[IEEE 2012 Annual SRII Global Conference (SRII) - San Jose, CA, USA (2012.07.24-2012.07.27)] 2012...

8
Towards Personalized Care Coordination Service Integrating Guideline-based Practice with Interactive Data-driven Analytics Pei-Yun Hsueh IBM T.J. Watson Research Center Hawthorne, USA [email protected] Sreeram Ramakrishnan IBM T.J. Watson Research Center Hawthorne, USA [email protected] AbstractTransforming clinical requirements into personalized care plans is a challenging task. Existing care coordination services account for the most common physiology -based needs addressed in clinical guidelines, but do not discern different individuals’ unique needs. Moreover, the individual difference-conferring signals are not explicitly recorded in patient health records and often can only be captured by integrating data sources obtained from a multitude of service providers. The quest for better personalized care coordination services leads to the exploration of two dimensions: what are the properties of tailoring matrix to personalize care coordination plans, and how to use the tailoring matrix to improve the relevance of care plans and provide individual feedbacks. In this paper, we present a personalization service framework that accounts for the two dimensions, leveraging the longitudinal records of other cohort patients whom have been dynamically identified as exhibiting patterns similar to the patient in focus. We summarize the design rationale and overall operation of the framework as well as details of the interactive analytics components as to how to capture individual differences and optimize expected utility of the coordinated care services in a continuous feedback loop. Keywords- personalization; sustainable operation; care coordination; interactive analytics; longitudinal health records I. INTRODUCTION Existing care coordination services account for the most common physiology-based needs addressed in clinical guidelines, but do not discern different individuals’ unique needs. Two main challenges have been observed arising. First, in the consultation stage, the lack of patient focus leads to gaps between practice and evidence. Worldwide healthcare quality surveys [1, 2] have shown that only 30- 40% of cases conform to latest evidence, rendering 20-50% of the care as either not needed or potentially harmful. Secondly, in the follow-up stage, the lack of patient-centric feedback and further guidance in extended care settings leave patients in a position follow the recommended care in a free style fashion. Intervention adherence has thus become an insidious problem on both sides of providers and patients. In fact, in existing practices only 50-55% of patients are found to adhere to recommendations after going home [3, 4]. These challenges together complicate care coordination processes and result in waste of resources and ineffectiveness of care plans. Reportedly, one in five admitted patients suffers from preventable readmissions [5]. In the US Medicare system alone this represents $100 billion of its $555 billion budget (as of 2011). Re-hospitalization also increases the risk of health complications, resulting in greater functional and cognitive impairment for patients. In the category of cardiovascular diseases, the problems lead to at least 125,000 unnecessary deaths every year in US [6]. Recently, regulatory measures have been imposed to penalize providers for high readmission rates with the aim of decreasing preventable readmissions. However, without capabilities to discern individual differences, it is difficult to bridge the gaps between best evidence and best practice in the consultation stage, let alone the insidious adherence problem in the follow-up stage. In this paper, we first focus on learning tailoring matrices for the task of transforming patients’ clinical requirements into their personalized service plans. The task brings challenges to existing service providers, as clinical requirements and patient health records did not explicitly indicate the individual differences key to personalization. Moreover, customization points needed to be personalized can often be revealed when integrating data from a multitude of service providers whom, when working together, create the needed continuous service cycle to sustain behavioral change on a daily basis. The quest of a better personalization scheme, which can work in a multi-provide environment typical in care coordination, leads to the exploration of two dimensions: (1) what are the properties of tailoring matrix to personalize care coordination plans, and (2) how to use the tailoring matrix to improve the relevance of care plans and provide individualized feedbacks. In the rest of the paper, we present a personalization service framework that includes data-driven analytic and interaction components that can infer proxy measures to indicate individual differences that are not explicitly recorded, and solicit necessary inputs to tailor personalized plan recommendations. In Section II-V, we introduce the framework from three perspectives: service, user, and system design. In section II, we summarize the design rationale and overall operation of the service framework; in Section III we demonstrate how this works in a user scenario; in Section IV and V, we detail the interactive analytics components as to how to capture individual differences that can help optimize the expected utility of coordinated care service plans and how to put these together in a system to provide sustainable personalized services. 2012 Service Research and Innovation Institute Global Conference 978-0-7695-4770-1/12 $26.00 © 2012 IEEE DOI 10.1109/SRII.2012.35 244 2012 Service Research and Innovation Institute Global Conference 978-0-7695-4770-1/12 $26.00 © 2012 IEEE DOI 10.1109/SRII.2012.35 250

Transcript of [IEEE 2012 Annual SRII Global Conference (SRII) - San Jose, CA, USA (2012.07.24-2012.07.27)] 2012...

