Interstudy heterogeneity of software |
| Face-to-face interview: 82% (a strong feature) |
No proper patient selection method in 39% of included studies |
| No description of age or sex ratio in 25 studies |
| 13 |
Sah et al [15], 2017. Bhopal, India |
Database created from headache diary and employed selection technique for analysis |
Work on migraine headache classification. Used: data mining classifiers K-NN, support vector machine (SVM), Random Forest, Naïve Bays. |
The best result was derived from the Naïve Bays classification. AUC 0.475, Precision 0.905 |
Data collected from headache diary |
[15] |
| 18 questionnaire |
| 14 |
Liu et al [16], 2022. Shanghai, China. School of medicine |
A cross-sectional study |
Used ML for identifying primary headaches. This is a cross-sectional study design. |
A logistic regression model was comparatively better. |
Only 2 types of headaches were worked on. |
[16] |
| N=173 patients ( 84:migraine, 89:TTH), collected information in neurology clinics using a questionnaire (19 questions) |
Logistic regression has an accuracy of 0.84 and an area under the receiver operating characteristic curve (ROC) of 0.90 |
Mild headache cases could not be included in this study as they did not come for medical advice. |
| Used: Decision tree, Random forest, gradient boosting algorithm, logistic regression, support vector machine (SVM) algorithms |
Helped distinguish migraine and TTH and their important symptomatic distinguishing features |
Small sample size |
| 15 |
Sanchez et al [17], 2020. Colombia |
The study was designed to test the classifier system to distinguish types of migraine |
Aimed at classifying migraine based on symptoms |
ANN provided excellent results with an accuracy of 97.5% and a precision of 97% |
[17] |
| N=400 retrospective medical records |
| Used set of 23 variables/questionnaire of symptoms or signs |
| Implemented artificial neural network (ANN) models, logistic regression models, SVM, nearest neighbor, decision tree |
| 16 |
Celik et al [18], 2017. |
A cross-sectional study to evaluate the accuracy of a classifier algorithm to diagnose primary headache type using web-based questionnaire |
Aimed to diagnose primary headache based on ant colony optimization algorithm. |
Classification Accuracy=96.9412% |
26 patients were misdiagnosed by the ant colony classification |
[18] |
| The web-based questionnaire system used www.migbase.com. Used MySQL database and PHP hypertext preprocessor (PHP) programming language, 40 attributes/questions were included |
Accuracies of migraine, TTH, and cluster headache were 98.2%, 92.4%, and 98.2% respectively |
A similar study was done in Turkey using the same set of patients and the same website for questionnaire but implemented the artificial immune algorithms for primary headache (2015) which reached an accuracy of 99.6471% (used AIRS2-Parallel algorithm) [19] |
| N=850 headache patients from 3 cities who visited a neurologist |
| Age range= 15 to 65 years. |
| 70% female and 30% male. |
Source
