| Background:In recent years,artificial intelligence technology has developed rapidly in the medical field,and more and more medical AI systems that simulate doctors’ diagnostic capabilities have been developed in order to provide patients with better medical services.However,the current clinical application of medical AI systems in the real world is not yet popular.So far,AI has experienced two generations of development.The first generation of AI is based on a knowledge-driven AI model,which aims to imitate the reasoning and decision-making capabilities of human experts.Its significant advantage is that the output results are interpretable.The main disadvantage is that when processing big data,it is time-consuming,labor-intensive and computationally expensive.The second generation of AI is a data-driven AI model represented by deep learning.This type of model can automatically extract rules from a large number of cases without the guidance of expert knowledge.Its advantage is that it has the advantage of big data processing,can automatically search and process a large number of variables,so as to reliably classify and predict the results.However,the lack of interpretability of the output results is the main problem it faces.Therefore,in recent years,AI scholars have proposed a new concept,that is,the third generation of AI.Generally speaking,it is to build a more powerful,interpretable and robust AI system by combining the advantages of knowledge-driven and data-driven methods,which promotes the innovative application of AI.Laboratory medicine is an important part of modern medicine,and 70% of the information required for clinical decision-making comes from laboratory.The laboratory test items cover thousands of indicators including clinical chemistry,toxicology,hematology,immunology and microbiology,etc.,providing clinicians with a large amount of objective data to support clinical decision-making.As we all know,a large number of common diseases,including diabetes,cancer,endocrine,infectious,and hereditary diseases,require laboratory tests to confirm or assist in the diagnosis.However,the current development trend of clinical specialization is inevitable.Clinicians are only familiar with the laboratory test items commonly used in their specialties,and it is easy to overlook key results and important parameters that are beyond the scope of their expertise.In addition,with the rapid development of laboratory medicine,there are more and more test items.Since disease affects the overall state of the body,key information or important trends of the disease are hidden among many abnormal laboratory data.In the process of disease diagnosis and treatment,when clinicians face abnormalities in multiple test items,it is difficult to extract the corresponding medical information from the abnormal test data.Even experienced doctors can easily ignore the internal relationship,resulting in missed or misdiagnosed disease.Although the correlation of comprehensive analysis and test data is of far-reaching significance for clinical disease judgment,this process involves complex logical reasoning and calculations,which brings huge challenges to human brain analysis.Therefore,the development of interpretable laboratory AI tools based on laboratory data through AI technology to mine the hidden value of laboratory data provides a huge application prospect for improving the value of laboratory diagnosis.Aims:The main purpose of this research is that using knowledge graphs and ML algorithms to train and analyze a large number of laboratory data and diagnosis data to establish a knowledge and data dual-driven AI system with robustness and interpretability,whose core function is to accurately diagnose diseases and provide reasonable explanations.At the same time,develop a multi-functional intelligent laboratory platform integrating laboratory data standardization,intelligent laboratory report interpretation and intelligent disease recommendation,so that the laboratory AI system can be used in clinical practice.Methods:Methods used in this topic are as follows:1.Construct a laboratory knowledge graph.The construction of laboratory knowledge graph mainly includes four aspects: conceptual design,knowledge construction,knowledge graph application and knowledge supplement.2.Data preprocessing.The data preprocessing process includes data cleaning,data integration,feature standardization and feature transformation.3.Data collection,study population and study design.Taking January 1,2020 as the time node,all qualified participants of outpatients and inpatients from January 1,2010 to October 31,2020 were divided into retrospective cohorts and prospective cohorts.Among them,the retrospective cohort is randomly divided into training set and validation set at a ratio of 8:2.The training set is used to train and build an AI system for laboratory.The prospective test cohort is used to test the disease diagnosis performance of the AI system.4.Construct a multi-label disease diagnosis system.Simulating the whole process of doctors’ reasoning about disease diagnosis,this research constructed a disease diagnosis system based on different organ systems.First,the diagnosis is divided into different organ system diagnosis.Then,each organ system is further divided into different disease diagnosis.5.Build a laboratory AI system.Three ML algorithms,LR,XGBoost,and ANN,are used to construct three data-driven models of laboratory AI systems.The integration of the knowledge-driven model based on laboratory knowledge graph and the data-driven model based on the ML algorithm presents the form and content of the dual-driven laboratory AI system of knowledge and data.6.Web application and graphical display of the AI system.Develop a smart laboratory platform.The core components mainly include four modules,including the intelligent assistant of the laboratory physician,the clinical laboratory data center,the inspection AI operation monitoring,and the laboratory big data board.7.Data analysis.The overall diagnostic performance of the model is evaluated using Recall and m AP.Results:1.Baseline characteristics.A total of 730,113 qualified participants from all outpatients and inpatients from January 1,2010 to October 31,2020 were included in the clinical study.Among them,there are 509,841 people in the training set,235,074 people in the validation set,and 69,101 people in the test set.2.LR data driven model.Based on the LR algorithm model,the accuracy of the AI system for predicting diseases(m AP=87.53%)is slightly better than that of ANN(m AP=86.83%)and stronger than XGBoost(m AP=84.41%).Therefore,this study chooses the LR algorithm to construct a data-driven model for laboratory AI system.3.Construction of knowledge and data dual-driven laboratory AI system.The laboratory AI system consists of two core models,namely,a knowledge-driven AI model based on the laboratory knowledge graph and a data-driven AI model based on the LR algorithm.In this study,a total of 4,147 knowledge-driven AI models and 168 disease data-driven AI models were established.4.Working ideas of the AI system.There are four working modes of disease diagnosis in the AI system.Mode 1 is called the knowledge-driven AI diagnosis mode of "gold standard" diagnosis;Mode 2 is called the superimposed knowledge-driven AI diagnosis mode of "classified diagnosis";Mode 3 is called the "knowledge + data driven" of "differential diagnosis" AI diagnosis mode;Mode 4 is called the "knowledge + data driven" AI diagnosis mode that "drives diagnosis with clinical features".5.Laboratory AI’s diagnostic performance evaluation for multiple systems and multiple diseases.The laboratory AI system has achieved a high level of accuracy and comprehensiveness at all levels of diagnosis.When predicting 10 organ systems,the m AP value is 95.19% and the Recall value is 100.00%;when predicting 10 specific diseases,the m AP value is 96.01% and the Recall value is 78.90%.6.Clinical application of the AI system.The laboratory physician’s intelligent assistant displays the model algorithm diagnosis recommendation based on the patient’s laboratory data with the diagnosis explanation of the disease.The clinical laboratory data center is used to realize the structure,standardization and normalization of laboratory data.Laboratory AI system operation monitoring can record and display the prediction results of laboratory AI system.Laboratory big data board displays to check the running status of the AI system,etc.Conclusions:In this study,the laboratory AI system driven by knowledge and data can automatically identify and comprehensively analyze 2,071 laboratory indicators,and complete the multiple relationship reasoning for the diagnosis of 10 kinds of organ system diseases including infection,respiratory system and blood system,and 441 kinds of specific diseases,and all the reasoning processes have good interpretability.The prediction of 10 organ systems and 10 specific diseases has very little missed diagnosis and high accuracy.The good disease diagnosis performance of the laboratory AI system helps clinicians make diagnosis and treatment decisions;it is also of great significance for the early diagnosis of complex and difficult diseases. |