Page 1: [IEEE 2012 Annual SRII Global Conference (SRII) - San Jose, CA, USA (2012.07.24-2012.07.27)] 2012 Annual SRII Global Conference - Towards Personalized Care Coordination Service: Integrating

Towards Personalized Care Coordination Service Integrating Guideline-based Practice with Interactive Data-driven Analytics

Pei-Yun Hsueh IBM T.J. Watson Research Center

Hawthorne, USA [email protected]

Sreeram Ramakrishnan IBM T.J. Watson Research Center

Hawthorne, USA [email protected]

Abstract— Transforming clinical requirements into personalized care plans is a challenging task. Existing care coordination services account for the most common physiology -based needs addressed in clinical guidelines, but do not discern different individuals’ unique needs. Moreover, the individual difference-conferring signals are not explicitly recorded in patient health records and often can only be captured by integrating data sources obtained from a multitude of service providers. The quest for better personalized care coordination services leads to the exploration of two dimensions: what are the properties of tailoring matrix to personalize care coordination plans, and how to use the tailoring matrix to improve the relevance of care plans and provide individual feedbacks. In this paper, we present a personalization service framework that accounts for the two dimensions, leveraging the longitudinal records of other cohort patients whom have been dynamically identified as exhibiting patterns similar to the patient in focus. We summarize the design rationale and overall operation of the framework as well as details of the interactive analytics components as to how to capture individual differences and optimize expected utility of the coordinated care services in a continuous feedback loop.

Keywords- personalization; sustainable operation; care coordination; interactive analytics; longitudinal health records

I. INTRODUCTION

Existing care coordination services account for the most common physiology-based needs addressed in clinical guidelines, but do not discern different individuals’ unique needs. Two main challenges have been observed arising. First, in the consultation stage, the lack of patient focus leads to gaps between practice and evidence. Worldwide healthcare quality surveys [1, 2] have shown that only 30-40% of cases conform to latest evidence, rendering 20-50% of the care as either not needed or potentially harmful. Secondly, in the follow-up stage, the lack of patient-centric feedback and further guidance in extended care settings leave patients in a position follow the recommended care in a free style fashion. Intervention adherence has thus become an insidious problem on both sides of providers and patients. In fact, in existing practices only 50-55% of patients are found to adhere to recommendations after going home [3, 4].

These challenges together complicate care coordination processes and result in waste of resources and ineffectiveness of care plans. Reportedly, one in five admitted patients

suffers from preventable readmissions [5]. In the US Medicare system alone this represents $100 billion of its $555 billion budget (as of 2011). Re-hospitalization also increases the risk of health complications, resulting in greater functional and cognitive impairment for patients. In the category of cardiovascular diseases, the problems lead to at least 125,000 unnecessary deaths every year in US [6]. Recently, regulatory measures have been imposed to penalize providers for high readmission rates with the aim of decreasing preventable readmissions.

However, without capabilities to discern individual differences, it is difficult to bridge the gaps between best evidence and best practice in the consultation stage, let alone the insidious adherence problem in the follow-up stage. In this paper, we first focus on learning tailoring matrices for the task of transforming patients’ clinical requirements into their personalized service plans. The task brings challenges to existing service providers, as clinical requirements and patient health records did not explicitly indicate the individual differences key to personalization. Moreover, customization points needed to be personalized can often be revealed when integrating data from a multitude of service providers whom, when working together, create the needed continuous service cycle to sustain behavioral change on a daily basis. The quest of a better personalization scheme, which can work in a multi-provide environment typical in care coordination, leads to the exploration of two dimensions: (1) what are the properties of tailoring matrix to personalize care coordination plans, and (2) how to use the tailoring matrix to improve the relevance of care plans and provide individualized feedbacks.

In the rest of the paper, we present a personalization service framework that includes data-driven analytic and interaction components that can infer proxy measures to indicate individual differences that are not explicitly recorded, and solicit necessary inputs to tailor personalized plan recommendations. In Section II-V, we introduce the framework from three perspectives: service, user, and system design. In section II, we summarize the design rationale and overall operation of the service framework; in Section III we demonstrate how this works in a user scenario; in Section IV and V, we detail the interactive analytics components as to how to capture individual differences that can help optimize the expected utility of coordinated care service plans and how to put these together in a system to provide sustainable personalized services.

2012 Service Research and Innovation Institute Global Conference

978-0-7695-4770-1/12 $26.00 © 2012 IEEE

DOI 10.1109/SRII.2012.35

244

2012 Service Research and Innovation Institute Global Conference

978-0-7695-4770-1/12 $26.00 © 2012 IEEE

DOI 10.1109/SRII.2012.35

250

Page 2: [IEEE 2012 Annual SRII Global Conference (SRII) - San Jose, CA, USA (2012.07.24-2012.07.27)] 2012 Annual SRII Global Conference - Towards Personalized Care Coordination Service: Integrating

II. FROM CLINICAL REQUIREMENTS TO

PERSONALIZATION SERVICE PLANS

To accommodate clinical requirements into patient care plan, in the past clinical practice guidelines (CPGs) have been introduced to translate new research findings into practice [7]. Researchers have developed computer-based programs to encode CPGs into electronic forms for later processing and better dissemination. Some of them include guideline-based decision support systems that can be used by clinicians to make differential diagnosis and patient-centric recommendations [8].

