MIT Sloan Health Systems Initiative
HSI Fall 2020 Research Updates
Fresh Food Farmacy: A Randomized Controlled Trial – Joseph Doyle
Addressing social determinants of health, including food insecurity, is a “big idea” in healthcare delivery—a re-focus toward prevention rather than being reactive. This project tests a nationally recognized approach that provides healthy food and education to low-income patients with diabetes to improve health and lower healthcare costs in a randomized, clinical trial. The trial now has 400 people enrolled and is on track to reach its target of 500 participants this spring. Despite the challenges posed by the pandemic, recruitment and delivery of the intervention have proceeded remarkably well, albeit with some adjustments. During the pandemic, the service moved to food pick up every other week and curbside pickup. Dietitian and RN advice moved to telephonic and video calls. In terms of impact, the team will compare treatment and control groups at 6 months and 1 year. Among the treatment group there are some success stories, including dramatic improvements in blood sugar levels
Predictive-Prescriptive Analytics to Address the Substance Abuse Crisis – Georgia Perakis and Dessislava Pachamanova
Professor Georgia Perakis and her team are combining probability, statistical learning and machine learning to create an interpretable model that predicts the best treatment for any patient. While this project started as a part of the Substance Use Disorder research with Lahey Health, its scope has greatly expanded. The new approach takes a holistic, dynamic view of a patient’s condition and treatment path: it personalizes the estimation of a patient’s risk in the context of learning treatment strategies that improve outcomes. The proposed approach, although motivated by treatment of substance abuse patients, turns out to be far more general. The COVID-19 pandemic unfortunately disrupted the team’s collaboration with Lahey Health, However, they continued developing the theoretical framework and the associated algorithm. Since the last update, they completed the theoretical aspects of the project, and tested the algorithm on well-known examples from the reinforcement learning (RL) domain. These experiments have confirmed their results and show that the algorithm learns a representation of the system. The next step is the development of novel approaches to apply this methodology for the application to real healthcare data.
Combining Machine Learning with Behavioral Insights to Provide Differentiated Digital Adherence Support – Jónas Oddur Jónasson, David G. Rand, and Erez Yoeli
This project focuses on patients enrolled in a program using technology provided by the digital health startup Keheala. These patients have all been prescribed a medication that must be taken daily over the course of several months. Keheala, installed on a basic mobile phone, enables patients to indicate that they have taken their daily medication. Some patients, however, need several reminder texts, and there are some whose cases are elevated to a healthcare team for more intensive follow-up. In this study, the researchers are combining machine learning and behavioral insights to identify at-risk patients who would benefit most from such personal engagement, and to estimate treatment effects using only data that are available pre-enrollment. With these insights, healthcare providers will be able to prioritize access to the treatment in resource constrained settings. They tested their model on data from a year-long RCT they ran in 2016 in Nairobi, Kenya. Their model was not terribly successful identifying appropriate patients based solely on pre-enrollment data; however, the model was tremendously successful at identifying such patients once they had been on the Keheala platform - even for only a few days. While the model is best able to predict who will need personal intervention the next day, it is also successful at making predictions about ultimate treatment success months in advance.
Their next step is to test the model on a more complex set of data from a three-year Kenya-wide RCT that they just completed. The goal is to enable Keheala and its field partners to provide differentiated care, allocate scarce resources and, more importantly, to provide a proof-of-concept for others.
Data Challenges, IT and Healthcare Challenges in Reporting Accurate Data in the COVID-19 Pandemic – Catherine Tucker
This research focuses on the issue of COVID data quality, including missing data, reporting delays, inaccurate data and inconsistent or unclear data. Tucker and her team gathered the data reported by state public health officials on their COVID-19 websites. They also include evidence from news reports and from their analysis of third-party data sources that track official updates from each state.Along with documenting the range and extent of inaccuracy, they also document how certain state regulations, such as privacy rules, or the use of outdated technologies, such as faxes, for reporting is contributing to the problems observed in data reporting.
Expected outcomes of this project are a whitepaper detailing the observed data issues and an empirical analysis, which will highlight how the use of fax machines to report data leads to such large data lags so as to render studies of non-pharmaceutical interventions unreliable.
Analytics-based Opioid Overdose Prediction – Jónas Oddur Jónasson and Nikos Trichakis
This project focuses on two questions: (1) How well can machine learning models predict an individual’s risk for adverse opioid-related outcomes? (2) How effectively can providers use these models to intervene with patients? The project’s data are from adjudicated data from health insurance claims and electronic health records (EHRs) of Medicaid beneficiaries in one Staten Island county.Since the previous update, the team has added data from diverse sources and honed their predictive models, so they are more powerful and better at identifying at-risk patients. Not only does the model identify patients at risk for opioid overdose, but it also captures other types of harm, such as poisoning. They have also conducted experiments to evaluate the usefulness of these models for intervention with patients.
They have found that their models remain robust even if there is a six-month delay in populating the database, which reflects real-world limitations in data systems. Other insights include that intervening in a small percentage of patients would still capture a substantial number of patients with any opioid-related outcome. Similarly, effective interventions at regular intervals can also prevent many adverse outcomes.
This project investigates how work redesign may improve the health and well-being of workers. Specifically, whether workers in fulfillment centers that establish a Health and Well-being Committee (HaWC) composed of frontline workers and middle managers who work together to address workplace concerns fare better than their counterparts in other fulfillment centers in the same company. The HaWC establishes a new channel for employees to share concerns, prioritize those concerns, brainstorm solutions, and carry out improvement projects that are oriented to workers’ well-being while recognizing operational needs.The central hypothesis is that this intervention will improve the mental health of frontline workers and may reduce injuries. Key organizational outcomes include improvements in absenteeism, turnover, and productivity.
This summer and fall, project activities focused on intervention development, research design, and securing buy-in from multiple levels of firm management. They also successfully carried out a baseline survey at the pilot site in early November. Unfortunately, the COVID-19 pandemic has just caused another delay because the pilot site managers and employees cannot take the time to work on this project given dramatic increases in order volume and increased absences. The pilot site managers and research team agreed to return to the next steps –a participatory co-design workshop to fine-tune the HaWC plans -- in January 2021, and then launch the full committee’s activities in February 2021.