P1813 - Multivariable Liver Fibrosis Staging Prediction Using Machine Learning in Nonalcoholic Fatty Liver Disease (AI-LF)

Monday, October 25, 2021

10:30 AM – 4:15 PM PT

Location: Shoreline Exhibit Hall

Has Audio

Presenting Author(s)

Sujoy Kar, MBBS, MD

Apollo Hospitals
Hyderabad, Telangana, India

Mahesh K. Goenka, MD, DM, FACG¹, Gajanan Rodge, MD², Enam Murshed Khan, MBBS, MD³, Usha Goenka, MBBS, MD³, Bharath Potla, ⁴, Raj Shukla, ⁴, Sujoy Kar, MBBS, MD⁴
¹Institute of Gastrosciences, Apollo Gleneagles Hospital, Kolkata, West Bengal, India; ²Institute of Gaastrosciences, Apollo Gleneagles Hospitals, Kolkata, West Bengal, India; ³Apollo Hospitals, Kolkata, West Bengal, India; ⁴Apollo Hospitals, Hyderabad, Telangana, India

Introduction: Nonalcoholic fatty liver disease (NAFLD) has emerged to be a global epidemic and is the most common liver disease worldwide, ranging from isolated steatosis to steatosis plus inflammation with or without fibrosis. The objective of this study is develop and validate a Machine Learning Model to identify and distinguish advanced fibrosis using Elastography and Liver Biopsy.

Methods: Data of 5150 NAFLD patients (Advanced Fibrosis F2-4 – 30.34%) was collected from Apollo Hospitals, Kolkata period 2011 to 2017 using standardized template and electronic medical records (ICD 10 codes). 25 Clinical and Laboratory parameters were studied along with patients’ Elastography reports and ARFI values. The Machine Learning (ML) modelling was performed using eXtreme Gradient Boosting (XGB) algorithm. The prospective validation cohort was selected of 1261 patients (F2-4 – 31.24%) from 2018 to 2020 and compared with Fib4 Score. This was further validated with 98 Liver Biopsies from validation cohort.

Results: Out of the 25 clinical and laboratory parameters, 11 variables including Age [Multivariate Odds Ratio (OR)– 3.39; 95%CI 2.99 – 3.84], History of Diabetes Mellitus [OR – 6.80, 95%CI 5.92 – 7.81], Albumin [OR– 3.70, 95%CI 3.25 – 4.20], Aspartate aminotransferase (AST) [OR- 3.65, 95%CI – 3.21 – 4.16}, Total Bilirubin [OR–3.13, 95%CI 2.76 – 3.56] and Platelet Count [OR–2.74, 95%CI 2.40 – 3.13] were found to be significant. The performance parameters of the development model is AUC ROC Score of 0.94 and validation cohort had the AUC and accuracy of 0.88. The AUC for 98 liver biopsy validation cohort was 0.83. The model performed better than Fib4 Score with Net Reclassification Improvement(NRI) at 0.499.

Discussion: The model comparing Advanced Liver Fibrosis (F2-4) from No or Low Fibrosis (F0/F1) provides insights into the Clinical and Laboratory Parameters and accurately predicts onset of liver fibrosis in NAFLD which could be useful as clinical decision support in low-cost settings.

Figure: Figure 1 - Left : Forrest Plot - Multivariate Odds Ratios of the 11 Predictors, for Development (top) & Validation Cohorts (bottom).
Right (Top) AUC ROC Graph for 98 Liver Biopsy Validation Cohort. Right (Bottom) Violin Graphs - Combination of Box Plots and Kernel Density Plots reflect on different clinical and lab variables and their representation for Advanced Fibrosis (F2-4) vs. No / Low Fibrosis in NAFLD Patients

	AILF		FIB4
Source of Algorithm / API	Internal		https://www.mdcalc.com/fibrosis-4-fib-4-index-liver-fibrosis
Confusion Matrix	F2-4	F0/F1	Fibrosis	No Fibrosis
Positive Cases (Advanced Fibrosis F2-4) : 394	291	103	191	203
Negative Case (No / Low Fibrosis F0/F1) : 867	86	781	74	793
Sensitivity	77.19% (72.62% to 81.33%)		72.08% (66.26% to 77.39%)
Specificity	88.35% (86.05% to 90.39%)		79.62% (76.98% to 82.08%)
Positive Likelihood Ratio (PLR)	6.62 (5.48 to 8.01)		3.54 (3.06 to 4.08)
Negative Likelihood Ratio (NLR)	0.26 (0.21 to 0.31)		0.35 (0.29 to 0.43)
Positive Predictive Value	73.86% (70.03% to 77.35%)		48.48% (44.90% to 52.07%)
Negative Predictive Value	90.08% (88.28% to 91.63%)		91.46% (89.81% to 92.88%)
Accuracy	85.01% (82.92% to 86.94%)		78.03% (75.64% to 80.29%)
Net Reclassification Improvement NRI NRI + (Events = F2-4) NRI - (Non Events = F0/F1)	0.499 0.196 0.302

Table: Comparison between AILF model vs FIB4

Disclosures:

Mahesh Goenka indicated no relevant financial relationships.

Gajanan Rodge indicated no relevant financial relationships.

Enam Murshed Khan indicated no relevant financial relationships.

Usha Goenka indicated no relevant financial relationships.

Bharath Potla indicated no relevant financial relationships.

Raj Shukla indicated no relevant financial relationships.

Sujoy Kar indicated no relevant financial relationships.

Mahesh K. Goenka, MD, DM, FACG¹, Gajanan Rodge, MD², Enam Murshed Khan, MBBS, MD³, Usha Goenka, MBBS, MD³, Bharath Potla, ⁴, Raj Shukla, ⁴, Sujoy Kar, MBBS, MD⁴. P1813 - Multivariable Liver Fibrosis Staging Prediction Using Machine Learning in Nonalcoholic Fatty Liver Disease (AI-LF), ACG 2021 Annual Scientific Meeting Abstracts. Las Vegas, Nevada: American College of Gastroenterology.