healthcare predicitive analytics for risk profiling in chronic care: a bayesian multitask learning...

Healthcare Predictive Analytics for Risk Profiling in Chronic Care: A Bayesian Multitask Learning Approach

Yu-Kai Lin (Flor ida State Universi ty)

Hsinchun Chen (Universi ty of Ar izona)

Randal l A. Brown (Southern Ar izona VA Health Care System)

Shu-Hsing Li (Nat ional Taiwan Universi ty)

Hung-Jen Yang (Stanford Universi ty)

5/27/2017 1Healthcare Predictive Analytics for Risk Profiling in Chronic Care

Background

Healthcare Predictive Analytics for Risk Profiling in Chronic Care: A Bayesian Mult itask Learning Approach

5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 2

How to improve chronic care?


Chronic Care ModelBodenheimer et al. (2002)

Chronic Disease ControlBrownson and Bright (2004)

“Technovigilance”Dixon-Woods et al. (2013)

• Community Resources and Policies

• Health Care Organization

• Self-management Support

• Delivery System Design• Decision Support

• Clinical Information Systems

Data and science-driven

decision-making

If one consistent message has

emerged from the literature on

improving quality and safety in

health care, it is that high-quality

intelligence is indispensable.

Healthcare analytics for clinical intelligence


Healthcare Analytics ⇔ Business Analytics in Healthcare

• Marketers vs. Clinicians (Fichman et al. 2011)

− Marketers:

◦ Consumer profiling for targeted marketing

◦ How likely a particular consumer will click an ad link, download an app, respond to a coupon, …

− Clinicians:

◦ Patient profiling for personalized care

◦ How likely a particular patient will develop a complication, experience an adverse medical event, respond to a treatment, …

Health analytics using EHR data


• Healthcare predictive analytics using electronic health records (EHRs) is a promising IS research direction

• Fichman, Kohli & Krishnan (ISR 2011: Healthcare IS)

“Using digital technology to enable

new kinds of mathematical

healthcare modeling … and how

they should be integrated with

electronic health records warrants

future research attention.”

• Chen, Chiang & Storey(MISQ 2012: BI & Analytics)

“Over the past decade, electronic

health records have been widely

adopted in hospitals and clinics

worldwide. Significant clinical

knowledge and a deeper

understanding of patient disease

patterns can be gleaned from such

collections.”

Research motivation


• Chronic Diseases (CDs) in the US

− Half of all adults (117 million people) had one or more CDs

− 86% of the nation’s health care costs are for treating CDs

− Seven of the top 10 causes of death in 2010 were CDs

− Maps on the prevalence of diagnosed diabetes:

1994 2000

<4.5% 4.5%–5.9% 6.0%–7.4% 7.5%–8.9% >9.0%

2014

Data source: https://www.cdc.gov/chronicdisease/

Research motivation (cont.)


• Patients with chronic diseases are often at risk for multiple complications

− Diabetes � stroke, heart attack, kidney failure, eye problems, and so on

• Surprisingly, almost all the existing clinical risk models are designed to focus on only a single outcome.

− Diabetes � stroke

− Diabetes � heart attack

− Diabetes � kidney diseases

Single-Task Learning (STL)


• In the usual approach, each event is modeled independently:

Predict Event 1

Predict Event 2

Predict Event K

logit �� = � + ��

��

logit �� = � � + ��

��…

logit �� = � � + ��

��

STL is fine, but…


• If the outcomes are related, the models are perhaps related (that is, coefficients/parameters are related).

• If the models are related, a model can perhaps “borrow” information from the other models.

� Spillover effect in model training

Multitask Learning (MTL)


Task 1

Task 2

Task K

Data

Data

Data

Training

Training

Training

Trained

Model

Trained

Model

Trained

Model

… … …

Task 1

Task 2

Task K

Data

Data

Data

Training

Trained

Model

Trained

Model

Trained

Model

… …

Single-Task Learning Multitask Learning

Research questions & relevance


• We are interested in studying:

− What are a patient’s risks to an array of events?

− How can we model multiple risks simultaneously?

− Does simultaneous learning of multiple event risks improve overall predictive performance of each event risk?

• Relevance to Information Systems (IS) research

− Healthcare IS (Fichman et al. 2011; Bardhan et al. 2014)

− Predictive analytics (Shmueli and Koppius 2011)

− Design science (Hevner et al. 2004; Gregor and Hevner 2013)

Model Development



Bayesian MultiTask Learning (BMTL) intuition


Ω1, … , ΩJ are K by K correlation matrices.

