Disease Inference System-Comparison Of Different Methods
Health plays an important role in ones happinessand well being. Automatic disease inference is mainly used tobridge the gap between what online health seekers with unusualsymptoms need and what busy doctors with biased expertisecan offer. One of the main challenges in health care sectoris to predict diseases by given symptoms. A disease inferencesystem analyze and categorize needs of health seekers andask for manifested symptoms. Basic steps for implementing thesystem is data preparation and processing,feature extraction,and applying Deep learning which predicts the correspondingdisease. In this paper we make a comparison of differentmethods which can be used to implement the disease inferencesystem.
Online health information includes both medical resourcesand patient community connections. it plays an importantrole in patient education and self care. it is seen that froma national consumer survey it shows that the averageconsumers spends nearly 52 hours looking for healthinformation on the internet. Online health resources canbe categorized into two categories. One is the reputableportals run by official sectors,organizations etc. Theyprovide accurate,well structured health knowledge onvarious topics. WebMD and MedlinePlus are some of theexamples. The next category is the community based healthservices.
A Community-based health care service is meantfor people of all ages who need health care assistance athome. Community-based health services are best thoughtof as a sub-system of the overall health system. HealthTapand HaoDF are some of the examples of community basedhealth services. These systems provide a interactive platform,where health seekers can ask the questions by providing thesymptoms while doctors can provide corresponding answersto the queries. There are some limitations to communitybased health care services. First,it is time consumingfor health seekers to get their answers. Second,doctorshave to cope up with heavy workload which reduce theefficiency. Third,since replies are conditioned on doctorsexpertise,conflicts among multiple doctors may occur.
The biggest stumbling block of automatic healthsystem is Disease Inference. Generally people search for:(1) Supplement clues of their diseases (2) Preventiveinformation (3) Possible Diseases by their manifestedsymptoms. The first two involve the exact disease name*This work was not supported by any organization1Aarathy Rajan is a PG Student doing Masters in Computer Science and
Engineering in Vidya Academy of Science and Technology. 2P. Ms. Nitha K P is Assistant Professor with the department of ComputerScience and Engineering,Vidya Academy of Science and Technologyand it can be automatically answered by matching thequestions in the achieved repositories from the structuredhealth portals. The last one seeks to predict the disease frommanifested symptoms. A robust diseases inference approachis the key to break the barrier of automatic wellness systems. Disease inference is a reasoning consequences basedon the given question,this task will be difficult due tofollowing reasons. First, Vocabulary gap makes the datainconsistent. For example,”shortness of breath and breathlesswere used by different health seekers to refer the samedisease named dyspnea. Second,health seekers gives theirquery in an incomplete manner ie they describe the problemin short questions. This factors limits the performancethat can be obtained by generating shallow learningmethods. Shallow learning methods include SVM,DecisionTree. In shallow learning methods,output of learning schemeis directly followed by a classifier as if the system has onlyone layer. In this paper,comparison among six approaches are stud-ied. Six approaches are SVM, KNN, MTSVM, DASVM,SASR, SCDL.
Different Methods discussed in this paper are SupportVector Machine, KNN, Multi-Switch Transductive SVM, De-terministic Annealing Semi-Supervised SVM, Stacked AutoEncoder- Softmax Regression, Sparsely Connected DeepLearning. A. Support Vector MachineA Support Vector Machine (SVM) is a discriminativeclassifier formally defined by a separating hyper plane. Ahyperplane is a subspace whose dimension is one lessthan that of its ambient space. It is a supervised learningmethod. Given labeled training data, the algorithm outputs anoptimal hyper plane which categorizes new examples. In twodimensional space this hyper plane is a line dividing a planein two parts where in each class lay in either side.
Mainly SVM is used for classification and regression. Classificationis the problem of identifying to which of a set of categoriesa new observation belongs, on the basis of a training setof data containing observations and Regression is a set ofstatistical processes for estimating the relationships amongvariables. The mappings used by SVM schemes are designed toensure that dot products of pairs input data vectors may becomputed easily in terms of the variables in the originalspace, by defining them in terms of a kernel function k(x,y)selected to suit the problem.
