Discussion and 3 comments

varunydd2

BigDataAnalyticsinHealthcareSystem-IBM.pdf

Home >Computer Science homework help >Discussion and 3 comments

Big Data Analytics for Healthcare

Chandan K. Reddy Department of Computer Science

Wayne State University

Jimeng Sun Healthcare Analytics Department IBM TJ Watson Research Center

2Jimeng Sun, Large-scale Healthcare Analytics

Healthcare Analytics using Electronic Health Records (EHR)

 Old way: Data are expensive and small – Input data are from clinical trials, which is small

and costly

– Modeling effort is small since the data is limited

• A single model can still take months

 EHR era: Data are cheap and large – Broader patient population

– Noisy data

– Heterogeneous data

– Diverse scale

– Complex use cases

3Jimeng Sun, Large-scale Healthcare Analytics

Heterogeneous Medical Data

DiagnosisDiagnosis

MedicationMedication

LabLab

Clinical notes

ImagesImages

Genetic data

4Jimeng Sun, Large-scale Healthcare Analytics

Challenges of Healthcare AnalyticsScalability ChallengesChallenges in Healthcare Analytics

Collaboration across domains

Analytic platform

Intuitive results

Scalable computation

PARALLEL MODEL BUILDING

6Jimeng Sun, Large-scale Healthcare Analytics

Motivation – Predictive modeling using EHR is growing

 Need for scalable predictive modeling platforms/systems due to increased computational requirements from:

– Processing EHR data (due to volume, variability, and heterogeneity) – Building accurate models – Building clinically meaningful models – Validating models for accuracy and generalizability

Explosion in interest

7Jimeng Sun, Large-scale Healthcare Analytics

What does it take to develop a predictive model using EHR?

Marina: IBM Analytics Consultant

2 3

4 5

Within 3 months, we need to 1. understand business case 2. obtain the data 3. prepare the data 4. develop predictive models 5. deliver the final model

David Gotz, Harry Starvropoulos, Jimeng Sun, Fei Wang. ICDA: A Platform for Intelligent Care Delivery Analytics, AMIA 2012

8Jimeng Sun, Large-scale Healthcare Analytics

A Generalized Predictive Modeling Pipeline

 Cohort Construction: Find an appropriate set of patients with the specified target condition and a corresponding set of control patients without the condition.  Feature Construction: Compute a feature vector representation for each

patient based on the patient’s EHR data.  Cross Validation: Partition the data into complementary subsets for use in

model training and validation testing.  Feature Selection: Rank the input features and select a subset of relevant

features for use in the model.  Classification: The training and evaluation of a model for a specific classifier.  Output: Clean up intermediate files and to put results into their final locations.

Model specification

9Jimeng Sun, Large-scale Healthcare Analytics

Cohort Construction

A ll

pa tie

nt s

Disease Target samples D1 Hypertension control 5000 D2 Heart failure onset 33K D3 Hypertension diagnosis 300K

Cases

Controls D3

10Jimeng Sun, Large-scale Healthcare Analytics

Feature Construction

 We define

– Diagnosis date and index date

– Prediction and observation windows

 Features are constructed from the observation window and predict HF onset after the prediction window

Index date Diagnosis date

Observation Window

Prediction Window

PARAMO: A Parallel Predictive Modeling Platform

Pipeline Specifications

Dependency Graph

Dependency Graph Generator

• Remove Redundancies • Identify Dependencies

Dependency Graph Execution Engine • Prioritization • Scheduling • Parallel Execution

Results

Parallelization Infrastructure

Pipeline Specifications

Dependency Graph

Dependency Graph Generator

• Remove Redundancies • Identify Dependencies

Dependency Graph Generator

• Remove Redundancies • Identify Dependencies

Dependency Graph Execution Engine • Prioritization • Scheduling • Parallel Execution

Results

Parallelization Infrastructure

An Example Set of Pipeline Specifications

 Cohort construction: One patient data set  Feature construction:

