Date: 09 Mar 2022

One major discussion at NASH-TAG this year was about the inconsistency in ballooned hepatocyte identification and how this inconsistency inflates screen fail rates and possibly placebo response across studies. This episode explores that issue in detail.

The episode is sponsored by HistoIndex. At the back of the episode is a discussion of how artificial intelligence-driven assistive technology can improve the consistency of ballooned hepatocyte scoring in advanced fibrosis and support the development of robust outcomes for fibrosis studies.

This conversation includes three of the paper’s authors, including last author Quentin Anstee. The conversation starts with Professor Anstee discussing how the paper came to be. He quickly jumps into the meat of the discussion, which is the relatively low concordance between highly skilled world-class hepatopathologists over how many ballooned hepatocytes they see on a slide of a patient who might have NASH. The results suggested significant differences between two world-class hepatopathologists looking at the same slide in terms of the presence and significance of hepatocytes.

The next stage of the episode consists of other panelists praising the study while asking questions about its implications. Stephen Harrison asks how we can take these findings into drug development. Mazen Noureddin asks whether these results suggest either that we should question using ballooned hepatocytes in drug development or, more likely, we should shift immediately to an AI-mediated solution to get more consistent results. Jörn Schattenberg asks whether liquid biopsies would improve prediction and, presciently, whether we are asking more from liver histology than it can deliver.

This points to a pivotal issue: liver histology was developed initially to characterize individual patients qualitatively. Today, we call these findings “semi-quantitative” and use them to analyze trial results.

Two problems: the error embedded in the differences between how two pathologists would interpret the same slide is far greater than the sampling error, leading us to undersample and overinterpret. Second, we try to “prove” the value of blood-based biomarkers and AI-based assistive technologies by comparing them to something that is deeply flawed in the first place.

The extrasode discusses how HistoIndex used created the qBallooning2 analysis to improve the common identification of ballooned hepatocytes in patients with advanced fibrosis. In addition, Mazen Nourediin discussed his AASLD Poster of Distinction from 2022, which looks at ways that artificial intelligence techniques can support a more robust analysis of improved outcomes in patients with cirrhosis.

In the end, though, no 4,000 character summary can do this episode justice. To fully get the impact of this critically important discussion, plan to listen at least twice. Listen to the first part to comprehend the scope of the issue around ballooned hepatocytes and the challenges it causes, and the second part to learn how inventive data scientists can create better solutions by processing large amounts of data and being willing to challenge historical assumptions.

This episode is sponsored by HistoIndex.