Challenges in using machine learning to direct healthcare services

Description

Zocdoc is an online marketplace that allows easy doctor discovery and instant online booking. Many of Zocdoc’s core product and business functions are fairly common in internet services, and each function can be cast as a well-defined and -researched machine learning opportunity. However, dealing with healthcare involves many constraints and challenges that render standard approaches to common problems infeasible. Brian Dalessandro surveys the various machine learning problems Zocdoc has faced and shares the data, legal, and ethical constraints that shape its solution space, explaining how regulatory-induced business model constraints translate into new formulations of common machine learning problems.

Healthcare service companies are required to comply with a number of regulations (HIPAA being the most notable). Operating in a HIPAA-regulated data environment imposes restrictions that contradict the “open data” cultural values of many data-driven organizations and limits the way certain data may be used. Nonetheless, HIPAA concerns can often be solved with existing technology solutions and policies. Other policies put constraints on how a company might monetize their service, creating constraints on the company’s business model. For instance, Zocdoc uses a subscription-based rather than a transaction-based model. A basic transformation in business model has many downstream effects on how machine learning problems are formulated and solved.

Beyond regulatory constraints, there are ethical considerations that are particularly important in making decisions related to healthcare. Brian walks you through the common machine learning development pipeline (from problem formulation to research, then to deployment and evaluation) and shows how the aforementioned constraints can be worked into the process, with a focus on how to adapt to new constraints and develop the appropriate patterns working within such a constrained data mining framework.