 Cross-validation: 2-fold cross-validation  Feature selection: Information Gain, Fisher Score  Classification: Naïve Bayes, Logistic Regression, Random Forest

Feature Type Aggregation Diagnoses Count Medications Count Symptoms Count Labs Mean

Dependency Graph Generator

 Input: Pipeline specifications

 Output: Dependency graph

 Function:

– Remove Redundancies

– Identify Dependencies

– Encode parallel jobs

Pipeline Specifications

Dependency Graph

Dependency Graph Generator • Remove Redundancies • Identify Dependencies

19Jimeng Sun, Large-scale Healthcare Analytics

Dependency Graph

20Jimeng Sun, Large-scale Healthcare Analytics

Dependency Graph Execution Engine

 Input: Dependency graph

 Output: Results (models, scores, etc.)

 Function: – Schedules tasks in a topological ordering of the graph – Prioritizes pending tasks using information from already completed tasks – Executes tasks in parallel via the parallelization infrastructure

Dependency Graph

Dependency Graph Execution Engine

• Prioritization • Scheduling • Parallel Execution

Results

Parallelization Infrastructure

Hadoop Cluster

Multi-Process Single Server

…Nimble

Dependency Graph Runtime Analysis (Hadoop, 20 Concurrent Tasks)

Experimental Data Sets

Data Set Years of Data

Number of Patients

Number of Features

Number of Records

Number of Cases

Number of Controls

Target Condition

Small 3 4,758 25932 3,312,558 615 949 Hypertension Control

Medium 10 32,675 46117 24,719,809 4644 28031 Heart Failure Onset

Large 4 319,650 49269 33,531,311 16385 164743 Hypertension Onset

23Jimeng Sun, Large-scale Healthcare Analytics

Experimental Pipeline Specifications

 Cohort construction: One patient data set: Small, Medium, or Large

 Feature construction:

 Cross-validation: 10 x 10-fold cross-validation

 Feature selection: Information Gain, Fisher Score

 Classification: k-NN, Naïve Bayes, Logistic Regression, Random Forest

Feature Type Aggregation Diagnoses Count Medications Count Procedures Count Symptoms Count Labs Mean

24Jimeng Sun, Large-scale Healthcare Analytics

Running Time vs. Parallelism level

 Small: 5,000 patients, Medium: 33K, Large: 319K

 10 times 10-fold cross validation

 Dependency graph: 1808 nodes and 3610 edges

1000

10000

100000

1000000

Serial 10 20 40 80 120 160

NumberofConcurrent Tasks

R un tim e (s )

Large Medium Small

9 days

3 hours

72X speed up

27Jimeng Sun, Large-scale Healthcare Analytics

Summary

 Predictive models in healthcare research is becoming more prevalent

 Electronic health records (EHR) adoption continues to accelerate

 Need for scalable predictive modeling platforms/systems

 PARAMO is a parallel predictive modeling platform for EHR data

 PARAMO can facilitate large-scale modeling endeavors and speed-up the research workflow

 Tests on real EHR data show significant performance gains

PATIENT SIMILARITY PLATFORM

29Jimeng Sun, Large-scale Healthcare Analytics

Patient Similarity Problem

Patient Doctor S

im ila

ri ty

se ar

30Jimeng Sun, Large-scale Healthcare Analytics

Patient Similarity Problem

Patient Doctor

31Jimeng Sun, Large-scale Healthcare Analytics

Modern setting

Traditional setting

Challenges of Patient Similarity

 As the size of a feature vector increases, it is hard to find exact match on all features

 How to find relevant patients to a query for a specific clinical context?

