Investigation of predictive factors for fatty liver in children and adolescents using artificial intelligence

Aliakbar Sayyari¹Amirhossein Hosseini¹Amin Magsudy²Tahereh hatamii¹Mohammadreza Arzaghi³Hamidreza Amiri⁴Fereshteh Sohrabivafa⁵AmirAli Okhovat⁶Naghi Dara¹Negar Imanzadeh³Farid Imanzadeh^1*Mahmoud Hajipour^1*

• ¹Pediatric Gastroenterology, Hepatology and Nutrition Research Center, Research Institute for Children's Health, Shahid Beheshti University of Medical Sciences, Tehran, Iran, Tehran, Iran
• ²Islamic Azad University, Tabriz, Tabriz, Iran
• ³Shahid Beheshti University of Medical Sciences, Tehran, Tehran, Iran
• ⁴Arak University of Medical Sciences, Arak, Markazi, Iran
• ⁵Dezful University of Medical Sciences (DUMS), Dezful, Khuzestan, Iran
• ⁶Tehran University of Medical Sciences, Tehran, Tehran, Iran

Childhood obesity is a growing problem worldwideglobal health concern, leading to contributing significantly to the prevalence of non-alcoholic fatty liver disease (NAFLD), which is the most common liver disease in children. Liver biopsy is the gold standard for NAFLD diagnosis. Machine learning algorithms could assist in an early diagnostic approach and leading to a favorable prognosis. Early diagnosis and intervention are crucial to prevent progression to more severe liver conditions. Artificial intelligence (AI) and machine learning (ML) offer promising approaches for predicting NAFLD in pediatric populations. Objective: This study aimed to identify predictive factors for NAFLD in children and adolescents using machine learning models, focusing on liver biopsy outcomes such as fibrosis, infiltrationfiltration, ballooning, and steatosis. Methods: Data from 659 children suspected of NAFLD, who underwent liver biopsy at Mofid Children's Hospital between 2011 and 2023, were analyzed. The dataset included both categorical and numerical variables, which were processed using one-hot encoding and standardization. Several machine learning models, including CatBoost, AdaBoost, Random Forest, and others, were trained and evaluated were trained and evaluated, including CatBoost, AdaBoost, Random Forest, and others. Model performance was assessed using cross-validation with metrics such as accuracy, precision, recall, F1 score, and ROC-AUCaccuracy, precision, recall, F1 score, and ROC-AUC metrics. Feature importance was determined through permutation analysis. Results: Among NAFLD patients, the CatBoost Classifier achieved the highest accuracy (91.8%) and ROC-AUC score (92.3%). Among NAFLD patients, the CatBoost Classifier achieved the highest accuracy (91.8%) and ROC-AUC score (92.3%) in cross-validation. In addition, the adjusted models showed better results. That is, the F1 for the CatBoost raised from 83% to 89% (AUC: 0.86 to 0.92), for the GradientBoosting from 76% to 81% (AUC: 0.81 to 0.85), and for Bernolli Naive Bayes from 78% to 82% (AUC: 0.82 to 0.85). Vitamin D, alanine transaminase (ALT), and patient height were identified as key predictors of fibrosis. Cholesterol was the most impactful feature for model differentiation, while abdominal pain was significant for the Bernoulli Naive Bayes model. Conclusion: Machine learning models, particularly CatBoost, demonstrated strong predictive capabilities for NAFLD diagnosis in children