Ω1

(1)

Ω2 ΩJ

(1) (1)

(2) (2) (2)

(K) (K) (K)

...Task 1

Task 2

Task K

β1

β1

β1

β2

β2

β2

βJ

βJ

βJ

...

...

...

...

Bayesian Analysis


� �|� = � �, �� = � �|� � �

� �� |� ∝ � �|� � �

• � is a random variable of interest; D is observed data

• We have a prior subjective belief about � � � �[as a part of model specification]

• We update our prior belief with the data to form posterior beliefs about � � � �|�[at a result of model fitting]

Bayes Rule:

Model Spec.


Para-

meter

Distribution/Function Form and

Supporting Reference

�� Cauchy distribution (Gelman et al. 2008):

��~Cauchy�0, 10� � Multivariate normal (MVN) distribution with

the horseshoe prior (Carvalho et al. 2010):

� = �� , �� , … , ��"

�~MVN $, %��Σ�%� Horseshoe prior (Carvalho et al. 2010):

%� = '�('� , (~Half-Cauchy�0,1�

Σ� Covariance matrix (Barnard et al. 2000)

Σ� = diag +� ∗ Ω� ∗ diag +�+� = .� , .� � , … , .� �

"

.� � Half-Cauchy distribution (Gelman et al. 2008):

.� � ~ Half-Cauchy�0, 2.5�Ω� Lewandowski, Kurowicka and Joe (LKJ, 2009):

Ω�~LKJ 2, 1

Note for the table: The index j ranges from 1 to J

(the total number of predictors), and the index k

ranges from 1 to K (the total number of tasks).

JK

σ Ω

Σ

β

τ

rψ

K

θ

N

αK

Model fitting


• No-U-Turn Sampler (NUTS; Hoffman and Gelman 2014)

− A variant of Hamiltonian Monte Carlo (HMC)

− Adaptively sets the algorithmic parameters in HMC

• 2 Markov chains, 1000 warm-up draws, 1000 sampling draws

• For each parameter, convergence was assessed using Gelman and Rubin’s (1992) diagnostic test, a.k.a., 34statistic, with the value less than 1.2.

Evaluations



Illustration of experiment design


v1 v2 v3 v4 v5 v6 v7 v8 Time

Step 1: Randomly sample a visit from the first half of the patient’s medical history.


Step 2: Use information available in and before the sampled visit for learning or prediction.


Step 3: Learn and predict if an event will happen in the next w years.

Event occurrence

Sampled visit, denoted by v0i

Summary of data


• 14,752 adult patients with type 2 diabetes

• Events / Complications

− stroke (henceforth denoted by STK), acute myocardial infarction (AMI; aka, heart attack), and acute renal failure (ARF)

− Exclude patients with all three events occurred before v0i

Before v0i

During v0i and v0i + w yearsAfter v0i + 5

w = 1 w = 2 w = 3 w = 4 w = 5

STK 1507 354 560 685 793 828 47

AMI 485 75 146 178 210 225 20

ARF 410 217 399 488 536 571 37

Variables in our analysis


• 179 variables in total

− Missing values are imputed using the column mean

• Examples

Category Example Variables

Patient information Age, body weight, male, smoking

Diagnoses Three digit ICD-9 codes, e.g., 401 for essential

hypertension and 427 for cardiac dysrhythmias

Treatments Aspirin, clopidogrel, insulin, isoket, metformin

Labs and exams CT scan, low-density lipoprotein cholesterol, serum

creatinine, systolic blood pressure

Note: ICD-9=International Classification of Diseases, Ninth Revision

Three sets of evaluations


1. BMTL vs. single task learning approaches

2. BMTL vs. other multitask learning approaches

3. Counterfactual analysis of practical use

Evaluations 1 and 2


• 10-fold cross validation

• Area Under the Curve (AUC)

− Ranges from 0.5 (a worthless model) to 1 (a perfect model)

− The DeLong test of AUC (DeLong et al. 1988)

Testing data

Training data

Fold 1 Fold 2 Fold 3 Fold 10

……

Ori

gin

al

Da

ta

Tru

e P

ositiv

e R

ate

False Positive Rate

Evaluation 1 (AUC; 10-fold CV)BMTL vs. STL approaches


Window

(w)

Task

(k)

Models

BMTL-Logit B-Logit Logit Logit-lasso

1 STK 0.747 0.725*** 0.723*** 0.735***

1 AMI 0.778 0.744*** 0.729*** 0.758**

1 ARF 0.863 0.855* 0.847** 0.849***

3 STK 0.742 0.724*** 0.722*** 0.728***

3 AMI 0.736 0.703*** 0.699*** 0.704***

3 ARF 0.833 0.823*** 0.819*** 0.823***

5 STK 0.739 0.724*** 0.723*** 0.727***

5 AMI 0.727 0.699*** 0.698*** 0.704***

5 ARF 0.820 0.812*** 0.809*** 0.814***

Note. Bolded values highlight the best AUC result in a row.