In Transductive SVM, we need to find out the decisionboundary given a set of labelled examples. This is based onMaximum Margin principle. Maximum Margin principle isdefined as, given an associated distance from the decisionboundary for each example. For this, we find out the lowlocal density. The point at which the low local densityis achieved the decision boundary is placed. Classificationmargin is decided by:Labeled examples,Class labels assignedto unlabeled data.
Deterministic Annealing Semi-Supervised SVM
Deterministic annealing is used for dealing with combi-natorial optimization problems. It is a semi-supervised Ker-nel machines where an easier problem is parametricallydeformed to the original hard problem and minimizers aresmoothly tracked. It involves two steps.
- Discrete variables are treated as random variables overwhich a space of probability distributions is defined.
- The original problem is replaced by a continuous op-timization problem of finding a distribution in thisspace that minimizes the expected value of the objectivefunction. To find the distribution which minimize the expected value ofthe objective function is performed by a Homotopy methodusing negative of the entropy of a distribution as the easy,convex function. Homotopy method is used fro optimization.
Stacked Auto Encoder-Softmax Regression
It is an unsupervised learning algorithm which usesback propagation and constrains the target values toapproximate inputs. Back propagation is a method usedin artificial neural networks to calculate a gradientthat is needed in the calculation of the weights to beused in the network. Autoencoder learns a representation(encoding) for a set of data, typically for dimensionalityreduction.
Reduction is the process ofreducing the number of random variables under considerationby obtaining a set of principal variables. It can be dividedinto feature selection and feature extraction.
Selection tries to find a subset of the original variables. Feature extraction builds derived values (features) intendedto be informative and non-redundant. Here we consider three hidden layers with random ini-tialization incrementally. Softmax classifier is chosen as heoutput layer. Softmax classifiers give you probabilities foreach class label while hinge loss gives you the margin. Itis a simple method since it only interpret probabilities ratherthan margin scores. This architecture is considered to be fullyconnected.
Sparsely Connected Deep LearningIn Sparsely connected Deep learning, consider L layerswith dl nodes in each layer. First layer contains the inputand the lth layer contains the output. Intermediate layers arehidden layers which are unseen from the data. Here nodes inthe higher layer connect to the nodes in the adjacent lowerlayer,rather than fully connected. In this architecture the lasthidden layer and output layer are fully connected.
Data is collected from Everyone Healthy, WebMedand MedlinePlus. It contains Question Answer pairs. Forcomparison we filter the whole dataset and consider certainclasses of diseases. These class of diseases contains onlyquestions where samples and tags were utilized to extractthe disease names. These methods are applied to this dataset. MTSVM and DASVM support binary classification iesupport two values either true or false. SASR and SDCLsignificantly outperform the other two supervised learningalgorithms. But results SDCL shows more accuracy andshow significant performance than SASR,this is due to thenode number in the hidden layer hence it will be hard toobtain optimal result.
After showing the basic comparison we got SDCL showsmore accuracy compared to different approaches. Now acomparison of hoe much hidden layers are also studied. Number of Layers Performance on DatasetStructure with One hidden layer 89. 00%Structure with Two hidden layer 93. 13%Structure with Three hidden layer 98. 21%Here we incrementally added hidden layers between theinput and output layer until it satisfies the convergent criteriaand found that three hidden layers are used to enhance thesystem. Convergent criterion is defined as the accuracy ofdeep leaning model with n hidden layers.
This paper performs study of comparison of six ap-proaches which can be used to implement disease inferencesystem. It shows that Sparsely Connected Deep learning ar-chitecture with three hidden layers shows the best accurateresults compared to other approaches. Therefore it is general-izable and scalable. Classical deep learning architectures aredensely connected and the node number in each hidden layersare adjusted. In contrast it sparsely connected deep learningwith improved efficiency and the number of hidden nodes isdetermined. Appendixes should appear before the acknowledgment.
Cite this Essay
To export a reference to this article please select a referencing style below