The episode and this conversation is sponsored by Histoindex. This specific conversation focuses on ways and Artificial Intelligence assistive technologies and analyses can improve our abilities to assess efficacy in drugs that treat advanced fibrosis and cirrhosis.
For this extra-sode, HistoIndex Chief Scientific Officer Dean Tai joins Quentin Anstee, Mazen Noureddin, Joern Schattenberg, and Roger Green to discuss how AI-based algorithms can support improved analysis of ballooned hepatocyte changes both in advanced fibrosis and cirrhosis patients.
The rest of the conversation probes how this work will affect diagnosis and drug development, what other tests panelists can foresee, and areas where Histoindex is looking to create new algorithms and improve existing ones.
This conversation starts with Histoindex Chief Scientific Officer Dean Tai discussing the approach his company takes to AI-assisted hepatopathology. Dean starts by discussing briefly how staining, which is extremely helpful when the goal is to define the individual patient, becomes one more source of error in the more quantitative approach necessary for drug trials. He goes on to point out that while they can achieve 90% success in reproducing an individual coder’s result using AI, their goal is to achieve 99% success. He finishes by defining the goal as “majority-agreed hepatocytes,” hepatocytes where 5 or more of the 9 pathologists in the initial exercise agreed that a hepatocyte had ballooned. As Dean puts it, with ballooned hepatocytes, “you are really trying to identify bad apples from all apples,” not “oranges from apples.”
In explaining the Histoindex approach, Dean describes some pathologists as “under-callers” who identify relatively few ballooned cells and others as “over-callers” who identify far more cells. The primary difference between the two groups was how large they needed a cell to be before they classed it as “ballooned.”
Because these differences were systematic and structural, Quentin questions whether we can ever “train” consistent responses. He suggests that consistency will grow for a while but then coders will revert more to their historical patterns. In response to a question from Mazen, Quentin goes on to note that the number of pathologists necessary to validate a ballooned cell will vary inversely with the number of cells identified. For example, a model based on a 7-member agreement produces results that are more specific, less sensitive. A model based on a 3-member agreement would produce more sensitivity, less specificity. As a result, the team settled on 5 (majority of 9) as the best-rounded number. Mazen responds that for drug trials, being more specific is preferable because it creates a better chance for the drug to appear efficacious when it is.
The panelists go on to note that this tool is extremely helpful today in Phase 2, but not in Phase 3 and Dean explains some of the concepts Histoindex is working on to support future use in Phase 3 trials.
At that point, the conversation shifts to having Mazen discuss his work in cirrhosis. Going back through historic work, and particularly separate work from Drs. Garcia-Tsao and Younossi, Mazen identified three features to track: septal thickness, nodular features, and fibrosis area (the SNOF score). Using these metrics Mazen sought data from the Galectin trials because these were among the few trials that measures portal pressures. The Galectin data allowed researchers to correlate these kinds of measures to a 20% change in portal pressures. This score wound up being reliable in detecting portal hypertension and in two particularly pivotal measures: detecting the presence of varices and changes of greater than 20% in portal pressures.
Guests include: Roger Green, M.B.A., Guest Panelist, Mazen Noureddin, Jörn M. Schattenberg, Quentin M. Anstee, Dean Tai