While the CPG-based research works focus on supporting the consultation stage, previous clinical trials have shown the value of personalized plans and feedbacks for chronic disease management in the follow-up stage. For example, Finland’s D2D, US’s DPP and UKPDS have all been employing certain types of manual processes to personalize feedbacks for diabetes management [12, 13, 14]. Three major issues have been identified:

• Systematically compare the benefits of various plans (i.e., configuration of service options);

• Accommodate individual differences; • Account for multiple criteria in the wellness decision

making process for making recommendations. Despite the potential of integrating computer-based CPG

content into clinicians’ workflows [9, 10], it is not sufficient to automate these key factors. More system support is therefore needed to sustain patient adherence to the recommended extended care in the follow-up stage[11]. In order to understand the needed support for sustaining such personalization tasks from the service perspective, in this section we examine the as-is service flow in current practice, gather requirements, and propose a to-be service flow.

A. As-Is Service Flow

Fig. 1 illustrates what takes place in the consultation stage: First, clinicians specify clinical requirements for what needs to be taken care of in extended care settings, based on a patient’s longitudinal records. Following the prescribed clinical requirements, care coordinators or patients themselves then determine the needed interventions at the time of care. One major deficiency in the as-is service flow is the lack of system support in personalization. There is currently no additional patient-specific information to guide coordinators and patients in the process. Nor is there adherence feedback channel for the clinicians and coordinators to provide their advices during the care coordination process.

Figure 1. As-Is Service Flow

B. To-Be Service Flow

To remedy the lack of system support, the design rationale of the personalization service framework is thus two-fold: first, to identify the customization points needed to offer computation, and then to provide computation for driving the dynamic generation of context-dependent, personalized plan and feedback. In response to the first part of design rationale, Fig. 2 shows the proposed service flow, which adds a patient-centric focus to the as-is flow in each possible customization point, going across the different stages of a personalized care plan: individualized guideline generation, user profile building, personalized plan generation, adherence risk mitigation, and personalized adherence follow-up plan adaptation. This has impacts on wellness service providers and healthcare providers who would like to become more accountable under new regulations and work with users in the extended care scenarios. In fact, this is applicable to many other stakeholders, such as payers, direct primary care providers, and ICT industry partners, whose market would be impacted by disruptive technologies that bring together large networks of engaged users and healthcare professionals.

Figure 2. To-Be Service Flow with Personalization Service Framework

C. Personalization Service System Requirements

The second part of design rationales is especially important when there are more than one service options available at a decision point. Specifically, this part of design involves capabilities to answer the following questions: First, how to craft the knowledge of a patient’s past history (from medical and behavioral history in the longitudinal records) and current status (clinical requirements from healthcare professionals) into actionable information that can be used to drive the generation of individualized guideline and personalized feedback?

Then, given the actionable information inferred from longitudinal records for a patient and the channel to interact with the patient to solicit input on personal constraints and preferences, how to systematically tailor clinical requirements into this patient’s individualized guideline?

These questions require the personalization service framework to incorporate features that can facilitate such sorts of personalization computation. Below are the three key

245251

Page 3: [IEEE 2012 Annual SRII Global Conference (SRII) - San Jose, CA, USA (2012.07.24-2012.07.27)] 2012 Annual SRII Global Conference - Towards Personalized Care Coordination Service: Integrating

features needed for supporting the three major issues important to the coordination of a personalized care plan.

• Risk Stratification, which provides the basis of comparing various plans at the sub-population level

• by inferring patient status through proxy measures (e.g., personal disease risk levels to be watched as specified in the clinical requirements);

• Interactive User Profiling, which accommodates individual differences by including users in the loop and capturing the differences in a personalization matrix;

• User-adaptive service selection and composition, which accounts for context-aware recommendation via personal multi-criteria decision making. This is achieved by calculating the expected utility of each plan at the personal level and recommending the top N plans that yield maximum utility.

With the three major features established, the system can then be used to add a layer of patient-centric focus to any existing professional-initiated services.

III. USER SCENARIO

In this subsection, we describe the scenario of individualized guideline generation, wherein a case manager works with the patient of focus in the consultation stage at the point of care to tailor the clinical requirements with respect to user need and context.

A. Personal Wellness Decision Support Program

Imagine the following use case. Mary was diagnosed with Type II Diabetes years ago. She is a career woman with two kids, working in a high-stressed job in international banking. She has lactose resistance to dairy food and an allergy to shellfish. She likes jogging but not swimming. She is often on-the-go with the need of attending business meals or eating in the airport so as to manage her nutrition intake a challenging task.

Recently, due to the pressure from a family tragedy, she has fallen behind her diabetes self-management plan, resulting in high HbA1c, weight gain, and an increasing risk of developing hypertension. In addition, she suffered a few dangerous hyperglycemia episodes.

Her family physician, Janet, realizes that Mary needs personalized consultation and signs her up in a personal wellness decision support (PWDS) program, which aims to help mid-career professionals to adjust their service plans with respect to their clinical and lifestyle need for better adherence and outcome. The program aims to help users at a recent rising risk who would like to work out a personal wellness management plan that can fit into their busy schedule. This sort of users prefers frictionless self management, but still has lifestyle factors they could not fully control or habits they would not give up. A case manager (e.g., community nurse) would be assigned to the participating patient’s case to handle wellness decision making and tailor individualized guidelines.