Query patient Feature vector

EHR databaseEHR database

…

Similar patients

No exact match >100k dimensional feature vector

32Jimeng Sun, Large-scale Healthcare Analytics

Our Approach

 For a clinical context,

1. What are important features?

2. What is the right similarity measure?

Query patient

Important features …

>100k dimensional feature vector

2. Patient similarity learning

1. Feature Selection & Generalization

33Jimeng Sun, Large-scale Healthcare Analytics

Healthcare Analytic Platform

Large-scale Analytics Platform

34Jimeng Sun, Large-scale Healthcare Analytics

Healthcare Analytic Platform

Healthcare Analytics Information Extraction

Data Mining Visualization

35Jimeng Sun, Large-scale Healthcare Analytics

Healthcare Analytic Platform

Healthcare Analytics

Information Extraction

Data Mining Visualization

Structured EHR

Feature extraction Feature

extraction

Patient representation

Unstructured EHR

Feature selection Feature selection

Patient similarity Patient

similarity

ContextContext

VisualizationVisualization

36Jimeng Sun, Large-scale Healthcare Analytics

Feature Extraction from Unstructured EHR data

Healthcare Analytics

Information Extraction

Data Mining Visualization

Structured EHR

Feature extraction Feature

extraction

Patient representation

Unstructured EHR

Feature selection Feature selection

Patient similarity Patient

similarity

ContextContext

VisualizationVisualization

37Jimeng Sun, Large-scale Healthcare Analytics

Motivations for Early Detection of Heart Failure

 Heart failure (HF) is a complex disease

 Huge Societal Burden

 Diagnoses are usually made late, despite there are symptoms documented in clinical notes prior

 Our method exacting HF symptoms achieves precision 0.925, recall 0.896, and F-score 0.910

Roy J. Byrd, Steven R. Steinhubl, Jimeng Sun, Shahram Ebadollahi. Automatic identification of heart failure diagnostic criteria, using text analysis of clinical notes from electronic health records. International Journal of Medical Informatics 2013

38Jimeng Sun, Large-scale Healthcare Analytics

Potential Impact on Evidence-based Therapies

0.00%

10.00%

20.00%

30.00%

40.00%

50.00%

60.00%

70.00%

Preceding Framingham diagnosis

After Framingham diagnosis

After clinical diagnosis

 Applying text mining to extract Framingham symptoms can help trigger early intervention

3,168 patients eventually all diagnosed with HF

Vhavakrishnan R, Steinhubl SR, Sun J, et al. Potential impact of predictive models for early detection of heart failure on the initiation of evidence-based therapies. J Am Coll Cardiol. 2012;59(13s1):E949-E949.

No symptoms Framingham symptoms

Clinical diagnosis

Opportunity for early intervention

39Jimeng Sun, Large-scale Healthcare Analytics

Knowledge plus Data Feature Selection

Healthcare Analytics

Information Extraction

Data Mining Visualization

Structured EHR

Feature extraction Feature

extraction

Patient representation

Unstructured EHR

Feature selection Feature selection

Patient similarity Patient

similarity

ContextContext

VisualizationVisualization

41Jimeng Sun, Large-scale Healthcare Analytics

Combining Knowledge- and Data-driven Risk Factors

Knowledge

Data

Combination Risk factor

augmentation Risk factor

augmentation Combined risk factors Combined risk factors

Knowledge base

Risk factor gathering

Risk factor gathering Knowledge

risk factors Knowledge risk factors

Clinical dataClinical data Data

processing Data

processing

Potential risk

factors

Potential risk

factors

Target condition

Risk factor augmentation

Dijun Luo, Fei Wang, Jimeng Sun, Marianthi Markatou, Jianying Hu,Shahram Ebadollahi, SOR: Scalable Orthogonal Regression for Low-Redundancy Feature Selection and its Healthcare Applications. SDM’12

Jimeng Sun, Jianying Hu, Dijun Luo, Marianthi Markatou, Fei Wang, Shahram Ebadollahi, Steven E. Steinhubl, Zahra Daar, Walter F. Stewart. Combining Knowledge and Data Driven Insights for Identifying Risk Factors using Electronic Health Records. AMIA2012

