Page 101 - AIH-2-2

Artificial Intelligence in Health Cirrhosis prediction in hepatitis C

A B C

Figure 3. Model performance when various percentages of the training and validation set was used for the labeled cohort. We tested the four developed
models using 10 identical testing sets, each with best hyperparameters selected during the validation process. To evaluate performance, we used three
different metrics: (A) AuROC; (B) Brier score; and (C) AuPRC. We then calculated the mean of these metrics across all 10 splits to represent the
performance of each method using a certain percentage of training and validation data from the labeled cohort.
Abbreviations: AuROC: Area under the receiver operating characteristic curve; AuPRC: Area under the precision-recall curve; RNN: Supervised recurrent
neural network; Semi-RNN: Semi-supervised recurrent neural network.

can accurately predict the progression to cirrhosis among
patients with CHC under both supervised and semi-
supervised learning. In particular, semi-supervised RNN
achieved a better predictive performance when we utilized
all of our unlabeled data. Unlike prior studies, our cohort
includes patients who had received prior HCV treatment,
given that the risk of fibrosis progression may persist after
antiviral therapy due to other patient cofactors.
Our results demonstrate that machine learning models
are feasible options for estimating cirrhosis progression
risk in large populations. We anticipate many potential
uses, especially in guiding outreach interventions for
people with the greatest risk for progression to cirrhosis.
Figure 4. Calibration curve of models when 100% of the training For example, using the RNN model, we determined that
and validation sets was used for the labeled cohort. We selected a 90% of all cirrhosis diagnoses occurred in samples with the
representative split that had an AuROC closest to the mean value across
10 splits for the RNN model. Then, we generated the calibration curves highest mean (SD) 67% (2.1%) of risk scores, whereas 80%
for the four models we developed, using 10 quantiles, and compared them of cirrhosis occurred in samples with the highest mean
to the perfect calibration line. (SD) 52% (1.8%) of risk scores. Therefore, the RNN model
Abbreviations: AuROC: Area under the receiver operating characteristic suggests potential benefit in focusing proactive outreach
curve; RNN: Supervised recurrent neural network; Semi-RNN: Semi- on the top 52% (or 67%) of samples with the highest risk
supervised recurrent neural network.
scores, where 80% (or 90%) of cirrhosis cases developed,
respectively (Table 2). Given that cirrhosis is the most
4. Discussion important risk factor for liver cancer, as well as the
Our findings suggest that machine learning methods cause for multiple disease complications in its own right,
could be useful in identifying patients at high risk for clinical decision support for cirrhosis screening could be
progression to cirrhosis, as defined by TE outcomes. We implemented based on individualized risk.
developed and compared four machine learning models The VHA system contains the largest cohort of CHC
for predicting cirrhosis development, including three patients in a single U.S. healthcare system and is an ideal
supervised learning models which were trained using only environment for developing machine-learning models for
patients with known TE outcomes, and a semi-supervised cirrhosis prediction. Nevertheless, our results should be
3
learning model that incorporated both patients with and interpreted within the context of several limitations. Most
without TE outcomes for training. We found that RNN importantly, we used TE as a proxy for cirrhosis, as opposed
models, which are good at processing sequential data, to liver biopsy – the historical gold standard. While TE

Volume 2 Issue 2 (2025) 95 doi: 10.36922/aih.4671