MIT CSAIL’s system defers can defer to experts when making predictions

Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory detailed a system that can work with human experts to arrive at a conclusion. …

A new study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) proposes a machine learning system that can examine X-rays to diagnose conditions including lung collapse and an enlarged heart. That’s not especially novel — computer vision in health care is a well-established field — but CSAIL’s system can novelly defer to experts depending on factors like the person’s ability and experience level.

Despite its promise, AI in medicine is fraught with ethical challenges. Google recently published a whitepaper that found an eye disease-predicting system was impractical in the real world, partially because of technological and clinical missteps. STAT reports that unproven AI algorithms are being used to predict the decline of COVID-19 patients. And companies like Babylon Health, which claim their systems can diagnose diseases as well as human physicians can, have come under scrutiny from regulators and clinicians.

CSAIL’s system aims to address this with a “classifier” that can predict a certain subset of tasks and a “rejector” that decides whether a given task should be handled by the classifier or an expert. The researchers behind the system claim the classifier is fairly accurate, achieving 8% better performance in the case of cardiomegaly (heart enlargement) compared with experts alone. But arguably its real advantage is customizability — the system allows a user to optimize for whatever choice they want, whether that’s prediction accuracy or the cost of the expert’s time and effort.

Efficiency is another advantage of the system’s approach. Through experiments on tasks in medical diagnosis and text and image classification, it was shown not only to achieve better performance than baselines but to do so with less computation and far fewer training samples.

The researchers haven’t yet tested the system with human experts — instead, they developed a series of “synthetic experts” so they could tweak parameters like experience and availability. The current iteration requires onboarding to acclimate to particular people’s strengths and weaknesses, but the team’s plans call for architecting systems that learn from biased expert data and work with (and defer to) several experts at once.

“There are many obstacles that understandably prohibit full automation in clinical settings, including issues of trust and accountability,” David Sontag, lead author and Von Helmholtz associate professor of medical engineering in MIT’s Department of Electrical Engineering and Computer Science, said in a statement. “We hope that our method will inspire machine learning practitioners to get more creative in integrating real-time human expertise into their algorithms.”

Live Updates for COVID-19 CASES