Healthcare Predictive Analytics for Risk Profiling in Chronic Care: A Bayesian Multitask Learning Approach Yu-Kai Lin (Florida State University) Hsinchun Chen (University of Arizona) Randall A. Brown (Southern Arizona VA Health Care System) Shu-Hsing Li (National Taiwan University) Hung-Jen Yang (Stanford University) 5/27/2017 1 Healthcare Predictive Analytics for Risk Profiling in Chronic Care
33
Embed
Healthcare Predicitive Analytics for Risk Profiling in Chronic Care: A Bayesian Multitask Learning Approach
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
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?
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 3
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
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 4
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
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 5
• Healthcare predictive analytics using electronic health records (EHRs) is a promising IS research direction
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
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 21
1. BMTL vs. single task learning approaches
2. BMTL vs. other multitask learning approaches
3. Counterfactual analysis of practical use
Evaluations 1 and 2
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 22
• 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
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 23
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
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 24
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
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 25
• 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.
Evaluation 3Counterfactual analysis
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 26
• 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
Evaluation 3Counterfactual analysis
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 27
• 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
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 28
A lower c (making fewer mistakes) is better
A higher d (supporting physicians) is better
Conclusions
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 29
Conclusions
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 30
• 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
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 31
• 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
5/27/2017 Healthcare Predictive Analytics for Risk Profiling in Chronic Care 32
• Building a better health system with IT and analytics