B. Questions posed by patients and stakeholders

In the consultation stage, this kind of scenario incurs questions from different stakeholders that cannot be answered in the current (as-is) practice. The patient would like to know what she should do or should not do; the case manager assigned to the case needs to know what factors to focus efforts to; and the physician needs to know what seems to work and when to change interventions.

Take Mary for example. In her situation, she would like to know the consequence of her inactions towards the self-management plan, whether there are any adjustments she should make at this point, and which new service plan (e.g., diet) will meet her clinical need without compromising her busy schedule and preferences. For Mary’s case manager, it is important to identify what are the prominent risk factors for Mary under the previous plan and what are the impacts of modifying one or more of these factors. In addition, Mary’s physician will also need information in the next visits about what service plan is working for Mary and be informed when there is a need to review her clinical requirements between her regular checkup visits.

C. Individualized guideline generation

Reflection on these questions suggest that the answers commonly require systematic comparison of available intervention plans, along with the capabilities of accommodating individual preferences and constraints (based on lifestyle need) and finding a balanced plan when there is more than one decision criteria involved. These requests are aligned with current thinking of our proposed personalization service framework (as shown in Fig. 2). The process of generating individualized guideline goes from the stage of clinical requirement to user profiling and ends with personalized service planning. Three additional steps are added to the as-is process in the to-be service flow:

1) Step 1 (Clinical requirement-based guideline refinement): Physician initiates predictive analytics support. In this step, the system scans through previous records to infer systematic measures of personal risk as the basis to compare various interventions at the sub-population level. In Mary’s case, the prediction results are used to generate an online HRA and compliance profile of Mary. Her physician can then comment on the risks to be watched and factors that have been shown as challenging to the previous plan. For example, her previous high-carb diet has caused her several hyperglycemida episodes.

2) Step 2 (Interactive user profiling): Case manager uses interactive tools to review clinical requirements with Patient and refine the prescribed guideline to accommodate both the patient’s clinical and lifestyle need. In this step, the system includes users in the loop and captures the differences in user models. This will provide initial answers for her questions such as “What are the risk factors I need to take care first?” and “What are in my intervention portfolio?” Then, her case manager solicits her input on what is important to her, selects the risk factors to be

246252

Page 4: [IEEE 2012 Annual SRII Global Conference (SRII) - San Jose, CA, USA (2012.07.24-2012.07.27)] 2012 Annual SRII Global Conference - Towards Personalized Care Coordination Service: Integrating

modified in the guidelines, and assesses the changes in terms of the expected utility with the refined guidelines. The inputs can help address questions like “What are the plans suitable for Mary given her needed risk adjustment?” Based on the input, her case manager can recommend a low-fat, high-carb diet over a low-carb diet for its benefits on preventing hyperglycemia episodes.

3) Step 3 (Professional advising for personalized service planning): To help with multi-criteria decision making, Patient’s profile on expected adherence level and utility is generated based on previous records and inferred status from population databases. In this step, the system adapts plan recommendations according to the expected utility at the personal level with multiple clinical and lifestyle decision criteria considered. Further questions include “What plan can work with Mary’s lifestyle need without compromising her clinical requirements?” Plotting the available personal service plans on the two dimensions, expected degradation of adherence and benefit on hypertension control, to understand the trade-off.

IV. DESIGN OF INTERACTIVE ANALYTICS COMPONENTS

These steps across the process of individualized guideline generation in the care coordination service cycle commonly involve at least three interactive analytics components. First, as the individual difference-conferring signals are not always explicitly recorded in patient health records, there is a need to look for proxy measures of indicative characteristics in patient-specific properties. Second, because the treatment effectiveness is not always universal, it is often necessary to infer a patient’s intervention effectiveness from the longitudinal records of a subset of cohort data. The selection of relevant patient cohorts for analysis can be achieved by either manual configuration or dynamic identification of those who exhibit patterns similar to the patient in focus. Third, to deal with situations where multiple criteria are involved in different decision points, interactive user profiling is developed to identify the source of variation in patient-centric recommendations. Last but not least, a unifying utility quantification framework that can integrate different types of data as personal features as well as personal constraints and preferences is employed to make personalized recommendations.

A. Risk Stratification

1) Data-Driven Risk Factor Analysis (Subpopulation): The personalization tasks involved in the service flow commonly require a systematic measurement of the risk and benefit of service plans. Historically, comparative effectiveness research achieves this by comparing the empirical risk reduction of different plans. For example, [15] compared the outcomes of the cohorts who were treated with individualized guidelines and those with the standard JNC7 high blood pressure guidelines, and showed that the individualized guideline yielded 13% of reduction in 5-year cardiovascular disease (CVD) risk. In the context of care coordination scenario, at the point of care, risk analytics is

extended to analyze cohort data in order to stratify individual risks and to provide health risk assessment (HRA) reports on patient-critical risk factors.