*** The AUC result is statistically significantly different from BMTL-Logit at α = 0.01.

** The AUC result is statistically significantly different from BMTL-Logit at α = 0.05.

* The AUC result is statistically significantly different from BMTL-Logit at α = 0.1.

Evaluation 2 (AUC; 10-fold CV)BMTL vs. other MTL approaches


Window

(w)

Task

(k)

Models

BMTL-Logit

(this study)

MTL-Logit

(Huang et al. 2012)

MTL-Tree

(Simm et al. 2014)

MTL-ANN

(Caruana 1997)

1 STK 0.747 0.746 0.717** 0.660***

1 AMI 0.778 0.767* 0.737** 0.686**

1 ARF 0.863 0.849* 0.831*** 0.650***

3 STK 0.742 0.730** 0.702*** 0.677***

3 AMI 0.736 0.693*** 0.727* 0.680***

3 ARF 0.833 0.816*** 0.787*** 0.763***

5 STK 0.739 0.719*** 0.686*** 0.670***

5 AMI 0.727 0.705** 0.692** 0.653***

5 ARF 0.820 0.809*** 0.770*** 0.703***

Note. Bolded values highlight the best AUC result in a row.

*** The AUC result is statistically significantly different from BMTL-Logit at α = 0.01.

** The AUC result is statistically significantly different from BMTL-Logit at α = 0.05.

* The AUC result is statistically significantly different from BMTL-Logit at α = 0.1.

Evaluation 3Counterfactual analysis


• Prediction of risk is not enough—we need evidence that

prediction can lead to actions that reduce risk beyond what

would occur without the prediction rule. (Grady and Berkowitz 2011)

− How to assess the practical value of a predictive model without actual use?

− Assumption for our counterfactual analysis:

Physicians will always provide guideline-recommended

preventive interventions if they believe a patient has a high

risk of STK/AMI/ARF.

− Among the positive cases (patients with the STK/AMI/ARF events between v0i and v0i + 5 years), what happened to them and what could happen to them given a prediction rule.



• Guideline-recommended preventive treatments

− Source: “Diabetes Comprehensive Care Plan Guidelines” from the American Association of Clinical Endocrinologists

Comorbidity Preventive Treatment

STK • Antihypertensive agents

• Antiplatelet therapy

AMI • Antihypertensive agents

• Antiplatelet therapy

• Lipid lowering therapy

ARF • Antihypertensive agents

• Angiotensin receptor blockers

• Angiotensin-converting-enzyme inhibitors



• From the positive cases, we are interested in the proportions who

− actually received preventive interventions at or before v0i?

− potentially could receive preventive interventions at v0i, given model predictions?

• Practically useful models: small c and large d

Predicted Risk

(from some model)

Low High

Preventive treatment

prescribed at/before v0i

Yes a b

No c dHigh/low risk cutoff level:

2% per year � 10% over 5 yrs

(Dhamoon and Elkind 2010)

Evaluation 3Summary of results


A lower c (making fewer mistakes) is better

A higher d (supporting physicians) is better

Conclusions



Conclusions


• The proposed BMTL approach outperforms the alternative models in risk profiling, and could support physicians to identify high risk patients.

• Multitask learning improves overall learning performance by sharing information across models

− Evidence for the spillover effect in model training

• Beyond healthcare

Practical implicationsRisk profiling in chronic care


• Hospitals: new healthcare delivery models

− Accountable care organizations; bundled payments

• Physicians: decision support at the point of care

− To err is human

• Patients: healthcare spending and # of conditions

− Medical Expenditure Panel Survey

To error is human


• Building a better health system with IT and analytics

Thank you

Please send comments to Yu-Kai Lin

[email protected]

5/27/2017 33Healthcare Predictive Analytics for Risk Profiling in Chronic Care

healthcare predicitive analytics for risk profiling in chronic care: a bayesian multitask learning...

Technology