42Jimeng Sun, Large-scale Healthcare Analytics

Risk Factor Augmentation

 Sparse learning objective formulation:

Model error Correlation among data-driven features

Correlation between data- and knowledge- driven features

Sparse Penalty

Dijun Luo, Fei Wang, Jimeng Sun, Marianthi Markatou, Jianying Hu,Shahram Ebadollahi, SOR: Scalable Orthogonal Regression for Low-Redundancy Feature Selection and its Healthcare Applications. SDM’12

44Jimeng Sun, Large-scale Healthcare Analytics

Prediction Results using Selected Features

+200+150

+100 +50

all knowledge features+diabetes

+Hypertension

CAD

0.5

0.55

0.6

0.65

0.7

0.75

0.8

0 100 200 300 400 500 600

A U C

Numberof features

Knowledge driven features improve AUC slightly

Data driven features improve AUC significantly

45Jimeng Sun, Large-scale Healthcare Analytics

Top-10 Selected Data-driven Features

 9 out of 10 are considered relevant to HF

Feature Relevancy to HF Dyslipidemia Thiazides-like Diuretics Antihypertensive Combinations Aminopenicillins Bone density regulators Naturietic Peptide Rales Diuretic Combinations S3Gallop NSAIDS

46Jimeng Sun, Large-scale Healthcare Analytics

Top-10 Selected Data-driven Features

 9 out of 10 are considered relevant to HF

 The data driven features are complementary to the existing knowledge-driven features

Feature Relevancy to HF Dyslipidemia Thiazides-like Diuretics Antihypertensive Combinations Aminopenicillins Bone density regulators Naturietic Peptide Rales Diuretic Combinations S3Gallop NSAIDS

Category Diagnosis Medication Lab Symptom

47Jimeng Sun, Large-scale Healthcare Analytics

Patient Similarity

Healthcare Analytics

Information Extraction

Data Mining Visualization

Structured EHR

Feature extraction Feature

extraction

Patient representation

Unstructured EHR

Feature selection Feature selection

Patient similarity Patient

similarity

ContextContext

VisualizationVisualization

48Jimeng Sun, Large-scale Healthcare Analytics

Patient Similarity through Locally Supervised Metric Learning

Query patient

Jimeng Sun, Fei Wang, Jianying Hu, Shahram Edabollahi: Supervised patient similarity measure of heterogeneous patient records. SIGKDD Explorations 14(1): 16-24 (2012)

Under a specific clinical context

49Jimeng Sun, Large-scale Healthcare Analytics

Patient Similarity through Locally Supervised Metric Learning

Query patient

Homogeneous neighborhood Heterogeneous

neighborhood

• Homogeneous neighbors: true positives • Heterogeneous neighbors: false positives

Jimeng Sun, Fei Wang, Jianying Hu, Shahram Edabollahi: Supervised patient similarity measure of heterogeneous patient records. SIGKDD Explorations 14(1): 16-24 (2012)

Under a specific clinical context

50Jimeng Sun, Large-scale Healthcare Analytics

Patient Similarity through Locally Supervised Metric Learning

Query patient

• Shrink homogeneous neighborhood • Grow heterogeneous neighborhood

Jimeng Sun, Fei Wang, Jianying Hu, Shahram Edabollahi: Supervised patient similarity measure of heterogeneous patient records. SIGKDD Explorations 14(1): 16-24 (2012)

Under a specific clinical context

51Jimeng Sun, Large-scale Healthcare Analytics

Locally Supervised Metric Learning (LSML)

Goal: Learn a generalized Mahalanobis distance for a specific clinical context (target label)

Jimeng Sun, Fei Wang, Jianying Hu, Shahram Edabollahi: Supervised patient similarity measure of heterogeneous patient records. SIGKDD Explorations 14(1): 16-24 (2012)

