Mining Dark Data for New Use Cases & Scalability

Posted June 26, 2018 at 10:00 AM by Yukun Chen

In January, a brilliant OpEd ran in the New York Times in response to an MIT Technology Review piece that warned about the lack of transparency in the decision-making process for AI systems. In addition to likening AI systems to “black boxes”, they culminated with a warning that “no one really knows how the most advanced algorithms do what they do”.  Talk to any data scientist who works with NLP and AI systems and they’ll fundamentally disagree with this stance (and likely side with our champion at the New York Times).

For AI systems to truly achieve optimal (and applicable) results in the hospital setting, we’ve always believed that it’s prudent to have a human-assisted machine learning throughout the whole life cycle of model development. At Pieces, we take that one step further by demanding that we have clinician-assisted machine learning, with our Clinician-in-the-Loop™ process.  

Beginning with initial model training and evaluation, to post-deployment auditing and continuous improvement, to transfer learning about implementation of a model from one hospital to another, Pieces-employed clinicians oversee the machine performance. 

Our clinicians and nurses can audit the live operation of all of our models (think: readmission risk scores, identifying discharge barriers, predicting the onset of disease or infection) in real-time and perform chart review and annotations, when the machine asks for them.

Similar to autonomous vehicle approaches, where the car will switch from autopilot to manual-driving in uncertain situations, Pieces can alert our clinicians with cases or situations that are too ambiguous for the machine to resolve during simulation exercises.

In these cases, Pieces clinicians will intervene by either accepting the machine’s predictions or correcting false predictions with proper signals delivered to our clients along with other high-confidence predictions.

Kermit

Annotated data from the ambiguous situations then trigger a model improvement processes in the DS Learning Lab. Like above, which allows Pieces to train and scale efficiently across our customers.

filed under: NLP, Clinical AI

Subscribe to Email Updates

Recent Posts