At Discern Health, we are harnessing the power of data and AI to improve health longevity and fulfillment. However, we know the responsible development and deployment of AI models, especially in healthcare, is paramount.
That’s why we follow best practices including clinical oversight, the ascertaining of data privacy and security, ensuring accountability and transparency, and the thoughtful identification of blind spots that may lead to bias. Moreover, we are aligning our work to evolving national regulations around AI as a medical tool.
Jacob Dodd, lead data scientist at Discern Health, guides our team to build models with all the aforementioned central tenets regarding responsible AI at the top of mind. Dodd found his passion for data and its power while in college when he worked on a project recording and analyzing whale sounds to determine patterns of communication. His acumen and understanding of data have evolved to its use in studying geographic spaces, water nutrients, finances and now healthcare outcomes and utilization patterns.
The Complexities and Opportunities with Health Data
Health datasets are unique in that they can be invaluable to drive positive change for health outcomes. Data comes in a myriad of formats and from a host of sources. Each of these representations completes the data picture of an individual or a population. An understanding of the sources and formats is critical to their use and to compare their semantics and syntax. Health data is also challenging because of issues surrounding data ownership rights and the need for privacy.
“Health data is interesting, but certainly can be a big challenge. The different types of data you can access are in different formats. We have our own dataset comprising claims data for millions of people and clinical data, both structured and unstructured, for an equally impressive number. Despite having three broad categories of data assets, the nuances continue. For example, not all claims are created equal. We have seen many different formats that claims are presented to us . . . we need to be thoughtful about how we prepare for the variability that is commonplace in health data.” Dodd said. “Claims variability is only the tip of the iceberg – data sets like labs, free text notes, etc. has even more variability in its makeup and despite standards for their exchange, that variability has not been eliminated.”
Keeping Humans in Healthcare
While AI solutions like the models Dodd’s team builds at Discern, show great promise to advance a more outcomes based and equitable healthcare future, particularly with meaningful preventative care to enhance lifespans, humans in the loop are of great import – both the humans that build and clinically validate and the humans whose clinical profiles they help address.
“We need to have humans in the loop, especially in healthcare,” Dodd continued. To date, Discern has about 30 models that predict a range of costly conditions and their impact on the most vulnerable, be it in terms of morbidity, mortality or financial costs. Discern sees the value in meaningful clinical-data science collaboration. This is critical to determine model relevance before the build and model accuracy after it. “We work with the medical team to determine the specific problem we’re trying to solve and make sure that the way we formulate the problem is defined with our data and actually solves that problem.”
In addition, Discern’s predictive models are not black boxes making automated decisions. “Accountability and transparency of the models is vital,” Dodd said. “Unchecked automated decision-making is risky and can perpetuate bias and marginalize those already at risk. At Discern, we specifically include rationale for all the predictions we are making. We identify the critical features that drive model performance and highlight them in a way that makes outputs clear and actionable for healthcare providers.”
Data Privacy and Security
Protecting sensitive health data is a top priority. Discern Health keeps client data separate and secure through approaches like federated learning and cloud-based containerization.
Speaking on model training, Dodd shared, “Our models are trained on a reference population that might be different from the client’s population. Despite similarities in the data, performance from data set to data set can vary. We always calibrate our models to our client’s dataset.”
This speaks to the deployment of Discern’s federated learning philosophy. Here is how it works at Discern Health:
- Build model on Discern’s reference population of 77 million patients.
- Send Discern’s models to the client’s environment.
- Train the models to calibrate with the client’s dataset.
- Bring only the models back to Discern and their learnings, not the client’s data.
“This approach helps our models improve and we can deploy a better model for not only that client, but other current and future clients as well. Despite the learning, a client’s data never leaves that secure area,” Dodd said.
With emerging data sets like wearables, epigenetics, genetics, pharmaceutical, etc. insights from a given client’s differentiated data set can be integrated into the base model to improve the model’s performance for not only this client but future clients as well.
Better Health Through Responsible AI
Ultimately, our goal is to positively impact health outcomes through the responsible use of data science and AI. For Dodd, success means being able to change health outcomes, to help patients stay as healthy as possible for as long as possible. As Dodd shares, “It would be incredibly meaningful to me personally if we could prevent something like a fall or alter a clinical outcome for a patient because of a model I helped build.”
At Discern Health, we are committed to pushing boundaries with AI while maintaining strict ethics, security, and human-centered processes. The path to better health runs through innovation and responsibility.