The choice of risk analytics as the intermediate layer of proxy measure is originated from decades of research in how to systematically account for prognostic differences. In fact, many multi-dimensional risk assessment schemes have been developed to predict the onset and progression of chronic diseases. Some estimate risks directly from personal medical records. E.g., Researchers analyzed more than 100,000 users of Veteran Health Administration (VHA) services to understand how to reliably stratify chronic disease risks [16]. Similarly, many others infer risks from self reports. E.g., Researchers of the National Heart Institute Framingham study analyzed the questionnaires of 5,000 participants over three generations to identify risk factors of cardiovascular diseases and extrapolated these factors to estimate risk of individuals for stratification purposes [17].

Risk assessment have also been found as effective for hospital management to lower discharge costs and improve healthcare service delivery quality. E.g., VHA used Chronic Disease Score (CDS) [18] to detect abnormality in healthcare delivery quality. Kaiser Permanente (KP) researchers analyzed the medical data of 1.39 million users to identify statistically significant associations between drug and disease risk. The results show a strong association between high doses of a branded painkiller and a three-fold risk of heart attacks, resulting in the drug being pulled from the market [19]. Other examples include co-morbidity indicators such as Charlson Comorbidity Index [20] and Elixhauser Index [21].

In recent years, more and more risk assessment schemes utilize not only the encoded medical records and self reports, but also external databases such as air pollution, pesticide use, and traffic density. E.g., KP BioBank researchers leverage a well-documented environmental database collected by the North Carolina government to study how environment exposure impacts health outcome and health behavior at a local community level [22].

With accumulation of individual genetic profile data, more and more risk assessment schemes have been proposed. The HapMap researchers examine thousands of profiles to discover disease-indicative genetic variations [23]. In the field of genomics, many molecular analyses are conducted to examine the relationship between individual variations (e.g., differences in genes, gene expression, protein expression, and metabolites) and clinical outcomes (e.g., disease onset and progression). Genome-wide association mining studies are conducted to capture weak signals hidden in a large and noisy feature space. The interactions between gene expressions and other external factors related to lifestyle interventions (e.g., physical activity, nutrition intake) are studied in emerging fields such as nutri-genomics.

In the context of care coordination, Fig. 6 (a) demonstrates the process of using data-driven risk models for risk-indicative factor analysis. The results are user-specific HRA profiles, which consists of patient-specific risk levels (relative to the best-performing and baseline group) and risk factors to be monitored. In those situations where in a patient’s health risk cannot be reliably inferred, an

247253

Page 5: [IEEE 2012 Annual SRII Global Conference (SRII) - San Jose, CA, USA (2012.07.24-2012.07.27)] 2012 Annual SRII Global Conference - Towards Personalized Care Coordination Service: Integrating

interactive risk analytics component is invoked to directly solicit input from the patient on those key attributes that need to be filled. (More details are described in [24].)

2) Risk-driven Patient Segmentation for Systematic Plan Evaluation (Subpopulation): However, these techniques are often designed to capture common risk factors in rare diseases. Yet common diseases, such as obesity and metabolic syndrome, involve complex changes in the expression of thousands of genes across multiple functional categories. New techniques are needed to narrow down on certain factors so as to reach targeted intervention. In the field of nutri-genomics, high-throughput screening techniques are currently being developed to identify “dietary signature”, which is a set of distinctive patterns in nutrients, non-nutritive food components and nutritional regimes that influence protein expression and regulate the progression of metabolic syndrome [25].

Beyond risk stratification, individual variations are also important analysis targets to determine suitable interventions for individuals as the basis of personalized medicine. E.g., understanding how individual genetic variations dictate the effect of physical activity on chronic disease management will have impacts on the choice of interventions [26]. For our task of transforming clinical requirements to personalized plans, it is important to identify “personal wellness signatures”, i.e., a set of patient-specific properties of which the summation of effects significantly impact intervention effectiveness. The identified signatures then lend support to adjust interventions for better outcomes.

In fact, certain risk factors only matter to a subset of individuals. For this reason, we use individual variation analysis to guide not only the selection of risk factors but also that of cohort data for analysis. The dynamic selection process relies on patient segmentation techniques that are sensitive to risk-indicative individual variations in patient-specific properties. While previous patient segmentation research focuses on finding similar patterns among patients for clustering purposes, the newly proposed scheme is tuned to discover quantifiable individual differences in the set of risks prescribed in the clinical requirements for monitoring.

In the context of care coordination, further added to the service framework is a risk-driven patient segmentation component. As illustrated in Fig. 6 (b), the component pinpoints both risk and baseline groups for analyzing risk-indicative feature, �, which represents the wellness signature that differentiates high risk groups from low risk ones. These are important indicators to be used for systematic evaluation of the different service compositions of a care plan.

