Page 102 - AIH-2-2

Artificial Intelligence in Health Cirrhosis prediction in hepatitis C

Figure 5. Feature attribution at two visits for an exemplary patient from the representative RNN model when 100% of the training and validation sets was
used for the labeled cohort. The length of each bar represents the absolute value of the feature gradient with respect to the loss function at the specific visit,
with the number next to each bar indicating the corresponding absolute gradient. The color of the bar reflects the centered feature value: red represents
values above the mean, and blue represents values below the mean. The deeper the color, the greater the deviation from the mean, indicating more extreme
values of the feature.
Abbreviation: RNN: Supervised recurrent neural network.

has robust performance and has largely supplanted liver our test data into different subgroups and evaluated the
biopsy in the assessment of fibrosis in CHC before antiviral optimal model selected through hyperparameter tuning.
treatment, controversy remains in the performance of TE Our findings indicate that the RNN model exhibited better
after antiviral therapy. predictive ability for female subjects (0.772) compared
Second, our data from the VHA population is to male subjects (0.745), for non-White subjects (0.748)
inherently imbalanced with a majority of male subjects, versus White subjects (0.737), and better predictive ability
potentially limiting the generalizability of results to other on subjects who did not achieve SVR (0.746) at time t
populations with CHC, as cirrhosis risk factors can vary compared to those who did (0.742), for the RNN model, as
significantly between genders. For example, women shown in Table 3. These results along with Table S1 suggest
tend to have a different metabolic response to alcohol that the model’s performance remains robust across
and certain hepatotoxic medications, leading to a lower these non-dominant features, mitigating concerns about
threshold for liver damage compared to men. In addition, generalizability. However, future work with more balanced
24
autoimmune liver diseases, such as primary biliary cohorts is needed to validate our machine learning models
cirrhosis and autoimmune hepatitis, are more prevalent in in external populations.
women and may influence their liver disease risk profiles. Third, our predictor variables were limited to those that
25
These gender-specific risk factors may not have been fully could be extracted from the VHA’s large administrative
captured in our analysis. Given the challenges in obtaining healthcare database. This prevented the inclusion of
electronic health records from external populations, alcohol use as a predictor. However, we were able to include
we addressed this concern by stratifying our predictive several serologic markers suggestive of ongoing alcohol
performance testing by gender, race, and SVR achievement use, including AST, AST:ALT ratio, and platelets. Similarly,
to assess the impact of sampling bias. Specifically, we divided diabetes was not specifically included as a predictor,

Volume 2 Issue 2 (2025) 96 doi: 10.36922/aih.4671