MACHINE LEARNING ADVANCED FIBROSIS IN NASH (ALADDIN) WITH WEB-BASED CALCULATION FOR PROBABILITY PREDICTION
Recommended Citation
Dunn W, Alkhouri N, Kundu R, Robert S, Nadeem R, Dunn N, Wong VW, Verma N, Yip TC, Loomba R, Abdelmalek MF, Diaz LA, Devuni D, Castera L, Noureddin M, Jafri S, Arab JP, Charlton MR, Wong GL, Yang L, Duseja AK, Chen V, Singal AK, Harrison SA, Al Yassin A, Hino K. MACHINE LEARNING ADVANCED FIBROSIS IN NASH (ALADDIN) WITH WEB-BASED CALCULATION FOR PROBABILITY PREDICTION. Hepatology 2023; 78:S829-S835.
Document Type
Conference Proceeding
Publication Date
10-25-2023
Publication Title
Hepatology
Abstract
Background: Nonalcoholic Fatty Liver Disease (NAFLD) prevalence poses a significant challenge, with fibrosis stage acting as a crucial prognostic indicator. Advanced fibrosis/cirrhosis (F3-4) cases often face swift disease progression. Non-Invasive Tests (NITs) for hepatologist referral or biopsy have limited accuracy. We propose a machine learning (ML) model offering a probabilistic prediction adaptable for both community and referral centers, where prevalence varies from 3.7 - 50%. Preliminary two-center data from the multi-center consortium inform this model. Methods: We collected retrospective data on patients diagnosed with NAFLD, Nonalcoholic Steatohepatitis (NASH), or cryptogenic cirrhosis; patients with other liver diseases were excluded. The primary outcome was advanced fibrosis (F3) or cirrhosis (F4). Data was divided into derivation (training) and validation (testing) cohorts. We employed Random Forest for ML, considering other models such as ElasticNet and Grant Boosting Machines. Results: The study incorporated 986 patients, with 269 having advanced fibrosis or cirrhosis. The mean FIB-4 was 1.76 (SD 1.52). The proposed ALADDIN score presented a 1 - Out-of-Bag (OOB) error rate of 78.3%, suggesting a strong fit to the training data. In the validation cohort, ALADDIN score demonstrated an AUC of 0.794 (95% CI 0.750 - 0.837), surpassing FIB-4 0.747 (0.697 - 0.798), p= 0.0039. With a 65% probability threshold, ALADDIN showed a PPV and NPV of 79%, against FIB-4's 65% PPV and 80% NPV at a 2.66 threshold. The Net Reclassification Improvement (NRI) of 0.379 and Integrated Discrimination Improvement (IDI) of 0.068 accentuate ALADDIN Score's enhanced reclassification and discrimination capacities over FIB-4. Figure 1 displayed ROC comparison of ALADDIN and FIB-4, and the top 20 variables. Conclusion: Preliminary data underscores the ALADDIN score's potential in outperforming the conventional FIB-4 score in a tertiary referral center setting. Patients with an ALADDIN score >65% might require liver biopsy, 15% - 65% may need additional noninvasive testing, and < 15% can be monitored in primary care. The ALADDIN score facilitates tailor-made care based on specific cirrhosis probability. With further data, the model can be calibrated for community settings. This is the first ML model with an online calculator for public use, enabling personalized care. The model is accessible at https://globalalchep.shinyapps.io/ALADDIN/.
Volume
78
First Page
S829
Last Page
S835