B. Interactive User Profiling

1) Interaction for Individual Differences (Personal): Personalized plan generation requires to coordinate around multiple criteria with reagrd to individual differences such as personal preferences, as the importance of the criteria varies widely with respect to individual patients’ clinical and lifestyle need. One direct way to resolve this is to include users in the loop by using their input on the

importance of different criteria and the tradeoff between expected benefits and to bootstrap the process of evaluation. Fig. 3 depicts an interactive user profiling component that incorporates data-driven analytics to digest various aspects of patient records. It is consisted of three major subcomponents: (1) Personal wellness risk profiling: Based on users’ longitudinal records, a pre-specified set of models are applied to infer sub-health status and generate HRA reports; (2) Visualization: Based on the inferred risk profile, design templates are invoked to choose an effective presentation strategy and generate interactive widgets; (3) User input solicitation: Users interact with the widgets to assess how the different options perform on risk mitigation and specify the importance of risk factors to be modified within the ranges permitted by their clinical requirements.

Figure 3. Interactive User Profiling

2) Active Scheme for Multi-criteria Decision Making (Personal): The interactive user profiling component is coupled with interactive risk analytics and risk-driven patient segmentation in the proposed scheme. This scheme is “active” because we expect to find personal wellness signatures to be dynamically changing with respect to changes in context and so are the selection criteria of risk factors and cohort data for analysis.

An additional issue facing the active scheme is inter-item comparison, e.g., the trade-off between risk reduction and adherence degradation. User profiles on inter-item evaluative importance are often non-existent in the first place and tedious to build from scratch. To bootstrap the user profiling process, it is beneficial to leverage similar users’ inputs. Memory-based collaborative filtering approaches are applied to aggregate importance scores from the top similar users of a patient to infer his score. In addition, the most similar users’ behavioral records are utilized to learn the adherence likelihood of the target patient on different service options, which we do not have records for.

Having built the user profiles of adherence likelihood and importance score, the inter-item evaluative importance can then be plotted to facilitate professional advising of personalized care planning.

248254

Page 6: [IEEE 2012 Annual SRII Global Conference (SRII) - San Jose, CA, USA (2012.07.24-2012.07.27)] 2012 Annual SRII Global Conference - Towards Personalized Care Coordination Service: Integrating

C. User-Adaptive Service Selection and Composition

1) Utility Quantification for User-adaptive Outcome-based Recommendation (Personal): To gear towards in-context recommendation for achieiving the best outcome, personalized plan generation services need to evaluate the best course of action for cases which involve more than one criterion for outcome evaluation and changing contexts.

The first part of resolving multiple personal outcome evaluation criteria is achieved by learning the empirical benefits on the outcome from past cohort data. We extend the risk-benefit analysis to generate personalized service plans. In particular, we apply a utility maximization mechanism to cast service selection and composition as a statistical decision problem, which takes a user’s clinical requirements as the input and selects the best performing set of service options as the output. User-specific information such as user profiles and service semantics are incorporated as parts of the features used to drive statistical decisions. Our task is thus to solve the maximization problem,

, (1)

where the utility function, ~

Util , is defined to estimate the empirical benefits of different service compositions from cohort data given the set of user-specific outcome evaluation criteria. Learning algorithms and Bayes decision rules are then applied to choose the composition

*

sc that yields the highest expected utility.

2) User-adaptive In-context Recommendation (Personal): The second part of the ncorporate the contextual factors into the utility maximizing personalization framework to generate personalized plans. The additional information of user context is translated into context-aware queries ,as the input to the framework to search for suitable plans and drive personalized recommendations accordingly. The expected utility is then estimated as in Formula (2) to take into consideration the effect of user context, which is denoted by cU

.

(2)

Each service composition is associated with the modified

gain function, ))(,,( ci UFscG�

θ , which depends on not only

the relevant user contextual factors )( cUF�

but also all the

parameters of ),}({ 1 Snjj θθθ == , where jθ is the

risk criteria model that generates jr . Furthermore, the posterior distribution can be calculated as in Formula (3) to allow efficient estimation.

(3)

The estimated expected utility, ),,|(~

SCrUscUtil c

, can then be used to rate available service compositions for follow-up recommendation. To overcome the data sparseness problem, we extend the concept of “like-minded” to incorporate user contextual factors, e.g., lifestyle, time, location, nutrition intake and activity history.

Figure 4. User-adaptive In-context Recomendation

V. PERSONALIZATION ENGINE

The goal of the design of personalization engine is to bring the synergy of the different components into personalized care planning – i.e., refining the recommended clinical guidelines into a follow-up personal care plan. Fig. 5 illustrates the information flow, specifically how the interactive analytics components are put together to sustain the proposed personalization service framework. First, the set of recommended clinical guidelines are input as a query to the plan generation engine. Given the input query, the engine applies the associated risk models to identify the risk levels of the target patient and select suitable care plans available in the service portfolio. When the needed risk models do not exist, the risk stratification component is invoked to learn from population data sources. The component also discovers quantifiable differences in patient-specific properties and identifies the target patients’ “personal wellness signature” (i.e., a set of patient-specific properties of which the summation of effects yields a significant effect on intervention effectiveness).

