Research On Opportunistic Hierarchical Screening Model For Osteoporosis Based On Multimodal Data

Osteoporosis is a skeletal disease characterized by reduced bone mass and deterioration of bone microstructure,leading to an increased risk of fracture.It’s most common in postmenopausal women and older men.Bone mineral density(BMD)measured on dual energy X-ray absorptiometry(DXA)is considered as "gold standard" for osteoporosis diagnosis.However,owing to the insufficiency of public awareness and the deficiency of osteoporosis prevention and treatment in primary medical institutions,the detection rate of DXA is still at a low level,which seriously affects the diagnosis and treatment of patients.Moreover,most of the studies are based on data from a single source to screen patients with osteoporosis and lack of consideration for the fusion of imaging data and clinical structured data.In this context,with the help of machine learning,this paper builds a hierarchical model based on demographic characteristics,routine laboratory test data and CT images,to opportunistic screen for osteoporosis patients age 50 and older.The main works of this paper are summarized as follows:Firstly,this paper analyzed the feasibility of demographic characteristics and routine laboratory test data on osteoporosis screening,respectively.Five machine learning methods,including logistic regression,support vector machine,multilayer perceptron,random forest and light gradient boosting machine,were used to build models based on demographic characteristics and routine laboratory test data,respectively.In addition,SHapley Additive exPlanations was used to interpret the built models.According to the experimental results,both of data were helpful in osteoporosis screening,which achieved an Area Under Curve(AUC)of Receiver Operating Characteristic was at least 0.746 and 0.686,respectively.Moreover,the built model based on demographic characteristics showed better performance than Osteoporosis Self-Assessment Tool for Asians(AUC=0.719).Among the included features,age and weight were the most important demographic characteristics,as well as alkaline phosphatase,high-density lipoprotein cholesterol and alanine transaminase were the most important routine laboratory test data.Then,the feasibility of lumbar CT image on osteoporosis screening was analyzed based on Radiomics.To be specific,an image segmentation model was constructed firstly based on U-Net to segment lumbar L1-L4.Then,texture features and shape features were extracted from the segmented images.Finally,models were built based on the extracted image features.The experimental results showed that the built image segmentation model could segment the lumbar vertebra precisely,which achieved a dice coefficient of 0.974 int test set.Texture features had better performances(AUC=0.965)than shape features(AUC=0.704).Moreover,it was difficult to improve the performances of the model by merging texture and shape features.Therefore,shape features were not used anymore.Among the extracted features,mean and standard deviation of gray value of image were the most important texture features.Compared with patients with normal BMD,the gray value of CT images of patients diagnosed with osteoporosis was lower and more discrete.Finally,based on the above research work,this paper proposed a model for osteoporosis screening by integrating demographic characteristics,routine laboratory test data,and CT images.According to the data availability,the model was divided into three layers:the first layer only used demographic characteristics,the second layer combined demographic characteristics with routine laboratory test data,and the third layer integrated all of the three-modal data.Experimental results showed that learning method(a decision-level multimodal data fusion method)could improve the performances of the built model,which achieved an AUC of 0.852,0.868 and 0.980 in three layers,respectively.In addition,a web application was developed based on the built hierarchical screening model to facilitate the application of the model.The proposed model integrating multimodal data shows great potential in opportunistic screening for osteoporotic patients age 50 and older.In primary medical institutions where DXA equipment is not accessible,it can be applied as an auxiliary tool to assist clinicians to discriminate patients at high risk of osteoporosis timely.Thus,the proposed model has certain application value.