The Role Of Machine Learning In The Early Diagnosis Of Primary Open-angle Glaucoma

Background and PurposePrimary open-angle glaucoma(POAG)is a disease characterizes as chronic progressive optic nerve damage.The pathogenesis of glaucoma currently accepted is about intraocular pressure exceeds the maximum that the intraocular tissues,especially the optic nerve,can withstand,which causes the retinal ganglion cells(RGCs)apoptosis,and lead to optic nerve damage and visual function defection.Therefore,it is manifested as thinning of the nerve fiber layer(RNFL)and progressive damage of visual field(VF),namely structural damage and functional damage.Due to the the high risk of blindness without obvious symptoms,early diagnosis and timely treatment are essential to maintain patients’ visual function and quality of life.The auxiliary examination of POAG in clinic usually consists of OCT,intraocular pressure and visual field examination.With the continuous development of computer technology and image segment technology,the application of Artificial Intelligence(AI)in the medical field is expanding,and the research on the diagnosis and treatment of eye diseases by machine learning and deep learning is increasing.At present,the application of AI in the diagnosis and treatment of glaucoma mainly focuses on the combination of auxiliary diagnostic images(fundus photography,SAP,OCT,etc.).With the continuous progress of information technology,AI will be widely used in the routine diagnosis and treatment of glaucoma.However,there is still a lack of tri-classification diagnostic model combining structural and functional examination.Therefore,in our study,the diagnosis of early glaucoma,suspected glaucoma and healthy eyes is classified into three categories.In this study,early POAG,suspicious POAG and healthy patients were differentiated through machine learning,and a relatively accurate model for early glaucoma diagnosis was established by comparing with the diagnosis results of clinicians.MethodsRetrospective cohort study.From September 2014 to September 2019,97 cases of early POAG eyes,86 cases of suspected glaucoma and a total of 291 cases of healthy eyes were collected from ophthalmology department of the Second Affiliated Hospital of Harbin Medical University,which met the diagnostic criteria for POAG.OCT retinal nerve fiber layer(RNFL)thickness scanning,Humphrey automated perimeter 24-2 program examination,and 24-hour intraocular pressure examination were performed to obtain 6 quadrant thickness values of RNFL,24-hour intraocular pressure values(mean value,highest value,lowest value and difference value),visual field examination parameters(VFI,MD,PSD)and age.Four different Machine learning methods,including Logistic Regression(LR),Random Forest(RF),Support Vector Machine(SVM)and Neural Network(NN),were used to differentiate these data by 5 fold-cross-validation.The accuracy of MLC(Machine Learning Classifier),which had the highest AUC performance,was compared with the average value of that of three glaucoma specialists and three general ophthalmologists.ResultsThe importance of each feature data to machine learning is different,the highest of which is the mean intraocular pressure(importance=0.88),the second is MD(importance=0.86),the lowest contribution to the model is VFI(importance=0.06),and the second is the temporal superior area.RNFL thickness(importance=0.07).All the four learning methods obtained excellent AUC results(0.958-0.984),and RF was the best(0.984).The sensitivity,specificity and accuracy of SVM were the highest(0.963,0.923 and 0.943,respectively).The diagnostic accuracy of glaucoma specialists was 90.38%,and that of general ophthalmologists was 84.53%,which was lower than that of RF with the highest AUC value(92.5%).The diagnosit accuracy of NN with the lowest AUC value(90.6%)was similar to that of the glaucoma specialist group,but still higher than that of the non-glaucoma specialist group.ConclusionsAll the four machine learning methods can effectively distinguish between early POAG and suspicious POAG,and SVM and RF are better,and the accuracy is higher than that of ophthalmologists.An early diagnosis model of POAG machine learning based on OCT,intraocular pressure and field of vision can be established clinically.