The next step is then to import the selected care plans into the personalization engine for user adaptation purposes, using information collected in the user profile and adherence history record. This leads to the need of invoking support from the user-adaptive service selection and composition component, which further ranks the selected plans with relevant contextual and personal wellness signatures in consideration.

In those cases where information is not available for some of the relevant factors, there emerges the need of invoking the interactive user profiling component (as shown in Fig. 3), which solicits inputs from the target patient for reliable adaptation. The active scheme automatically determines what other inputs are essential to be solicited from users and updates the assessment with respect to the contextual changes, based on an incoming activity stream.

∏=

∝n

jcjjiSjc UrpscpSrUp

1

),|()|(),,|(��

θθθ

~

maxarg Util

θθθ dSCrUpUFscGSCrUscutil ccici ),,|())(,,(),,|(~ ���

�Θ

=

249255

Page 7: [IEEE 2012 Annual SRII Global Conference (SRII) - San Jose, CA, USA (2012.07.24-2012.07.27)] 2012 Annual SRII Global Conference - Towards Personalized Care Coordination Service: Integrating

Figure 5. Personalization Engine

I. CONCLUSION AND FUTURE WORK

In this paper, we propose a personalization service framework and discuss its implementation from three perspectives: service, user and system design. This framework extends evidence-based medicine and risk-benefit analysis to generate personalized service plans by employing clinical practice guidelines and data-drive analytics to answer questions such as what are the properties of tailoring matrix to personalize care coordination plans, and how to use the tailoring matrix to improve the relevance of care plans and provide individual feedbacks.

This is especially important in the context of care coordination services. Its benefits are at least three-fold. First, personalized care coordination services enable providers and payers to become more accountable by bringing outcomes into the center when devising incentive schemes for care management in the extended care scenarios.

Secondly, personalized care coordination services are more patient-centric by predicting personal wellness outcome and driving service recommendation with the extra knowledge of the target patient in consideration, including their personal preferences/constraints and adherence history.

Finally, personalized care coordination services make existing practice better coordinated by bringing in system support for affordable case management, as the implementation of personalization engine is expected to down cost the expensive personalization procedures, which are currently done with manual processes.

As shown in the personalization service framework, the task of transforming clinical requirements into personalized care plans only cover the part of customization points in the consultation stage. There are certainly many challenges remaining in the follow-up stage for monitoring and personalized care plan adaptation. Also, as many individual difference-indicative signals can only be captured by integrating the data sources obtained from a multitude of service providers, the framework also needs to come with a better data sharing, integration, ownership management, and privacy-preserving scheme [27, 28]. Finally, as witnessed in

other domains that have been powered by data-driven intelligent analysis [29], the curated analytic models and best practices for care coordination plan generation are expected to foster the healthcare community’s orientation towards an “evidence-based” paradigm. The resulting repository amasses not only risk models for care plan generation, but also interactive analytics tooling and best practices for individualized guideline adaptation. The opportunity to introduce interactive analytics into the computerization of workflow and care pathways will remedy the long-lasting problem of computer-based CPGs and decision support systems [30] and bring about transformational impacts.

REFERENCES

[1] R. Grol, “Successes and failures in the implementation of evidence-

based guidelines for clinical practice,” Medical Care, vol. 39, pp. 46-54, 2001.

[2] M. Schuster, E. McGlynn, and R. Brook, “How good is the quality of health care in the United States?” Milbank Quartly, vol. 76, pp. 517-563. 1998.

[3] E. A. McGlynn, S. M. Asch, and E. A. Kerr, “Quality of health care delivered to adults in the United States – Reply,” New England Journal of Medicine, vol. 349, pp. 1867-1868, 2003.

[4] World Health Organization, “Adherence to Long-Term Therapies: Evidence for Action,” Report of World Health Organisation, 2003.

[5] S. F. Jencks, M. V. Williams, and E. A. Coleman, “Rehospitalizations among Patients in the Medicare Fee-for-Service Program,” New England Journal of Medicine, vol. 360, pp.1418-1428, 2009.

[6] P. Lamy, “Drug Use and the Elderly: Some Observations and Recommendations," The Maryland Pharmacist, vol. 64(3): 12, 1998.

[7] M. J. Field and K. N. Lohr (Eds), Clinical Practice Guidelines: Directions for a New Program. Washington, DC: Institute of Medicine, National Academy Press, 1990.

[8] R. Graham, M. Mancher, D. M. Wolman, S. Greenfield, and E. Steinberg, Clinical practice guidelines we can trust, Institute of Medicine, Washington, DC: National Academy Press, 2011.

[9] A. Balas, S. M. Austin, J.A. Mitchell, B.G. Ewigman, K.D. Bopp, and G.D. Brown, “The clinical value of computerized information services: a review of 98 randomized clinical trials,” Arch Fam Med, vol. 5, pp. 271-278, 1996.

[10] D. Wang, M. Peleg, S.W. Tu, A.A. Boxwala, R.A. Greenes, V.L. Patel, and E.H. Shortliffe, “Representation primitives, process models and patient data in computer-interpretable clinical practice guidelines: a literature review of guideline representation models,” Int J Med Inform, 2002.

