Healthcare Insights AI Lab
Deep Learning based AI is a particularly good fit for health system data, as the volume and diversity of this data has surged in recent years. Many of these health system data sources are unstructured or semi-structured. While traditional methods for analysis are struggling to provide meaningful insights, Deep Learning based AI offers tremendous potential. The Healthcare Insights AI Lab was launched in 2017 under the direction of Professor Gordon Gao and is staffed by CHIDS-affiliated faculty scientists and graduate students, who work in close collaboration with Inovalon (NYSE: INOV) and the New Jersey Institute of Technology under the direction of Professor Yi Chen to envision, create and grow impactful AI-based projects. Projects have focused on use of Deep Learning for reading and analyzing unstructured clinical notes, optimizing interventions and forecasting resource needs.
Gordon Gao, Yi Chen, Weiguang Wang, Jinhe Shi, Kenyon Crowley, Margret Bjarnadottir, Ritu Agarwal
Chronic opioid therapy (COT) has been associated with serious adverse outcomes and the social and economic impact of continuing opioid treatment is sizeable. The net effect of COT on a given patient's health – beneficial, adverse, or neutral – may be difficult to determine ex ante, and affected by many unobservable factors. Given the risks and adverse outcomes associated with COT, many consider COT a care choice of last resort. Thus, to the extent that clinicians need to make decisions regarding whether opioid use will begin or continue, decision support on the individual-patient probability of COT appears critical. Personalized guidelines, built on decision support systems (DSSs), have the potential to influence care at the point of service. Building models that can serve as the foundation of such systems can therefore contribute to changes in physician prescribing. As a result, we aim to study personalized pain management-related decision support for opioid prescribing. Specifically we will apply state of the art machine learning algorithms, as well as more traditional models, to study the feasibility and potential impact of a COT DSS, using a large data set from the U.S. Army. We will further investigate the economic impact of such decision support. The National Institute for Health Care Management (NIHCM) is helping support this work through its Research Grants program.
Margret Bjarnadottir, Ritu Agarwal, Kenyon Crowley, Al Nelson, Kislaya Prasad
Digital Therapeutic Assessment and Development
Mobile personal devices, health engagement technologies and analytics have the potential to play a transformative role in improving healthcare delivery and health outcomes by facilitating greater patient engagement and involvement in their own health, with personalization having effects. CHIDS is working with mutliple industry partners that have developed digital solutions targeted at chronic diseases like diabetes and congestive heart failure. CHIDS is working with these partners to evaluate its product in field trials and help foster machine learning models that can classify users along various attributes and help predict effective engagement strategies.
Gordon Gao, Kenyon Crowley, Michelle Dugas, Ritu Agarwal, Margret Bjarnadottir, Weiguang Wang
The Dark Side of Social Influence
Social influence in the form of social norms has been widely used to transform behaviors, and is argued to be especially efficacious in the context of health related activities. However, can such externally induced compliance produce negative outcomes? When individuals feel compelled to conform to the behavior of the majority, does it lead to an unexpected backfire effect? We conducted a randomized field experiment of more than 10,000 individuals for a two-month period on an online physical activity community to examine if there is a dark side to social influence. We studied the effect of social norms on users’ goal setting and goal achievement behavior. While social influence increases the rate of goal setting, strikingly, we also observe a dark side to social influence in that such influence yields lower rates of goal achievement. Our findings have important implications for the design of interventions in the context of mHealth technologies. Read paper.
Che-Wei Liu, Gordon Gao, Ritu Agarwal
Vocal Minority and Silent Majority: How Do Online Ratings Reflect Population Perceptions of Quality?
Consumer-generated ratings typically share an objective of illuminating the quality of a product or service for other buyers. While ratings have become ubiquitous and influential on the Internet, surprisingly little empirical research has investigated how these online assessments reflect the opinion of the population at large, especially in the domain of professional services where quality is often opaque to consumers. Building on the word-of-mouth literature, we examine the relationship between online ratings and population perceptions of physician quality. Our study builds on prior work by leveraging a unique dataset which includes direct measures of both the offline population’s perception of physician quality and consumer-generated online reviews. As a result, we are able to examine how online ratings reflect patients’ opinions about physician quality. In sharp contrast to the widely voiced concerns by medical practitioners, we find that physicians who are rated lower in quality by the patient population are less likely to be rated online. Although ratings provided online are positively correlated with patient population opinions, the online ratings tend to be exaggerated at the upper end of the quality spectrum. This study is the first to provide empirical evidence of the relationship between online ratings and the underlying consumer-perceived quality, and extends prior research on online word-of-mouth to the domain of professional services.
Gordon Gao, Brad Greenwood, Ritu Agarwal
Understanding Sticker Prices: An Analysis of CMS Provider Charge Data
This year, the Centers for Medicare and Medicaid Services (CMS) released data containing information on provider charges (i.e., sticker prices) and CMS payments for the 100 most common inpatient services and 30 common outpatient services. Our objective is to evaluate how CMS payment models explain payments and sticker prices for both inpatient and outpatient services. We have found that payments for inpatient and outpatient services are explained extremely well by their respective models, but these payment models do not explain sticker prices well. We are currently applying advanced machine learning methods to an augmented data set to improve our understanding of sticker prices. Our analysis demonstrates that providers are using additional (and potentially diverse) information to determine their prices. Ongoing research investigates whether higher sticker prices are associated with higher quality providers, higher profits, and other potential provider-based outcomes.
Margret Bjarnadottir, Sean Barnes
Data Driven Detection of Episodes-of-Care
Healthcare reimbursement is at the forefront of healthcare reform debates in the United States. Bundled payment systems have been proposed as a practical and promising reimbursement alternative to induce incentives that lower healthcare expenditures. Bundled payments reimburse a single amount for an episode of care in contrast to separate payments for every single medical service rendered in the current fee-for-service model. Significant evidence suggests that bundled payment systems achieve improved quality outcomes, and better coordinated care at a lower cost. The present process of defining an episode of care relies on the consensus of expert panels on which services constitute an episode. This manual approach is inconsistent and not systematic. The approach is labor-intensive and inflexible and is inefficient for dynamically incorporating heterogeneity in patients’ health conditions, comorbidities, and medical practices across communities. Additionally, these manual approaches have focused on individually definitions of episode of care, which does not allow communities to analyze the collective impact of moving to a bundled payment system on their overall expenses and healthcare outcomes. Consequently, it is currently impossible to analyze and design the necessary conditions that will facilitate the adoption of bundled payments in a community, and across the country. Although the expert panel approach is clinically meaningful, the process of defining an episode of care could be enhanced and accelerated by employing sophisticated data mining and optimization models. This study is developing data mining algorithms that deal with large unstructured and heterogeneous data to characterize episodes of care for different groups of patients. This study uses a large repository of claims data, over 110 million claim records of 660,000 insured customers, to derive service distributions for different episodes of care, using advanced data mining methods.
Margret Bjarnadottir, Sean Barnes