Our investigators are a part of an AIM-AHEAD infrastructure core that will enable a coordinated data and computing infrastructure that enhances the interoperability of large-scale data resources with data that are maintained, governed, and prepared by individual institutions to preserve privacy and autonomy. Below are some of the innovation tools which we are working on building for the program:
- Trustworthy AI/Federated learning
In order to build a trustworthy AI platform, we have utilized blockchain as the underlying model sharing platform (on-chain), and secure storage and distributed training/federated learning environment as the off-chain system to train the model. All the trained AI models will be stored in an immutable ledger sharing with different organizations on chain.
- AI Bias Assessment and Reduction
We have developed a technique for measuring bias of AI algorithms for imaging diagnosis across multiple hospital institutions, as well as an algorithmic technique to correct for biases through unsupervised domain adaptation.
- Semi-Supervised Learning
We have developed an algorithmic technique for semi-supervised learning, i.e. learning with missing data labels, that has led to several peer-reviewed publications. We have analyzed the performance of this methodology with medical imaging data for early screening lung cancers, although in theory it can be applied to other data modalities including structured and unstructured EHR data fields. We are currently working toward the integration of semi-supervised learning with NLP analytics to enable robust inference across EHR.
- Blockchain Data Sharing
A secure data sharing platform is extremely important for sensitive data, such as patient personal information. Blockchain plays an important role in sharing the data with permissioned members in a blockchain network. We plan to build a data-sharing platform including images, all kinds of string data, or patient records with permissioned blockchain.
- Secured Blockchain with Reinforcement Learning
Previously, the blockchain systems were believed to be secure with its default proof of work (PoW) consensus protocol, with which cyber attacks are never beneficial as the attacker can never achieve more than they contribute. It was shown later that it is possible for the attackers to benefit from selfish mining attacks, where the attackers pretend to contribute to the blockchain system while withholding important transactions. We plan to use reinforcement-learning-based solutions to detect such selfish behaviors by autonomously interpreting participants’ intentions from their observable operations.