[11] R. Grol and J. Grimshaw, “From best evidence to best practice: effective implementation of change in patients' care,” Lancet 2003; 362: pp. 1225-1230.

[12] UK Prospective Diabetes Study (UKPDS) Group, “Intensive blood-glucose control with sulphonylureas or insulin compared with conventional treatment and risk of complications in patients with type 2 diabetes (UKPDS 33),” Lancet, vol. 352 (9193): pp. 837–853, 1998.

[13] US Department of Health and Human Services, “DPP (Diabetes Prevention Program),” NIH Publication, no. 09-5099, 2008.

[14] J. Tuomeilehto et al, “Prevention of Type II diabetes mellitus by changes in lifestyle among subjects with impaired glucose tolerance,” New England Journal of Medicine, vol. 344(18), pp.1343-1350, 2001.

[15] D. Eddy et al, “Individualized Guidelines: the Potential for Increasing Quality and Reducing Costs,” Annals of Internal Medicine, vol. 154: pp. 627-634, 2011.

250256

Page 8: [IEEE 2012 Annual SRII Global Conference (SRII) - San Jose, CA, USA (2012.07.24-2012.07.27)] 2012 Annual SRII Global Conference - Towards Personalized Care Coordination Service: Integrating

[16] K. L. Sloan et al, “Construction and characteristics of the RxRisk-V: a VA-adapted pharmacy-based case-mix instrument,” Medical Care vo. 41 (6), pp.761-774, 2003.

[17] R. B. d'Agostino, R.S. Vasan, M. J. Pencina, P. A. Wolf, M. Cobain, J. M. Massaro, and W. B. Kannel, “General cardiovascular risk profile for use in primary care: the Framingham Heart Study,” Circulation, vol. 117 (6), pp. 743-753, 2008.

[18] C. Macknight and K. Rockwood, “Use of the chronic disease score to measure comorbidity in the Canadian Study of Health and Aging,” International Psychogeriatrics, vol. 13 Supp 1, pp. 137-142, 2001.

[19] D. H. Solomon, S. Schneeweiss, R. J. Glynn, Y. Kiyota, R. Levin, H. Mogun, and J. Avorn, “Relationship Between Selective Cyclooxygenase-2 Inhibitors and Acute Myocardial Infarction in Older Adults,” Circulation, vol. 109 (17), pp. 2068-2073, 2004.

[20] A. Elixhauser, C. Steiner, R. Harris and R. Coffey, “Comorbidity measures for use with administrative data,” Medical Care, vol. 36 (1), pp.8-27, 1998.

[21] M. E. Charlson, P. Pompei, K. L. Ales, and C. R. Mackenzie, “A new method of classifying prognostic comorbidity in longitudinal studies: development and validation,” Journal of Chronic Diseases, vol. 40 (5), pp. 373-383, 1987.

[22] C. Schaefer, “The Kaiser Permanente Research Program on Genes, Environment and Health: A Resource for Genetic Epidemiology in Adult Health and Aging,” In the Proceedings of 17th Annual HMO Research Netowrk Conference, 2011.

[23] G. A. Thorisson, A. V. Smith, L. Krishnan, and L. D. Stein, “The International HapMap Project Web Site,” Genome Research vol. 15 (11), pp. 1592–1593, 2005.

[24] P. Hsueh, R. Lin, M. Hsiao, L Zeng, S Ramakrishnan, and H Chang, ”Cloud-based platform for personalization in a wellness management ecosystem: Why, what, and how,” In Proceedings of IEEE Collaborative Computing: Networking, Applications and Worksharing (CollaborateCom), 2010.

[25] H. M. Roche, “Nutrigenomics—new approaches for human nutrition research,” Journal of the Science of Food and Agriculture, vol. 86 (8), pp. 1156–1163, 2006

[26] M. Mori et al, “Genetic basis of inter-individual variability in the effects of exercise on the alleviation of lifestyle-related diseases,” Journal of Physiology ,vol. 587 (23), pp. 5577–5584, 2009.

[27] P. Hsueh, T. Grandison, X. Zhu, H. Pai, and H. Chang, “Challenges and Requirements on Privacy in Enabling Evidence Use Service on Wellness Cloud,” Frontiers in Service Conference, 2012.

[28] T. Grandison, P. Hsueh, L. Zeng, H. Chang, “Privacy Protection Issues for Healthcare Wellness Clouds,” Book Chapter In Privacy Protection Measures and Technologies in Business Organizations, IGI-Global, 2011 .

[29] J.R. Cooper, "Curing Analytic Pathologies: Pathways to Improved Intelligence Analysis," Center for Study of Intelligence,Central Intelligence Agency,Washington,DC, no.20505, 2005.

[30] P. Gooch and A. Roudsari, "Computerization of workflows, guidelines, and care pathways: a review of implementation challenges for process-oriented health information systems," JAMIA, 2011.

(a) (b)

Figure 6. Data-driven Risk Stratification

251257