Segmed Ethics Committee Recap: Promoting AI Research Integrity

Anonymized medical data sharing is necessary for the advancement of medicine, but it requires a tremendous amount of trust from all stakeholders. Patients need to trust that their healthcare providers are utilizing their data to advance innovations that will help them in the future. Healthcare providers such as health systems and hospitals must trust external stakeholders to use that data for actual good. Transparency is what allows for accountability, with data sharers and data users providing checks and balances.

Radical transparency is at the heart of Segmed’s identity, which is why we formed an ethics committee. The committee will be a quarterly gathering of our data partners, discussing issues pertinent to medical data sharing, sharing their thoughts and providing guidance on the way in which their data will be used. 

Our first committee meeting was centered around patient opt-in or opt-out consent for data sharing, with Aline Lutz (Medical Director), Akemi Leung (Lead Designer), and Jie Wu (Chief Data Officer) moderating the discussions. We invited representatives from our data partners, including legal counsel, to provide further guidance. The attendees included a COO and Chief Counsel from a large imaging group in the Northeast, a VP of Innovation from a health system in the South, and a Information Systems Manager from a large imaging group in the Southwest. Some highlights of the meeting are gathered below: 

Should patients have to actively consent to share their anonymized medical data? Or should they have an option to opt-out?  

Opt-in consent requires patients to actively check “yes” to anonymized medical data sharing, whereas opt-out provides an optionality for patients to say “no” to have their data removed from data sharing. The general consensus is opt-out consent has provided better and faster care for more patients. To our data partners, data sharing is inevitable for things such as diagnostic support, so the key is to include as many patients as possible while ensuring that the anonymization is sufficient. “We are always asking how the data could be used to give back to the community,” said one VP of Innovation.

How does your organization handle patient consent for data sharing?   

Data partners use a variety of methods to solicit patient consent as part of the private practices policy. One partner allows the patient to opt-out in the online patient portal at any given time. A common finding among partners: very few people opt-out of the data sharing consent. This low opt-out rate has actually improved the healthcare quality. A partner mentioned that their data is fed into a cancer-finding program, which has caught several instances of early-stage cancer that was otherwise undetected. 

What lessons have you learned to make data sharing work?

Partners unanimously cited a strong relationship between the exporting and intaking teams as essential to data sharing success. Another important step: involving as many internal stakeholders as possible. This includes legal, finance, clinical, compliance, research IT, strategy, innovation, and more. 

How does opting out influence the diversity of data representation in AI training? How can we make sure that vulnerable and underrepresented populations are included?

A study by Kaushal, Langlotz et. al found that AI training data is extremely underrepresented across most US States.1  Data partners were troubled by the lack of diversity in AI training data. One data partner suggested that there should be an initiative to reach underrepresented populations while teaming up with local research groups. On the other hand, potential drawbacks of reaching extremely underrepresented populations include the re-identifiability of patient data based purely on ZIP code for rural areas. We heed this potential risk at Segmed and do not identify our data partners beyond state/geographical regions.  Further discussions and learning on this topic will continue between us and our partners. 

The Segmed team would like to thank all of our data partners for taking their time and exploring these issues with us. We plan on including insights gained from this discussion into further privacy and data anonymization development and considerations. 

Ref:`Kaushal, A, Langlotz, C., (2020). Geographic Distribution of US Cohorts Used to Train Deep Learning Algorithms. JAMA: The Journal of the American Medical Association, 324(12), 1212-1213.