On April 23, 2016, over 300 developers from around the country descended on San Francisco for the weekend to tackle some of the hardest challenges facing the nation. The event was called BayesHack, sponsored by the nonprofit Bayes Impact. There were representatives from 7 cabinet-level federal agencies present to set up the 11 “prompts”, mentor the teams and judge the entries. The prompts for the U.S. Department of Health and Human Services and the Department of Veterans Affairs asked challenging questions on how to leverage existing datasets…
- How can data connect individuals with the health providers they need?
- How can data get help to sufferers of opioid addiction?
- How can data predict and prevent veteran suicide?
- How can data tackle End Stage Renal Disease (ESRD) and Chronic Kidney Disease (CKD)?
As part of the judging process, the teams had to pitch their solutions to both agencies and private sector judges, such as partners at Andreessen Horowitz. All teams submitted their code to the event’s github account, so that it could be used for judging, as well as ensuring that it will be available in the public domain. For hackathons such as this one, it’s important to recognize that even if there are already similar commercial products, getting solutions into the public domain makes it possible for others build on later. (Incidentally, this focus on actual working prototypes via GitHub is surprisingly lacking from many hackathons. Bayes did a great job focusing on potential implementation beyond just the weekend.)
Of particular focus was the “How can data connect individuals with the health providers they need?” prompt, since this data has only recently become available, due to a regulatory requirement. This data consisted of commercial healthcare provider networks for plans on ACA insurance marketplaces, including plan coverage, practice locations, specialties, copays and drug formularies. There were 7 team submissions, most of which produced solutions focusing on usability for consumers and advanced analytics to policy makers. Some teams expanded the scope to include not just insurance selection, but access to care in general.
To summarize some of the novel ideas in the solutions…
- Simplified mobile-first user experience, resembling TurboTax for health selection
- Visualizations and what-if analysis for policy makers
- Voice recognition and NLP, as in Google freeform search instead of menus and buttons
- Ranking algorithms and recommendation engines
- Ingesting additional 3rd party information (such as Vitals, Yelp, and Part D claims) for consumers who need additional information before they can make an informed choice
- Providing an API for other apps to leverage
- Enabling self-reporting of network accuracy, like GasBuddy for health plan coverage
Here are some notable entries for this prompt:
|The Hhs-marketplace team created an app that leverages chart visualizations to let a consumer compare plan attributes against benchmarks, such as state averages. The example below shows a user entering a zip code and the specialists they’re interested in seeing. The app finds the plans that meet those criteria, displays cost comparisons for them and a graphical comparison of the options.
|The Fhir salamander team created mobile-first responsive web front end that takes the user through a series of simple menu choices to get them to recommended plans. Along the way, for convenience and efficiency, it enables the user to click a button to place a telephone call to the plans (to ensure that the doctor they want is taking new patients from that plan) or to view the summary plan description files.
|The Hhs insights team produced an interactive provider density map. Their approach was to target policy makers, rather than consumers. For that purpose, they built aggregate analytics and related visualizations. For example, their code uses DOL MSA (Metro Statistical Area) for GeoJSON calculations and visualizations. In order to enable the needed analytics, they had to take on the challenge of normalizing the JSON schema of provider networks into a tabular format, as well as pre-calculating several aggregate metrics.
|The Hhs marketplace finder team created an app that displays the pros-cons of the top 5 plan option for the
user, along with visualizations for making quantitative comparisons easy to understand. Bad choices are suppressed to avoid screen clutter. It starts with less than 10 simple questions. Then adds a prediction of the user’s healthcare needs, which was determined based on statistics by age, gender, preexisting conditions and location. Finally, it would eventually make it possible for a user to estimate their total cost based on different events, such as hospitalization or illness.
|A data science team from Berkeley calling themselves Semantic Search, submitted an extremely ambitious
project. Basically, creation of a Google Pagerank for healthcare decisions. Instead of the menus and buttons of a traditional app UI, this solution used a freeform field for a user to indicate what they were looking for. The goal is to let a consumer who is not tech saavy explain their situation in a natural way, without the interface and technology getting in the way. Under the covers it uses natural language processing, ranking algorithms and a recommendation engine. The user is ultimately presented with the top couple plans, along with explanations of why they were recommended. To make the solution possible, this app has to collect behavioral data logs, use logistic regression to predict the probability that a certain plan would work, and leverage the LETOR ranking mechanism to provide answers.
|Disclaimer: The challenge prompt used for HHS does not constitute the department’s official stance or endorsement of this activity. It was used in an unofficial capacity only and intended to take advantage of data newly available from industry due to changes in regulations of the health insurance marketplace.|