The following blog was co-written by Prashant Natarajan, Product Director, Oracle Healthcare Solutions.
In our book, "Demystifying Big Data and Machine Learning for Healthcare," we expose the ways that new advanced analytics capabilities, including big data, machine learning and artificial intelligence (AI) can be used by healthcare organizations to help humans who work in the healthcare system do what they do, only better. In the course of the book, we explain the key technologies that allow us to understand clinical data, including unstructured text and images, to build intelligent tools and to deploy them at scale to address real challenges in healthcare. Our hypothesis is that the biggest opportunity for advanced analytics is to use new technologies to augment human capabilities, not to replace humans.
The biggest concern we hear from our readers, and from the healthcare community at large, is, "we are already struggling to get clinicians to use the decision support tools we already have. They have alert fatigue. We don't need more technology." Our response is, "we agree."
Alert fatigue happens when technology doesn't understand the very complex context of each clinical encounter. Furthermore, the opportunity to move the needle for clinical outcomes and for health system operational efficiency is greatest for situations in which there is not already a highly trained expert managing the process. For example, the impact of technical information delivered to a clinician in the midst of a typical encounter is not high. On the other hand, interactive communication of information that will help a patient navigate a complex care trajectory, or allow a case manager to quickly identify the critical post-acute care that a patient needs following hospital discharge, can have a huge impact of outcome and cost. Hence the need for Contextually Intelligent Agents, or CIAs.
Contextually Intelligent Agents
CIAs leverage the new and powerful ability of machine learning and AI systems to understand human spoken and written communication, and are increasingly sensitive to subtle contextual information that can make technical information both timely and effective.
OK, so you're convinced that well-implemented CIAs could be valuable to guide patients and care providers through the complexities of modern healthcare. Your next concern is data quality. Algorithms are only as good as the data that goes into them. To use CIAs, don't we need to guarantee that all the data going into our algorithms is pristinely pure and error free? And if we are using unstructured data to feed our algorithms, how can we even be sure that our inferences from that data are 100 percent accurate?
Our response is that you need to think about data fidelity, not absolute data accuracy. This is the insight behind the NRF Framework, which we explain in the book. The idea is that each application uses a variety of data sources, each of which has some non-zero uncertainty. Further, each application has a different sensitivity to the uncertainty in its incoming data. To communicate the next step in a patient's post-acute physical therapy, and to encourage that patient to promptly follow through on recommended care, does not require 100 percent data accuracy. It requires an understanding of the situation and the ability to interpret responses from the patient. Furthermore, modern CIAs have the ability to clarify incoming information when the data fidelity is not up to snuff.
Data Fidelity Concept
The data fidelity concept works for input to critical treatment decisions, as well as it applies to patient coaching: In cases where errors can be dangerous, high data confidence is required. Note that most clinical information systems currently assume that system data about patients is 100 percent reliable. Ironically, studies show very high rates of both false positives and false negatives in clinical structured data, so implementation of the NRF Framework actually addresses both new analytical capabilities and current clinical risks to patients due to avoidable errors.
The key element of the NRF Framework is that each data stream has a measurable quality. When the data fidelity is not high enough for the application, then measuring data quality creates the opportunity to improve quality through both data quality processes, human driven quality review and algorithmic inference from other sources of data (for example, when a coded diagnosis can be shown to be a false positive or incorrect coding by performing machine learning inference on clinical text and other measurement results.
We see a bright future for humans in healthcare, in which patients have better outcomes and providers, caregivers and other actors in the health system can take advantage of CIAs to interactively provide them with the contextually appropriate, actionable information they need, when and where they need it, and all with appropriate data fidelity, as prescribed by the NRF Framework. Humans and machines, rejoice!