| Meningitis is mainly characterized by headache,fever,neck rigidity,vomiting,and seizures.It is a common and frequently occurring disease in the clinic and has high morbidity and mortality.Early definitive diagnosis and treatment are the most important factors to reduce the mortality and neurological sequelae in patients.Clinically,the most common kinds of meningitis are tuberculous meningitis,bacterial meningitis,cryptococcal meningitis,viral meningitis,autoimmune encephalitis,parasite meningitis,etc.The early clinical manifestations of these types of meningitis are atypical.Besides,laboratory tests are not sensitive and the positive rate of etiology detection in cerebrospinal fluid is low.Therefore,it is difficult for clinicians to identify the cause of meningitis at early stage,which may easily lead to misdiagnosis,missed diagnosis,and delay in treatment,which may affect the prognosis of patients.In recent years,with the rapid development of artificial intelligence technologies in the medical field,the "artificial intelligence + medical" model has been broadly used in clinical diagnosis and treatment,providing a new direction for the solution of meningitis diagnosis.However,there is no artificial intelligence system applied to meningitis at home and abroad.ObjectiveBy constructing the expert system and machine learning algorithm for meningitis,we established an intelligent auxiliary diagnosis system for meningitis and preliminary verified the system.In addition,the system will collect clinical data from a large number of meningitis patients to provide clinical evidence for deep learning and big data analysis.MethodsData of clinically diagnosed meningitis patients were collected,of which 449 cases were used for optimal debugging of the expert system,machine learning and initial validation of the system.38 prospective cases were used for further clinical validation of the system algorithms.Ten cases were extracted from the above data for man vs.machine competition to verify the clinical diagnostic efficacy of the system.According to the clinical diagnosis procedure of meningitis,the interface of the intelligent auxiliary diagnosis system for meningitis was designed.Combining related literature,textbooks,expert knowledge,domestic and international guidelines of meningitis,we constructed a knowledge base of meningitis and designed diagnostic rules,thereby initially established a meningitis expert system.At the same time,the case data were used for machine learning to screen the optimal machine learning methods for identifying meningitis.Combining machine learning algorithms with expert system,a complete meningitis intelligent auxiliary diagnosis system was constructed and deployed on the network platform.Finally,we compared the diagnose accordance rate among machine learning,expert system and clinicians.The man vs.machine races about etiological subtypes of meningitis were organized to verify the diagnostic efficacy of the system.Results 1.The clinician's initial diagnosisIn the 449 cases of meningitis patients whose data were retrospectively included,the results of initial diagnosis by clinicians were as follows: Of the 127 TBM patients,1 case was diagnosed before admission,and 126 were diagnosed correctly in 15 cases.The correct rate was 11.9%.Among the 140 cases of BM,8 cases were diagnosed before admission and 132 cases were diagnosed.The initial diagnosis was correct in 72 cases with a correct rate of 54.5%;in 61 cases of CM,18 cases were diagnosed before admission,43 cases were diagnosed correctly in 4 cases,and the correct rate was 9.3%;of the 121 cases of VM,the initial diagnosis was correct in 60 cases.The correct rate is 49.6%.The total accuracy of the initial diagnosis by clinicians was 35.8%.2.The intelligent auxiliary diagnosis system for meningitisAn intelligent aided diagnosis system for meningitis has been established.The system contained two diagnostic models: one was an expert system based on knowledge,and the other was an intelligent system based on machine learning algorithms.The expert system basically covered all clinical meningitis,including VM,TBM,BM,CM,autoimmune meningitis,cerebral parasite infectious diseases,cancerous meningitis,and set special labels to diagnose specific meningitis.However,due to limited learning data,machine learning algorithms currently only could recognize four types of meningitis: TBM,BM,CM,and VM.At present,the system has been deployed on the Internet and could be downloaded from mobile phones.3.The diagnostic accordance rate of expert systemIn 449 cases of meningitis,the diagnosis was consistent with 288 cases,not 161 cases,and the total coincidence rate was 64.1%.The TBM,BM,CM,and VM accordance rates were 40.9%,77.1%,42.6%,and 84.3%,respectively.4.Machine learning resultsFinally,we chose the random forest algorithm to identify meningitis.According to the random forest algorithm,the 15-dimensional parameters for identifying meningitis were ranked in order of importance: erythrocyte sedimentation rate,duration of onset,total leukocyte count in cerebrospinal fluid,cerebrospinal fluid lymphocyte ratio,age,cerebrospinal fluid neutrophil granulocyte ratio,IgA,cerebrospinal fluid glucose/blood glucose,IgM,IgG,cerebrospinal fluid sugar,lumbar puncture pressure,cerebrospinal fluid protein,blood glucose,cerebrospinal fluid monocyte ratio.Through ten times of ten-fold cross validation results,the total recognition rate of random forest in 15 dimensions was 81% and the standard deviation is 5.5%.The recognition rates of TBM,CM,VM,and BM were 75%,68%,90%,and 85%,respectively.5.Clinical validation of the systemBy comparing the compliance rates of the expert system,random forest algorithm,and the initial diagnosis of clinicians,we could obtain the fact that random forest algorithm outperform the others,and the initial diagnosis of clinicians had the lowest rate in the diagnosis of meningitis disease.When the results were prospectively verified,the coincidence rates of the expert system and the machine learning algorithm were both 78.9%,and the difference was not statistically significant.The diagnostic accuracy of both of them was low for CM.The results of the man vs.machine competition had shown that the system ranked first with an accuracy rate of 70% compared with 38 participating doctors(the highest accuracy rate for doctors was 60%).In addition,the system was also faster than human doctors.It took an average of 60 seconds per question,while the participating doctors spent an average of 150 seconds per question.Therefore,it was initially demonstrated that the system was superior to the clinicians in terms of diagnostic accuracy and speed.ConclusionThe design of the intelligent auxiliary diagnostic system for meningitis was based on clinical guidelines,norms,and consensus.It includes two diagnostic models—— expert systems and machine learning.It can be used in clinical practice to solve the problem of meningitis diagnosis.After the system is launched on the mobile phone,it will provide clinical decision support for clinical young doctors and primary physicians.It will also assist clinicians in making more scientific and rational clinical decision of meningitis.In addition,the system is also conducive to doctors to standardize the inquiry process and learn new treatment programs.It is also conducive to the collection of a large number of meningitis records,in order to provide a basis for deep learning and big data analysis.At the same time,this study also created a precedent for artificial intelligence + meningitis,which provided new ideas for the development of medical decisions. |
PDF Full Text Request