Healthcare Analytics: Case Studies

Healthcare Analytics: Case StudiesTrần Thế Truyền & Shivapratap GopakumarCenter for Pattern Recognition and Data AnalyticsDeakin University, AustraliaEmail: truyen.tran@deakin.edu.auURL: truyen.vietlabs.comHUST, Dec 2013risk prediction/stratification2readmissiondeathtoxicitystressquality-of-lifeprogression to advanced stageslength-of-stayside effectssuicide attemptsHeart Failure Readmission Prediction3Why predict heart failure(HF) readmission ?Heart Failure (HF): inability of the heart to pump blood efficientlyNo permanent cureOnce hospitalized, need subsequent visits Leading cause of hospitalization in older adults50% of HF patients re-hospitalized in 6 months70% of rehospitalisation because of worsening of previous HF conditionsWhy predict heart failure(HF) readmission ?Readmission frequency = measure of hospital qualityReadmissions are costly for patient and hospitalHow to reduce preventable readmissions ?Identify the causes of HF readmissionsAssess the risk in individual patientsHF Readmission Models Difficult task: huge number of socio-demographic variables 6 different models that exclusively predict heart failure readmissions [1]Use limited subset of variables for predictionLimited subset: from clinical expertise, prev studiesDo not share final predictorsCan we have a readmission prediction model for heart failure without any prior hypothesis (clinical experts, previous findings)?Use all available data from EMR database?Our Proposed ModelUse all available information with no bias Perform a multivariate regression analysisUse strong feature selection techniquesEnsure stability of the resulting feature subsetData CollectionData from local hospital: 1405 patients with HFTemporal Split for training and testing Feature ExtractionConverts inpatient time-stamped events into high-dimensional feature vectorMultiple time scales taken into accountOne sided convolutional filter bank:Summarizes event statistics over multiple time periodsModel TrainingLasso regularized logistic regressionWhy lasso?Effective in handling high dimensional variablesStrong feature selectionWeak features derived to zeroResults in Sparse modelBut sparsity invites instabilityThe Instability ProblemEMR derived features are:Weakly predictive of some tasksHighly correlatedApplication of Lasso RegularizationChooses one of strongly correlated feature pairsInconsistent feature weightsLasso ShrinkageHigh Dimensional EMROne Solution: Encourage statistical strength among correlated featuresThe Instability ProblemLasso regularizationLaplacian regularization with Feature GraphsFeature graphNodes: featuresEdges: relationship between featuresRelationship TemporalStructural: ICD-10 codesModel TrainingLasso regularized logistic regressionModified by feature graph regularization intoModel EvaluationValidation data: data from futureModel Discrimination (AUC):0.65 for 6 months readmission0.66 for 12 months readmissionMeasuring StabilityStability[3] was measured using Consistency Index: corrects overlapping due to chanceJaccard Index: fraction between cardinalities of intersection and union subsetsTop Predictors for HF readmissionRanking based on Importance.Feature importance = learned weight of feature * std_dev(feature)Our findingsIt is possible to have a fully automated readmission model for HF from EMR data without prior knowledgeSuch model has similar discrimination compared to existing modelsFeature graphs can be applied for stabilizing clinical modelsHandwritten Digit Recognition20Data: MNIST60,000 for training10,000 for testing28x28 B&W images10 Categories {0->9}21Model ChoicesMulticlass logistic regression, aka:Maximum Entropy classifier (or MaxEnt)SoftmaxMulticlass SVMAdaBoost.MHRandom Forests22Data Preprocessing ChoicesRaw Pixel: 784 dimsPCA, e.g., using 100 first eigenvectorsMore recently:RBM, e.g., using 500 hidden unitsNRBM (Nonnegative RBM)AutoencodersDBN (Deep Belief Networks)DBM (Deep Boltzmann Machines)23Visualization Tools in 2D: PCA24Visualization Tools in 2D: t-SNE25Heritage Health Prize2627