Maximize the total margin

Total distance to heterogeneous neighbors

Margin for xi

Total distance to homogeneous neighbors

Homogeneous neighborhood for xi

Heterogeneous neighborhood for xi

56Jimeng Sun, Large-scale Healthcare Analytics

Patient Similarity Experiment Design

20 0K

s am

pl es

195 features

Disease Target samples D1 DM with Acute Complications 4,392 D2 DM without Complications 10,734 D3 Depression 6,794 D4 Heart Failure 5,262 D5 Asthma 6,606 D6 Lung Cancers 1,172

Cases

Controls D6

57Jimeng Sun, Large-scale Healthcare Analytics

Prediction Results on Patient Similarity  Baselines:

– EUC: Euclidean distance

– PCA: Principal component analysis

– LDA: Linear discriminant analysis

 Observations:

– LDA does not perform well, because of the resulting dimensionality is too low

– LSML algorithm performs the best among all

DM with Acute Complications

DM without Complications

Depression Congestive Heart Failure

Asthma Lung Cancers

Euclidean 0.539 0.638 0.609 0.688 0.602 0.645

LDA 0.541 0.604 0.589 0.564 0.584 0.595

PCA 0.57 0.639 0.625 0.697 0.617 0.664

LSML 0.576 0.669 0.632 0.723 0.625 0.677

58Jimeng Sun, Large-scale Healthcare Analytics

Clinical Relevancy of Patient Similarity Results

 Retrieve the top ranked comorbidities among similar patients

 over 80% of those are considered as relevant to target disease

possible possible

possible

possible possible possible

no no no no no

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

DM with Acute Complica ons

DM without Complica ons

Depression Heart Failure Asthma Lung Cancers

yes possible no

59Jimeng Sun, Large-scale Healthcare Analytics

Summary on Patient Similarity

 LSML learns a customized distance metric

 Extension 1: Composite distance integration (Comdi) [1]

– How to combine multiple patient similarity measures?

 Extension 2: Interactive metric update (iMet) [2]

– How to update an existing distance measure?

1. Fei Wang, Jimeng Sun, Shahram Ebadollahi: Integrating Distance Metrics Learned from Multiple Experts and its Application in Inter-Patient Similarity Assessment. SDM 2011: 59-70 56

2. Fei Wang, Jimeng Sun, Jianying Hu, Shahram Ebadollahi: iMet: Interactive Metric Learning in Healthcare Applications. SDM 2011: 944-955

60Jimeng Sun, Large-scale Healthcare Analytics

Visualization

Healthcare Analytics

Information Extraction

Data Mining Visualization

Structured EHR

Feature extraction Feature

extraction

Patient representation

Unstructured EHR

Feature selection Feature selection

Patient similarity Patient

similarity

ContextContext

VisualizationVisualization

61Jimeng Sun, Large-scale Healthcare Analytics

Visualization

FacetAtlas (InfoVis’10) SolarMap(ICDM’11)

DICON (InfoVis’11) MatrixFlow (AMIA’12)

62Jimeng Sun, Large-scale Healthcare Analytics

Conclusions

 Scalable healthcare analytic research platform

– Enables efficient collaboration across domains

– Extensible analytic platform

– Intuitive analytic results

– Scalable computation engine

Information Extraction

Data Mining Visualization

Structured EHR

Feature extraction Feature

extraction

Patient representation

Unstructured EHR

Feature selection Feature selection

Patient similarity Patient

similarity

ContextContext

VisualizationVisualization

63Jimeng Sun, Large-scale Healthcare Analytics

Future of Healthcare Analytics

Research Challenges

 Data analytic techniques for each data modality

 Privacy preserving data sharing

 Visual analytic techniques

Genomic data

Clinical data

Behavior data

Social data

Health Analytic Apps

Heart disease predictor for $5.99 Analytic cloud

Privacy engine

Visualization

User

Data miner