Skip to main content
ACG 2023 Annual Meeting Main banner

Sunday Poster Session

Category: Liver

P0942 - Explainable Artificial Intelligence Accurately Predicts High-Risk Nonalcoholic Steatohepatitis and Identifies New Subphenotypes in Lean Individuals

Sunday, October 22, 2023
3:30 PM - 7:00 PM PT
Location: Exhibit Hall
ACG Award Winner
Award: Presidential Poster Award

Basile Njei, MD, MPH, PhD1, Jonathan Dranoff, MD2, Joseph Lim, MD1
1Yale School of Medicine, New Haven, CT; 2Yale University, New Haven, CT

Introduction: Lean status (body mass index: BMI < 25 kg/m2, or < 23 kg/m2 for Asians) is associated with a 90% increased risk of all-cause mortality in patients with Nonalcoholic Steatohepatitis (NASH). Lean individuals with NASH are also less likely to be screened because they may present with normal liver enzymes and without metabolic disorders. We used explainable artificial intelligence (XAI) to develop an algorithm integrating clinical and laboratory data for predicting high-risk NASH in lean individuals.

Methods: Data were derived from the National Health and Nutrition Examination Surveys (NHANES) 2017-March 2020. We used FibroScan-AST (FAST) score ≥0.35, to identify individuals with high-risk NASH. Extreme gradient boosting (XGBoost) was used to establish the prediction. The training set contained 50% of included patients. The optimized model was then validated by the independent holdout test set, on the remaining 50%. We illustrated the clinical feature’s importance and provided visualized interpretation using SHapley Additive exPlanations (SHAP). To identify high-risk NASH subphenotypes, an unsupervised distance-based clustering method was used.

Results: There was a total of 860 lean subjects meeting the inclusion criteria. The mean age was 51 (range 18-80 years), and 416 (48%) were women (Table 1). Subjects with high-risk NASH comprised 2.3% and 2.5% of training and test set, respectively. Figure 1A shows the areas under the receiver operating characteristic curve (AUC) of our model for predicting high-risk NASH. The optimized XGBoost model demonstrated a superior ability to predict high-risk NASH vs. FIB-4, and NAFLD fibrosis scores (AUC of 0.98 vs. 0.74, & 0.72, respectively, all p < 0.001). To explain the high performance of the XGBoost model, we found that the top 6 predictors of high-risk lean NASH were ALT, ferritin, insulin resistance, male gender, low HDL-cholesterol, and diabetes (Figure 1B). In addition to the predictive factors identified, 3 novel subphenotypes of high-risk lean NASH were identified by XAI: (1) men who smoke; (2) females with ferritin > 300 ng/mL; (3) and patients with low HDL-cholesterol (< 40 mg/dl).

Discussion: We used an XAI approach to develop a clinically applicable model that outperforms commonly used clinical risk indices and could increase the identification of high-risk lean NASH patients in resource-limited settings.

Figure: Figure 1A. AUROC of XGBoost model to identify lean high-risk NASH.
Figure 1B. Summary SHAP Plot of Predictor Importance and Impact on Model Output. Distribution of Individual Shapley Additive Explanation (SHAP) values and predictor values used in the model. SHAP values represent the contribution of each predictor to the model output, with positive and negative values indicating High-Risk NASH and No High-Risk NASH prediction, respectively. Feature values are color-coded to represent higher and lower predictor values.

 

 

 

 

 

 

  Overall No High-Risk NASH High-Risk NASH p-value

n

 

860

839

21

 

Age, mean (SD)

 

51.2 (20.5)

51.2 (20.5)

51.7 (20.3)

0.909

Race, n (%)

Hispanic

64 (7.4)

62 (7.4)

2 (9.5)

0.497

Other Hispanic

71 (8.3)

69 (8.2)

2 (9.5)

 

Non-Hispanic White

313 (36.4)

309 (36.8)

4 (19.0)

 

Non-Hispanic Black

200 (23.3)

193 (23.0)

7 (33.3)

 

Non-Hispanic Asian

176 (20.5)

172 (20.5)

4 (19.0)

 

Other

36 (4.2)

34 (4.1)

2 (9.5)

 

Gender, n (%)

Male

444 (51.6)

426 (50.8)

18 (85.7)

0.003

Female

416 (48.4)

413 (49.2)

3 (14.3)

 

Average drinks per day, mean (SD)

 

1.2 (0.3)

1.2 (0.3)

1.5 (0.4)

0.028

Hemoglobin A1C (%), mean (SD)

 

5.6 (0.9)

5.6 (0.9)

5.9 (1.0)

0.272

BMI, mean (SD)

 

22.2 (2.0)

22.2 (2.0)

22.2 (2.1)

0.944

Waist circumference (cm), mean (SD)

 

83.0 (8.0)

82.9 (8.0)

84.2 (9.2)

0.549

Triglyceride (mg/dL), mean (SD)

 

85.7 (42.1)

85.3 (41.7)

100.0 (55.8)

0.246

HDL (mg/dL), mean (SD)

 

58.6 (15.5)

58.9 (15.5)

48.7 (13.0)

0.002

Glucose (mg/dL), mean (SD)

 

105.7 (24.9)

105.6 (24.9)

111.8 (26.1)

0.293

AST (U/L), mean (SD)

 

21.4 (12.4)

20.2 (7.1)

68.1 (45.5)

< 0.001

ALT (U/L), mean (SD)

 

18.0 (13.0)

16.9 (9.2)

61.2 (40.8)

< 0.001

Ferritin (ng/mL), mean (SD)

 

144.2 (158.7)

138.4 (147.1)

377.3 (340.7)

0.004

Insulin (μU/mL), mean (SD)

 

7.2 (3.5)

7.1 (3.4)

9.7 (6.4)

0.088

Platelet count (1000 cells/μL), mean (SD)

 

233.1 (60.7)

233.6 (60.5)

212.4 (64.6)

0.151

Serum albumin (g/dL), mean (SD)

 

4.2 (0.3)

4.2 (0.3)

4.2 (0.4)

1.000

HOMA-IR, mean (SD)

 

1.9 (1.2)

1.9 (1.1)

3.0 (2.9)

0.091

 

 

 

 

 

 


Table: Table 1. Characteristics of Lean Adults Older than 18 Years with High-Risk and No High-Risk NASH in the National Health and Nutrition Examination Survey between 2017 and March 2020. Homeostasis Model Assessment of Insulin Resistance [HOMA-IR = insulin (μU/ml) × fasting glucose (mmol/L)/22.5].


Disclosures:
Basile Njei indicated no relevant financial relationships.
Jonathan Dranoff indicated no relevant financial relationships.
Joseph Lim: Gilead – Grant/Research Support. Intercept – Grant/Research Support. Inventiva – Grant/Research Support. Novo Nordisk – Grant/Research Support. Pfizer (Any division) – Grant/Research Support. Viking – Grant/Research Support.


Basile Njei, MD, MPH, PhD1, Jonathan Dranoff, MD2, Joseph Lim, MD1. P0942 - Explainable Artificial Intelligence Accurately Predicts High-Risk Nonalcoholic Steatohepatitis and Identifies New Subphenotypes in Lean Individuals, ACG 2023 Annual Scientific Meeting Abstracts. Vancouver, BC, Canada: American College of Gastroenterology.